From the individual, to the team, to the organization — and how change spreads, sticks, or quietly doesn’t

Version 2 · 10 August 2026 · Lucian Cosinschi Supersedes v1 of 1 August 2026 (corrections logged in Appendix B)

Purpose: academic grounding Stance: rigorously critical — evidence separated from folklore throughout

---

How to read this document

The field of change management has an unusual pathology. It is enormous, commercially lucrative, taught in every business school, and a substantial portion of its most confidently repeated content has no empirical basis whatsoever. Meanwhile two parallel bodies of work — behavioural science and implementation science — have produced genuinely rigorous findings that rarely reach the corporate change practitioner.

This report holds all three. Every substantive claim carries an evidence grade:

Grade	Meaning
[A]	Meta-analysis, systematic review, or replicated experimental evidence
[B]	Multiple primary studies, reasonable methods, some contest
[C]	Single study, correlational, theoretical, or contested construct
[D]	Practitioner benchmarking, consultancy data, self-interested source
[X]	Widely believed, evidentially unsupported — folklore

A [D] is not worthless. Prosci’s benchmarking is the largest dataset in existence on what change practitioners actually do, and its central findings are broadly consistent with the academic work. But it is collected by a firm selling the methodology it validates, and that must be visible in the grade.

A [X] does not mean useless. Some folklore encodes real practitioner wisdom in a form that happens to be false as stated. The grade concerns the evidence, not the utility — but you should know which you are trading on.

Navigation. Part 0 is the compressed version. Part 15 is the folklore table — the highest-utility single page for client work. Part 13 is the practical process. Part 18 is how to deploy this commercially. Appendix A is the verification log; Appendix B lists what changed from v1 and why.

---

Part 0 — The compressed version

Twenty-three findings that survive scrutiny.

1. The dominant change models are not evidence. Lewin never wrote unfreeze–change–refreeze [A]. Kotter’s eight steps have never been validated as a sequence [A]. The “70% of change fails” statistic has no source study [A]. The Kübler-Ross change curve is a double extrapolation from a grief model that was itself never validated [B]. These frameworks are serviceable checklists. They are illegitimate as proof.

2. Behaviour change is a context problem before it is a motivation problem. Habits are cue–response associations bound to environments. Change the environment and you change the behaviour more reliably than by changing the person’s mind [B]. The strongest single confirmation comes from an unexpected direction: Szulanski’s study of 122 best-practice transfers across eight companies found that, contrary to conventional wisdom blaming motivation, the major barriers were knowledge-related — the recipient’s absorptive capacity, causal ambiguity, and an arduous relationship between source and recipient [B].

3. The intention–action gap is the central individual-level problem, and planning closes part of it. Implementation intentions (“when X happens, I will do Y”) show d ≈ 0.65 across 94 tests in the foundational meta-analysis, updated across 642 tests in 2024 [A]. It is the cheapest high-yield intervention in the literature.

4. Nudge, as a general category, is much weaker than advertised — but friction is real. Mertens et al. reported a pooled d = 0.43; Maier et al. showed no evidence of effect surviving publication-bias correction; a 2025 second-order meta-analysis put the aggregate at d = 0.27, falling to d = 0.004 after adjustment [A]. What does survive is structural: defaults, and the removal or addition of friction. Sunstein’s “sludge audit” — cataloguing and costing unjustified friction — is the durable practical residue and has been formalised internationally, including in OECD guidance [B].

5. Training transfer, not training quality, is where value is lost. The work environment dominates instructional design as a determinant of transfer [A]. Meta-analytic SEM across 32 studies and 5,487 participants found the social-support model accounted for 32% of transfer variance, with peer support the strongest predictor of motivation to transfer and feedback/coaching the strongest predictor of transfer itself [A].

6. Behaviour modelling training is the best-evidenced method for interpersonal skill. 117 studies, effect sizes slightly above 1.0 SD against controls, with skill and job-behaviour effects stable or increasing over time even as declarative knowledge decayed [A]. Demonstrate, practise, feedback, plan for use. It has been the answer for forty years.

7. Leadership training works better than cynics claim, when designed properly. 335 studies: δ = 0.73 learning, 0.82 transfer, 0.72 results, with fifteen significant design moderators [A]. Needs analysis, spacing, practice and feedback do most of the work.

8. Feedback is not automatically good. 607 effect sizes, mean d = 0.41, and over a third of feedback interventions reduced performance [A]. Feedback that directs attention to the self rather than the task is the failure mode. Thirty years old, still routinely ignored.

9. Consequences change behaviour, and money is a narrow instrument. Organizational behaviour modification — antecedents, behaviour, consequences applied systematically — produced d = 0.51, roughly a 17% performance increase, across two decades of field research [A]. Financial incentives correlate .34 with performance quantity and essentially not at all with quality (.08) [A]. And tangible, expected, performance-contingent rewards reliably undermine intrinsic motivation (d ≈ −0.28 to −0.40 across 128 studies) [A]. Pay for volume, not for judgement; and never pay for something you want people to care about.

10. Teams are where change is retained or lost. Psychological safety is the best-evidenced team-level enabler — 136 samples, 22,000+ individuals [A] — and, contrary to what was long true, it is now also manipulable: a field RCT covering roughly 7,000 employees in 1,000 teams raised it through a low-touch intervention on the content and frequency of manager one-to-ones [B].

11. Behaviour spreads differently than information. Complex contagion requires redundant social reinforcement, so behaviour spreads faster through clustered networks than through networks rich in weak ties [B]. Seed dense clusters, not well-connected individuals.

12. The diffusion literature already answered “why won’t they adopt it?” Rogers’ five attributes — relative advantage, compatibility, complexity, trialability, observability — explain a large share of variance in adoption rates across product categories [B], and Greenhalgh et al.’s systematic review adapted them for service organizations [A]. Most corporate “resistance” is an innovation-attribute problem misdiagnosed as an attitude problem.

13. Implementation science is the most rigorous body of change knowledge, and corporate practice ignores it. CFIR (five domains, 39 constructs), ERIC (73 named strategies), RE-AIM (five evaluation dimensions), NPT (four embedding mechanisms) [A/B]. Built where failed implementation kills people. Free, and unknown in the commercial market.

14. Sustainment is a separate problem with separate causes. Roughly half of evidence-based interventions are not sustained [B]. Every popular model treats “make it stick” as a final step; it is the majority of the work.

15. Organizations adopt practices for legitimacy, not only effectiveness — and then decouple them from real work. Meyer and Rowan’s account of formal structure as myth and ceremony, and DiMaggio and Powell’s three isomorphic pressures (coercive, mimetic, normative), explain more about why unevidenced change methodologies persist than any critique of the methodologies themselves [B]. Abrahamson’s management-fashion research documents the characteristic bell-shaped popularity curve empirically [B].

16. Organizations are running more change than people can absorb. Change fatigue is a portfolio-capacity problem, not a resilience deficit [B/D]. Stopping things is a change intervention, and usually the neglected one.

17. Popular L&D and change instruments are largely unevidenced. Learning styles [X], MBTI for anything consequential [A against], the learning pyramid percentages [X], andragogy as a validated theory [C/X], Tuckman’s stages as a reliable sequence [C], grit as distinct from conscientiousness (ρ = .84) [A against], growth mindset interventions at scale [A, contested]. Knowing this precisely is a professional differentiator — deployed with tact.

18. AI adoption in 2025–26 is a live, well-documented case of everything in this report. Randomised trials show real productivity gains inside the capability frontier and degraded accuracy outside it [A]. MIT’s 2025 analysis reported ~95% of enterprise GenAI pilots delivering no measurable P&L impact, alongside ~90% of surveyed workers using personal AI tools daily — a textbook decoupling of official structure from actual practice [C/D].

19. The nameable artefact is never the mechanism. The checklist, the keystone habit, the kata card, the nudge — each is the visible object around which a social and structural programme was built. Michigan proves it three times over: the same checklist worked in Michigan ICUs, produced nothing in Michigan surgery, and underperformed in England [A].

20. Small changes do not accumulate into transformation. Romanelli and Tushman’s empirical test of punctuated equilibrium found that most organizational transformations occurred through rapid, discontinuous change across most or all domains, and that small changes in strategy, structure and power distribution did not accumulate into fundamental transformation [B]. Incremental methods are right for embedding practice and wrong for reorientation — and confusing the two is a major source of disappointment in change work.

21. Stopping is a skill with its own literature. De-implementation research shows that the strategies which reliably reduce low-value practice are contextual tailoring, stakeholder relationships, and infrastructure and workflow change — not education or exhortation — and that replacing a stopped practice works better than merely removing it [B].

22. Almost all of this evidence comes from a narrow slice of the world. Roughly 70–80% of organizational research samples are drawn from the US and Europe, representing about 11% of the global population [B]. Mechanisms travel better than techniques; state the population when you cite the effect.

23. What recurs across every level is a small set of mechanisms — cue and context; friction; capability; absorptive capacity; social proof; observability; visible authority commitment; participation; task-directed feedback; consequences; routine embedding; and finite attention. If a change effort cannot be described in those terms, it is probably a communications campaign.

---

Part 1 — Why this literature is a mess

1.1 Three literatures that barely speak

Organizational behaviour / organization development (OD). Origins in Lewin, the National Training Laboratories, and post-war group dynamics. Produced Kotter, Beer, Schein, Burke, Bridges, ADKAR. Its practitioner wing is enormous and commercial; its academic wing is respectable but has struggled to test its own core models. Characteristic method: case study, survey, occasional longitudinal field study.

Behavioural science / health psychology. Origins in social and cognitive psychology. Produced COM-B and the Behaviour Change Wheel, the BCT taxonomy, implementation intentions, self-determination theory, habit research, and the nudge movement. Characteristic method: randomised experiment. Much higher internal validity, much lower organizational realism, and a live replication crisis of its own.

Implementation science. Origins in healthcare, evidence-based medicine and international development. Produced CFIR, Normalisation Process Theory, RE-AIM, PARIHS, the ERIC compilation, the NASSS framework and the sustainment literature. Characteristic method: hybrid effectiveness–implementation trials, mixed methods, explicit determinant frameworks. The most methodologically serious of the three, and effectively unknown to corporate change practitioners.

Two further literatures sit adjacent and are routinely missed:

Diffusion and knowledge transfer — Rogers, Greenhalgh, Szulanski, Centola. Answers the question “why doesn’t this spread?” better than anything in OD.

Institutional theory and critical management studies — Meyer & Rowan, DiMaggio & Powell, Abrahamson, labour process theory. Answers the question “why do organizations keep buying things that don’t work?” These are not cynical asides; they are the best available explanation of the field’s own pathology.

The single most valuable move available to a change practitioner in 2026 is to import implementation-science rigour and diffusion vocabulary into an OD-dominated commercial field. Nobody in the room will have heard of CFIR or NASSS, and they are better than everything the room has heard of.

1.2 The evidence quality problem

Stouten, Rousseau and De Cremer’s 2018 Academy of Management Annals review is the honest map. They compared widely used practitioner models against scholarly evidence and concluded that practitioner reliance falls on models that “more often cite expert opinion as their foundation rather than scientific evidence,” producing an integrative summary of what is known, contested, untested and underused [A]. Their ten evidence-based steps appear in Part 5.

Four structural reasons the field stays poor:

Commercial incentive. Change models are products. A model that says “it depends on context, effects are modest, and half of what you’re doing is theatre” does not sell. A model with eight numbered steps and a failure-rate statistic that implies you need help does.

Unfalsifiable formulation. Most popular models are specified so loosely that failure is always attributable to poor execution rather than to the model. If a Kotter implementation fails, the diagnosis is “insufficient urgency” — never “the model is wrong.”

Citation laundering. Hughes’ 2011 investigation of the 70% claim traced it through five prominent sources and found each either asserted the number without evidence or cited another source that asserted it without evidence [A]. Assertion becomes fact through repetition.

Institutional demand for legitimacy over efficacy. Covered fully in Part 7. Organizations adopt visible, recognisable practices because doing so signals competence to boards, regulators, clients and staff. Efficacy is a secondary and often unmeasured consideration. This is not a scandal; it is a well-documented sociological regularity, and it means that critique alone will never displace a popular model.

1.3 What good evidence in this field looks like

Realistically you will rarely get randomised trials of organizational change. What you can demand:

• A specified behaviour. “Improve collaboration” is not a behaviour. “Post a written update in the shared channel before Thursday standup” is.
• A stated mechanism. Which of the twelve in Part 12 does this operate on?
• A counterfactual. Comparison sites, staggered rollout, stepped wedge, or at minimum pre/post with a named control unit.
• Behavioural, not attitudinal, outcomes. Satisfaction surveys measure the intervention’s popularity.
• Fidelity measurement. Did the thing actually get delivered as designed? Without this you cannot distinguish a bad idea from a bad rollout.
• Follow-up beyond three months. Almost all decay happens after the reporting period closes.

Stepped-wedge designs deserve emphasis: most organizations roll out in waves anyway for operational reasons, which means a credible quasi-experiment is available at near-zero marginal cost and is almost never taken. The obstacle is not methodological. It is that nobody asks in the planning meeting.

1.4 Whose behaviour was actually studied

A caveat that applies to essentially everything in this report, and that most syntheses omit.

The sampling problem. Organizational and psychological research draws overwhelmingly on WEIRD populations — Western, Educated, Industrialized, Rich and Democratic. Analyses of sample diversity in organizational research find something on the order of 70–80% of samples drawn from the United States and Europe, regions containing roughly 11% of the world’s population [B]. Within those, samples skew further toward students, large firms, white-collar work and major cities. Constructs that feel universal — trust, voice, autonomy, fairness, appropriate deference to authority — are culturally shaped, and the interventions built on them inherit that shaping.

The Hofstede problem, in both directions. The standard corrective is to reach for national culture dimensions, but that framework has its own well-documented weaknesses [B]: deriving stable national profiles from a single company’s employee survey data, using national averages that conceal large intra-national variation, treating culture as static, an instrument originally administered in English and developed against Western reference points, and an implicit developmental hierarchy in how the poles are described. Using Hofstede to correct for WEIRD bias imports a second bias.

The specific risk for change management. Organization development emerged in societies holding a strong belief that people can control their own destinies — participation, voice, self-directed development, “surfacing” disagreement. In more fatalistic or higher power-distance settings, those same interventions may be read as naïve, as a trap, or as an abdication of leadership responsibility. The practices are not culturally neutral technologies; they encode assumptions about agency and authority.

Practical handling. Not paralysis, but three habits:

1. State the population the evidence came from when you cite an effect. “Meta-analytic evidence, predominantly US and European white-collar samples” is a more honest sentence and costs nothing.
2. Treat mechanisms as more portable than techniques. Social proof, friction, capability and consequences travel better than the specific practices built on them. That people adopt what credible peers visibly do is robust; who counts as a credible peer is local.
3. Pilot for cultural fit explicitly, and read low adoption in a new context as a compatibility signal (Rogers) before reading it as resistance.

This also flags where the field’s evidence is thinnest and where genuinely original work is available — see Part 16.

---

Part 2 — The individual

2.1 What a habit actually is

The technical definition constrains what interventions can work. A habit is a learned association between a context cue and a response, acquired through repetition in a stable context, which comes to be triggered with minimal conscious deliberation. Wood and Neal’s context-cue account is the key theoretical statement: habits are not strong intentions, they are automated responses that can run counter to current intentions [B].

Three consequences follow, each contradicting something popular:

Motivation is required to start a habit and largely irrelevant to maintaining one. Interventions that keep pumping motivation into an established behaviour waste effort; interventions that fail to secure stable context waste everything.

Habits are things environments produce, not things people have. Move the person and a substantial share of their behaviour resets.

Willpower models are largely the wrong frame. The ego-depletion literature that underwrote a decade of willpower advice has not survived scrutiny: a preregistered multi-laboratory replication (k = 36, N = 3,531) found a non-significant effect of d = 0.06, and a separate 23-laboratory replication with 2,141 participants found the effect negligible [A]. Some researchers now argue for revised dynamic-systems formulations rather than abandonment [C], but the practical instruction is unchanged: design the cue, not the resolve.

2.2 How long habit formation actually takes

The 21-day figure is folklore [X], derived from Maxwell Maltz’s 1960 observation about plastic-surgery patients adjusting to their appearance — a clinical impression about adaptation, not a study of habit formation.

The real number, from Lally et al. (2010): the median time to reach 95% of asymptotic automaticity was 66 days, range 18 to 254 days, across 96 participants tracked over 12 weeks [B]. Three findings matter more than the headline:

• Missing a single day did not break the formation curve. Perfect streaks are not required, and streak-based tools may cause more abandonment than they prevent.
• Behaviours with strong contextual cues and immediate rewards automated fastest.
• Roughly half the participants had not reached habit-level automaticity even after 84 days.

Sample caveat, stated for honesty: 96 university-affiliated volunteers choosing simple self-selected health behaviours. Generalising the exact number to organizational behaviour is unsupported. Generalising the shape — asymptotic curve, enormous individual variance, months not weeks — is reasonable.

The uncomfortable programme-design translation: any change effort expecting behavioural automaticity within a quarter is expecting something that happens to about half of people at best, for simple daily behaviours, under favourable conditions. Complex workplace behaviours performed weekly rather than daily will take substantially longer.

2.3 Breaking habits is a different problem from building them

Under-covered in most practitioner treatments. Wood and Neal’s model identifies three approaches [B]:

Cue disruption — remove or alter the environmental trigger. Most effective, least used, because it requires changing something structural rather than something personal.

Environmental re-engineering — add friction to the unwanted behaviour, remove friction from the alternative, or coach people to redesign their own environments.

Vigilant monitoring — deliberate inhibitory attention (“don’t do it”) at the moment of temptation. This is the strategy people spontaneously reach for, and it does work — studies find it effective for curbing junk food, smoking and excess drinking — but it is cognitively expensive and therefore fragile under load [B].

The ordering matters for organizational work: if you want a behaviour to stop, changing what triggers it beats asking people to resist it. Most workplace “compliance” campaigns are vigilant-monitoring interventions applied at scale, which is the most effortful and least durable of the three.

2.4 The habit discontinuity hypothesis

Verplanken and colleagues’ finding: context change temporarily disrupts habitual patterns, making behaviour-relevant information more salient and creating a window in which new choices are actually deliberated [B]. Field tests around residential moves and workplace relocation support it, though later work using UK Understanding Society data suggests the window’s influence diminishes over time [C].

This is one of the most practically actionable findings in the report and is almost never used deliberately. Organizational life is full of context discontinuities — reorganisations, office moves, system migrations, team changes, new managers, onboarding, return from leave. Each is a naturally occurring window in which behaviour is unusually malleable, and each is typically treated as a disruption to be minimised rather than an opportunity to be used.

Practical rule: attach behaviour change to a discontinuity you already have. The cheapest change programme in existence is one that rides an office move or a system migration.

2.5 Habit stacking and chaining: what’s real

“Habit stacking” — anchoring a new behaviour to an existing one — is popular via BJ Fogg’s Tiny Habits and James Clear’s Atomic Habits. Assessed honestly:

• The underlying mechanism is implementation intentions, which are extremely well evidenced [A]. Gollwitzer & Sheeran found d = 0.65 across 94 independent tests and over 8,000 participants; Sheeran, Listrom and Gollwitzer’s 2024 update aggregated 642 tests and reframed the effect as conditional on plan format and motivational context rather than a single universal value [A].
• The specific claim that an existing habit is a superior anchor to a time or location cue is plausible but not well tested [C]. What is established is that a specific, salient, reliably-occurring cue beats a vague intention.
• Fogg’s “make it tiny” recommendation has weaker direct evidence than its popularity implies [C], though it is consistent with self-efficacy theory: early mastery experience is the strongest efficacy source, and efficacy predicts performance.
• Atomic Habits is a good synthesis with weak sourcing. Its four laws (obvious, attractive, easy, satisfying) map almost exactly onto the Behavioural Insights Team’s EAST framework and onto operant conditioning. Use it as a mnemonic, never as a citation.

Chaining in the technical behaviour-analytic sense — building complex behaviour by linking discrete steps, often taught backwards — comes from applied behaviour analysis, has strong evidence in skill acquisition [A], and is a different thing from habit stacking. It is directly applicable and under-used in workplace procedural training: teach the final step first so every practice attempt ends in successful completion.

2.6 COM-B and the Behaviour Change Wheel

Michie’s COM-B model holds that behaviour requires Capability, Opportunity and Motivation, each split in two (physical/psychological capability; physical/social opportunity; reflective/automatic motivation). Around it sits the Behaviour Change Wheel: nine intervention functions — education, persuasion, incentivisation, coercion, training, restriction, environmental restructuring, modelling, enablement — and seven policy categories — communication/marketing, guidelines, fiscal measures, regulation, legislation, environmental and social planning, service provision [verified].

Assessment:

• As a diagnostic it is the best available tool at the individual level [B]. It forces the analyst to ask whether the problem is can’t, no chance to, or won’t. Most corporate change defaults to persuasion when the actual deficit is opportunity or capability — a bias Szulanski’s data independently confirms.
• As a predictive theory it is weak. COM-B is a framework, not a falsifiable model; it cannot give you an effect size.
• The associated BCT Taxonomy v1 (Michie et al., 2013) catalogues 93 behaviour change techniques in 16 clusters, explicitly to standardise how interventions are reported [A for taxonomy reliability; B for evidence attached to individual techniques]. It is the closest thing the field has to a periodic table, and it is more useful to a consultant than any change model, because it lets you specify precisely what you are doing.

Practical chain: diagnose deficit (COM-B) → select intervention function (Wheel) → specify technique (BCTTv1). More rigorous than anything in mainstream change management, and it takes about an hour.

2.7 Motivation: self-determination theory

SDT distinguishes autonomous motivation (intrinsic and fully internalised extrinsic) from controlled motivation (external and introjected regulation), and posits three basic psychological needs — competence, autonomy, relatedness [A].

The workplace evidence is strong and consistent: need satisfaction predicts autonomous motivation, which predicts performance, persistence, learning, creativity, wellbeing and retention [A]. Meta-analytic work on leader autonomy support establishes that manager behaviour is a meaningful lever on need satisfaction [A]. Recent meta-analytic work distinguishes further: intrinsic motivation associates most with proactive performance and affective commitment, while identified regulation associates most strongly with overall work performance and citizenship behaviour [B] — a useful nuance, since fully-internalised extrinsic motivation is far more achievable in most jobs than genuine intrinsic motivation.

Two cautions:

• Much of the workplace SDT evidence is correlational and cross-sectional. Autonomous motivation and performance are mutually reinforcing; the causal arrow is less clean than advocates present.
• “Autonomy” in SDT means volition and self-endorsement, not independence or absence of structure. This is routinely mistranslated into “let people do what they want,” which is neither the theory nor the evidence.

The change-relevant implication is precise: participation in how a change is implemented supports autonomy even when whether is not negotiable. That is a real, available lever in mandated change, and it is the strongest theoretical justification for the participation findings in Part 5.

2.8 Self-efficacy and goals

Self-efficacy — belief in one’s capability to execute a specific behaviour — correlates with work performance at r = 0.38 across 114 studies, 157 effect sizes and 21,616 participants [A]. Bandura’s four sources, in descending strength: mastery experience, vicarious experience, verbal persuasion, physiological state. Note the ordering: watching a credible peer succeed beats being told you can do it, and doing it yourself beats both. This is why behaviour modelling training works and pep talks don’t.

Goal setting theory (Locke & Latham) is among the most replicated findings in organizational psychology: specific, difficult goals outperform vague or easy goals, given ability, commitment and feedback [A]. The evidence spans tens of thousands of participants and time horizons from one minute to 25 years.

The essential corrective is Ordóñez, Schweitzer, Galinsky and Bazerman’s “Goals Gone Wild” (2009), documenting systematic side effects of over-prescribed goals: narrowed focus, increased unethical behaviour, risk distortion, reduced intrinsic motivation, and cultural corrosion [B]. Locke and Latham responded sharply in the same issue. The synthesis to hold: goal setting is real and powerful, and that power is exactly why misapplication causes harm. Stretch goals specifically require both slack resources and a recent history of success; without both they produce escalation of commitment or corner-cutting.

For learning behaviours, learning goals (“master three techniques”) outperform performance goals (“hit 12% improvement”) when the task is complex and the person lacks the requisite strategies. Setting performance goals on a task nobody yet knows how to do is a common and costly design error.

2.9 Feedback: the intervention that can go backwards

Kluger and DeNisi (1996) analysed 607 effect sizes from 23,663 observations. Mean effect: d = 0.41. Share of interventions that reduced performance: over one third (commonly reported as 38%) [A].

Feedback Intervention Theory explains the mechanism: feedback directs attention. Directed at the task and task-learning processes, performance improves. Directed at the self — praise, criticism, normative comparison, anything identity-relevant — performance often falls, because cognitive resources shift from the task to self-regulation.

Rules that follow, and that most performance systems violate:

• Feedback on the task and the process, not on the person.
• Normative comparison is a self-directing intervention and is high-risk.
• Feedback without a clear next action is more likely to harm than help.
• Multi-source (360) feedback deserves scepticism: Jones et al.’s coaching meta-analysis found the use of multisource feedback associated with smaller coaching effects [A].

2.10 Stages of change: a model to retire

The Transtheoretical Model (Prochaska & DiClemente) — precontemplation, contemplation, preparation, action, maintenance — is probably the single most widely taught individual-change model outside psychology departments, and it is in poor evidential health.

Robert West’s 2005 Addiction editorial, “Time for a change: putting the Transtheoretical (Stages of Change) Model to rest,” argued the problems are serious enough that it has held back the field and should be discarded [B]. His specific charges:

• Stage boundaries are arbitrary. Why six months for precontemplation and thirty days for preparation? The cut-offs have no theoretical or empirical justification.
• Transitions are not sequential. People skip stages; many smokers quit abruptly with no contemplation or preparation phase at all.
• It over-weights conscious rational deliberation and neglects automatic and affective processes — which, given everything in 2.1–2.3, is a serious omission.

The decisive test is whether stage-matched interventions outperform the same intervention unmatched. A Cochrane review covering 41 stage-based trials and more than 33,000 smokers found that in the four trials directly comparing the same intervention in standard versus stage-based form, the stage-based version was neither more nor less effective [A].

Verdict: the model’s descriptive vocabulary (“they’re not ready yet”) is useful shorthand. Its prescriptive claim — that you should tailor intervention to assessed stage — is not supported. If you use it, use it as language, not as a targeting algorithm.

2.11 Motivational interviewing

Worth knowing because it is the best-evidenced conversational method for individual change, and because it directly contradicts the standard change-management instinct to persuade.

Lundahl et al.’s 2010 meta-analysis of 119 studies found MI produced statistically significant, durable results in the small effect range against weak comparison groups (average g = 0.28) and non-significant results against specific active treatments (g = 0.09) [A]. Moderators included feedback, delivery time, manualisation, delivery mode and ethnicity.

The honest reading: MI is better than doing nothing or than generic advice, and not demonstrably better than other structured active approaches. Its transferable core — eliciting the person’s own change talk rather than supplying arguments, and treating ambivalence as normal rather than as resistance — is consistent with SDT and with the resistance-reframing literature in Part 5. Its evidence base outside clinical settings is thin; workplace application is a reasonable extrapolation, not an established finding [C].

2.12 Individual level — evidence summary

Claim	Grade	Note
Habits are cue–context associations; context change disrupts them	B	Wood & Neal; Verplanken
Median 66 days to automaticity, range 18–254	B	Lally 2010, small volunteer sample
21 days to form a habit	X	Maltz 1960, not a study
Cue disruption > environmental re-engineering > vigilant monitoring for breaking habits	B	Wood & Neal
Life/work transitions open a behaviour-change window	B	Habit discontinuity hypothesis
Implementation intentions raise goal attainment (d≈0.65)	A	Gollwitzer & Sheeran 2006; 642-test update 2024
Ego depletion / limited willpower as originally formulated	X	d=0.06 in multi-lab preregistered replication
Self-efficacy predicts work performance (r=0.38)	A	Stajkovic & Luthans 1998
Specific difficult goals beat vague goals	A	Locke & Latham
Over-prescribed goals cause systematic side effects	B	Ordóñez et al. 2009
Feedback averages d=0.41 but harms in >1/3 of cases	A	Kluger & DeNisi 1996
Stage-matched interventions outperform unmatched	X	Cochrane: no difference in direct comparisons
Motivational interviewing beats no-treatment, not other active methods	A	Lundahl et al. 2010
Nudges reliably change behaviour at scale	C	Collapses toward zero after publication-bias correction
Defaults and friction changes work	B	The durable residue of the nudge programme

---

Part 3 — Learning that changes behaviour

Separated from Part 2 because the L&D literature has its own evidence base, its own folklore, and its own commercial pathologies, and because “we ran training” remains the single most common organizational response to a behaviour problem.

3.1 What is well supported

Spacing / distributed practice [A]. Cepeda et al. synthesised 839 assessments across 317 experiments in 184 articles. Distributing practice beats massing, and — the finding practitioners consistently miss — the optimal inter-study interval grows with the retention interval you care about. If you need recall in a year, the gaps should be weeks, not days. Large-scale workplace datasets covering 10,000+ employees in pharmaceutical, education and retail settings replicate the effect outside the lab.

Retrieval practice [A]. Testing yourself outperforms restudying, and combines multiplicatively with spacing. Systematic review evidence in health professions education supports both, separately and together.

Behaviour modelling training [A]. Taylor, Russ-Eft & Chan: 117 studies, mean population effect slightly above 1.0 SD against control groups for interpersonal knowledge and skill. The moderator pattern is the important part — effects were largest for learning outcomes, smaller for job behaviour, smaller again for results; and while declarative knowledge effects decayed over time, effects on skills and job behaviour remained stable or increased. Structure: model the behaviour, provide practice, provide feedback, plan for transfer.

Worked examples, faded as competence grows [A]; interleaving for related skill categories [A].

Leadership training, when properly designed [A]. Lacerenza et al., 335 studies: δ = 0.63 reactions, 0.73 learning, 0.82 transfer, 0.72 results. Note that transfer exceeds learning — unusual, and a signal that the well-designed programmes in this set were doing work-environment intervention, not just instruction. Fifteen significant moderators; the ones that matter most are needs analysis, spaced rather than single-session delivery, multiple delivery methods, practice with feedback, and face-to-face or on-site delivery.

Gamification, modestly and conditionally [A]. Sailer & Homner’s meta-analysis found significant effects on cognitive (g = 0.49), motivational (g = 0.36) and behavioural learning outcomes (g = 0.25). Critically, the moderator analysis found game fiction and social interaction were significant moderators of behavioural outcomes — reward-and-status mechanics alone were far less effective than designs with challenge, meaningful goals and narrative. Most corporate gamification is points and leaderboards, which is the part that doesn’t work.

3.2 What is not supported

Learning styles [X]. Pashler et al. found ample evidence that learners have preferences and no adequate evidence that matching instruction to those preferences improves learning. Kirschner’s “stop propagating the learning styles myth” remains correct. A 2024 meta-analysis of the matching hypothesis revisited the question and did not rescue it. Surveys continue to find majority belief among educators and L&D professionals — which makes accurate knowledge of this a genuine differentiator, though one to deploy gently, since many people have built practice on it.

The learning pyramid / “we remember 10% of what we read, 90% of what we do” [X]. Fabricated percentages with no traceable source study.

Andragogy as a validated theory [C/X]. Knowles’ assumptions were derived from experience, observation and theoretical influence rather than empirical research. Reviews find the evidence for andragogical approaches weak and inconclusive, with andragogical methods sometimes performing no better and occasionally worse than traditional approaches. Knowles himself eventually described it as a set of assumptions or a conceptual framework rather than a theory. The practical principles (relevance, experience, self-direction, problem-centred) are reasonable design heuristics; the claim that adults learn by fundamentally different mechanisms is not supported.

Neuro-branded learning products, left/right brain learners, VAK [X].

3.3 Contested: deliberate practice

Macnamara, Hambrick and Oswald’s 2014 meta-analysis found deliberate practice explained 26% of performance variance in games, 21% in music, 18% in sports, 4% in education and less than 1% in professions [A for the analysis].

Ericsson’s objection is substantive rather than defensive: many included studies did not measure deliberate practice as he defined it in 1993, using general practice measures instead, which would attenuate the estimate.

The defensible reading: structured, feedback-rich, effortful practice matters greatly for skill acquisition; the claim that it is the dominant determinant of expertise is unsupported, particularly in professional domains where tasks are ill-defined and feedback is delayed and noisy. For workplace application the practical instruction survives regardless of who wins the argument: practice with immediate corrective feedback on well-defined sub-skills, and be sceptical of anyone promising expertise through hours logged.

3.4 70-20-10

Originates with Morgan McCall, Michael Lombardo and Robert Eichinger at the Center for Creative Leadership in the 1980s, popularised in The Career Architect Development Planner (1996). Roughly 200 successful executives were asked to retrospectively self-report the sources of their development.

Alan Clardy’s 2018 review in Human Resource Development Review examined five literature traditions that appear to converge on the 70% claim and concluded the evidential basis is weak and that HRD policies relying on it are likely to be misleading [B].

The design problems, worth being able to state precisely because you will be in rooms where this is quoted as fact:

• Retrospective self-report of learning sources, which people are demonstrably poor at.
• Sample of already-successful executives — survivorship bias with no comparison group.
• No measurement of learning outcomes, only attributions.
• Descriptive summary repurposed as a prescriptive budget allocation.

Keep: the underlying observation that most capability develops through work rather than in classrooms, which is consistent with the transfer literature. Drop: the ratio, the budget-allocation use, and any citation of it as research.

3.5 The transfer problem

This is where learning meets organizational reality, and where most L&D value is destroyed.

Baldwin and Ford’s 1988 model remains the organising frame: transfer is a function of trainee characteristics, training design and work environment, operating through generalisation and maintenance.

Blume, Ford, Baldwin and Huang’s 2010 meta-analysis of 89 studies is the key quantitative statement [A]:

• Positive relationships between transfer and cognitive ability, conscientiousness, motivation, and a supportive work environment.
• Predictors matter more for open skills (leadership, interpersonal, judgement) than for closed skills (software, procedures). Crucial design distinction: closed-skill training can succeed largely on instructional quality; open-skill training cannot succeed without work-environment intervention.
• Transfer relationships are systematically inflated when predictor and outcome come from the same source in the same context — meaning most published training evaluation overstates its case.

On the environment: meta-analytic SEM across 32 studies and 5,487 participants found a four-factor social-support model (supervisor support, peer support, supervisor sanctions, feedback/coaching) accounting for 32% of transfer variance, with peer support the strongest predictor of motivation to transfer and feedback/coaching the strongest predictor of transfer [A]. Earlier estimates put supervisor support (ρ ≈ .31) above peer support (ρ ≈ .14); the estimates disagree in magnitude but agree in direction — the social environment after training matters more than the training.

The single most important operational implication in this report: if you can only intervene in one place, intervene in the weeks after the learning event, in the team, through the manager and peers — not in the design of the event.

Diagnostic instrument. Holton and Bates’ Learning Transfer System Inventory (LTSI) is a 16-factor validated instrument covering eleven programme-specific and five general factors, with validation across 5,990 participants in 17 countries and translated versions replicating the factor structure [A]. Seven factors show significant correlations with transfer, including learner readiness, motivation to transfer, transfer design, opportunity to use, and the two expectation factors. It is the best available off-the-shelf pre-diagnostic for a training-linked change effort, and almost nobody uses it.

3.6 Evaluation: Kirkpatrick and its successors

Kirkpatrick’s four levels (reaction, learning, behaviour, results) remain near-universal despite well-established criticism [A for the critique]:

• Alliger and Janak showed the assumed hierarchy — that each level yields more valuable information than the last — lacks empirical support.
• Bates and others argue it oversimplifies, ignoring individual and organizational moderators.
• The assumed causality between levels is difficult to demonstrate; levels 1 and 2 can be independent.

The practical damage is specific: Level 1 is easy and gets measured, Level 3 is hard and does not, and the field ends up optimising for participant satisfaction — which is uncorrelated with learning and can be negatively correlated with the desirable difficulty that produces it.

Three better options:

LTEM (Thalheimer, 2018) — an eight-tier framework designed as a corrective, distinguishing attendance and activity from knowledge, decision-making competence, task competence, transfer, and effects. Its lower tiers explicitly represent methods inadequate to validate learning success, which is the design feature that makes it useful: it tells you when your measurement is worthless [C for validation; B for design logic].

Success Case Method (Brinkerhoff, 2005) — deliberately samples the most and least successful cases and investigates both in depth, on the assumption that any initiative contains both. Not a substitute for a counterfactual, but excellent for generating mechanism hypotheses and for surfacing the conditions under which the intervention worked [C].

RE-AIM — see Part 9. The most rigorous of the three, and the only one that forces you to report reach, adoption and maintenance rather than effectiveness alone.

Phillips ROI methodology adds a fifth level converting results to monetary value. Treat with caution [C/D]: the isolation step — estimating what share of the improvement is attributable to training — usually relies on participant or manager estimation, which reintroduces exactly the common-method inflation Blume et al. identified. An ROI figure produced this way is a persuasion artefact, not a measurement.

3.7 Learning level — evidence summary

Claim	Grade	Note
Spacing improves retention; optimal gap scales with retention interval	A	Cepeda et al.; workplace replications
Retrieval practice beats restudy, combines with spacing	A
Behaviour modelling >1.0 SD for interpersonal skill	A	Taylor, Russ-Eft & Chan 2005
Well-designed leadership training transfers (δ≈0.82)	A	Lacerenza et al. 2017
Gamification helps, but via challenge/narrative/social not points	A	Sailer & Homner 2020
Learning styles matching improves outcomes	X	Pashler et al.; repeatedly falsified
Learning pyramid percentages	X	Fabricated
Andragogy is an empirically validated theory of adult learning	C/X	Knowles’ own later position was weaker
Deliberate practice dominates expertise	C	<1% variance in professions; contested
70-20-10 as an evidence-based ratio	X	Clardy 2018
Work environment dominates design for transfer	A	Blume et al. 2010
Peer support strongest predictor of motivation to transfer	A	Meta-analytic SEM, 32 studies
Kirkpatrick’s hierarchy and causality assumptions	X	Alliger & Janak; Bates
LTSI is a validated 16-factor transfer diagnostic	A	5,990 participants, 17 countries
ROI figures from participant-estimated isolation	D	Persuasion artefact

---

Part 4 — Groups and teams

4.1 Why the team is the real unit of change

The individual literature tells you how a person changes. The organizational literature tells you how a programme is run. Neither describes the level at which change succeeds or fails: the working group of five to fifteen people whose shared expectations govern what its members actually do on Tuesday morning.

Four independent lines converge here:

• Transfer research. The post-training environment dominates transfer outcomes [A]. That environment is the team.
• Diffusion research. Behaviour requires redundant social reinforcement and propagates through dense local clusters [B]. Those clusters are teams.
• Implementation research. CFIR’s inner-setting domain — culture, implementation climate, readiness, networks and communications — consistently discriminates high- from low-performing sites [B].
• Knowledge transfer research. Szulanski found the “arduous relationship” between source and recipient among the top barriers to internal best-practice transfer [B]. Relationships are team-level.

Yet team-level change remains the least developed of the three literatures. There is far more written about running an enterprise transformation than about changing the behaviour of one team — which is the thing that would need to happen eighty times over for the transformation to be real.

4.2 Psychological safety

Edmondson’s construct — a shared belief that the team is safe for interpersonal risk-taking — is the best-evidenced team-level variable in the change literature.

Frazier et al.’s 2017 meta-analysis aggregated 136 independent samples covering over 22,000 individuals and nearly 5,000 groups, documenting consistent associations with voice behaviour, learning behaviour, creativity and performance [A]. It also found transformational leadership’s effect on team performance runs substantially through psychological safety — a useful reframing of leadership from trait to mechanism.

On producing it — an update worth having. For most of the construct’s history the honest position was that the evidence for what psychological safety predicts far exceeded the evidence for how to produce it, with identified antecedents (supportive leader behaviour, relationship networks, role clarity, work design) established only correlationally. That has changed. Castro, Englmaier and Guadalupe ran a field randomised controlled trial across roughly 7,000 employees in 1,000 teams in a large organization, using the content and frequency of manager one-to-one meetings as the conduit for changing leader behaviour. The intervention itself was strikingly light — a small number of emails and reminders with PDF attachments that managers were free to act on [B]. Separately, a cluster RCT has tested a group-based psychological safety intervention for preventing workplace bullying and sustaining engagement [C], and the CIPD published a psychological safety evidence review in February 2024 [B].

The practical significance is large and specific: the highest-evidenced route to team psychological safety runs through the structure and content of the manager’s routine one-to-ones, not through workshops or awareness campaigns. That is a routine modification in Feldman and Pentland’s sense, delivered at near-zero cost, and it is a directly sellable intervention.

Three cautions remain:

Frequently misdefined as comfort. Edmondson’s own framing pairs safety with accountability; high safety with low standards produces the comfort zone, not the learning zone. Programmes that raise safety without raising standards produce pleasant teams that don’t improve.

Usually measured badly. It is a shared group-level belief. Averaging individual survey scores without checking within-group agreement is standard practice and technically invalid — you are measuring individual perceptions, not team climate. If you measure it, report agreement statistics.

4.3 How behaviour spreads: complex contagion

The default organizational mental model is epidemic: find influencers, seed the message, watch it spread. That model comes from information diffusion and is wrong for behaviour.

Centola’s experimental work — 1,528 participants randomly assigned to artificially constructed clustered or random online networks, tracking adoption of a health behaviour — found behaviour spread farther and faster in clustered networks than in random networks with more long-range weak ties, because clustering delivers social reinforcement: multiple independent contacts exhibiting the behaviour [B].

This is complex contagion: adoption requires repeated, redundant exposure from several sources, unlike simple contagion where one exposure suffices.

Practical implications, and they are substantial:

• Do not start with the most connected individual. Start with a dense cluster where members can see each other adopting.
• Redundancy is a feature. Hearing it from three peers differs qualitatively from hearing it from one leader three times.
• Weak ties spread awareness; strong clustered ties spread behaviour. Communication strategy and adoption strategy have opposite network requirements.
• Pilot whole intact teams, not scattered volunteers. The scattered-champion model is the most common design and, on this evidence, close to the worst.

Caution: these are artificial online networks with a single simple behaviour. Later work — including 2025 causal-evidence studies and modelling of the reach-versus-reinforcement trade-off — suggests the picture is more conditional than “clustered beats random” implies [C]. The mechanism (social reinforcement) is well supported; exact network prescriptions are not.

4.4 Norms

Cialdini’s focus theory distinguishes descriptive norms (what people do) from injunctive norms (what people approve of). Norms interventions work by correcting misperception.

Honest evidence position: real effects in some domains (littering, tax compliance, energy use), disappointing in others [B/C]. Two failure modes matter organizationally:

The boomerang effect. Telling people they are already better than the norm moves them toward it — degrading your best performers. Cialdini’s fix, pairing descriptive information with an injunctive signal of approval, is supported but rarely implemented.

Norms must be true and locally credible. “87% of colleagues complete this by Friday” fails if everyone can see their own team doesn’t. Organization-wide norm statements are usually not credible at team level, which is the level that matters.

The under-used application is pluralistic ignorance: situations where most individuals privately dislike a practice but each believes others endorse it. Extremely common in meeting culture, email response expectations, safety shortcuts and working hours — and one of the few places where simply surfacing accurate information about the group’s actual preferences changes behaviour quickly and cheaply.

4.5 Routines: the real unit of team behaviour

Feldman and Pentland’s reconceptualisation is the most underrated theoretical tool in this literature.

They split a routine into two aspects [B]:

• The ostensive aspect: the abstract pattern — rules, the idea of the routine, what people would describe if asked how it works.
• The performative aspect: specific actions by specific people at specific times that actually enact it.

The ostensive guides and legitimises performances; performances create, maintain and modify the ostensive. The relationship between them is a continuous variation–selection–retention loop, which is why routines are simultaneously a source of stability and of change.

Why this matters more than it appears: almost every organizational intervention operates exclusively on the ostensive — new process documents, policies, operating models, training. The performative is what people actually do, and it drifts continuously regardless. A change is embedded only when the performative has shifted; it is merely documented when only the ostensive has.

This yields a diagnostic question that cuts through most change theatre: “Show me the last three times this actually happened.” If nobody can, the change exists only ostensively.

It also explains why the most effective embedding mechanisms are mundane: changing the agenda template, changing what the standing meeting reviews, changing what the handover form requires, changing what triggers an escalation. These alter performance directly rather than describing it.

4.6 Interventions at team level: what has evidence

Coaching [A]. Jones, Woods & Guillaume (2016): positive effects overall δ = 0.36; skill-based 0.28; affective 0.51; individual-level results 1.24. Notable moderators: internal coaches outperformed external coaches, and use of multisource feedback reduced effects. Format (face-to-face, blended, e-coaching) and duration (number of sessions, longevity) showed no significant moderation — commercially inconvenient and worth knowing. The high results figure rests on fewer studies; treat cautiously.

Behaviour modelling training delivered to intact teams combines the strongest method with the strongest transfer environment [A].

Team building [B]. Neuman, Edwards & Raju’s meta-analysis of 126 OD intervention studies found team building and laboratory training most effective for changing attitudes, and — the more important finding — that multi-faceted interventions outperformed single-technique interventions. That result recurs across the literature and is the most durable design principle in it: bundles beat single levers.

Communities of practice [C]. Wenger’s construct is theoretically rich and empirically loose. Reviews find the concept has diverged so widely that structure and function vary enormously, making effectiveness hard to evaluate; the definitional inconsistency is itself the main finding. Storberg-Walker’s theory-building critique found most of Wenger’s analytical elements failed tests of definitional clarity and level of analysis. Conditions associated with functioning CoPs — perceived legitimacy of the knowledge, member integration, clear shared purpose, trust, reciprocity — are sensible but not causally established. Treat CoPs as a plausible structure, not an evidence-based intervention.

Action learning (Revans) [C]. Long practitioner history, thin comparative evidence. Its mechanism — real problem, small group, questioning insight, action, reflection — is consistent with everything else here. Evidence for it as a distinct superior method is absent.

Positive deviance [B/C]. Developed by Jerry and Monique Sternin with Richard Pascale; first applied at scale to child malnutrition in Vietnam through Save the Children in the 1990s, since applied in over 60 countries. Mechanism: identify individuals achieving better outcomes with the same resources, have the community discover and codify those behaviours themselves, spread through practice rather than instruction. Documented mechanisms of change: social mobilisation, information gathering, and behaviour change through community adoption.

Theoretically it satisfies almost every mechanism in Part 12 at once — in-group social proof, participation, self-discovery supporting autonomy, locally feasible solutions by construction, and diffusion through dense networks. Evidence is largely case-based rather than comparative, and it is vulnerable to selection effects. The Cambridge Elements volume is the most rigorous available treatment.

4.7 The manager

Across the literature the line manager is the highest-leverage single actor at team level, for structural rather than heroic reasons: the manager controls opportunity to perform, allocates time, sets the agenda, notices, and reinforces.

What the evidence supports the manager doing:

• Pre-brief — a conversation before the intervention about what will be used and where. Motivation to transfer predicts transfer and is partly set beforehand.
• Explicit opportunity to perform — assigning the work in which the new behaviour is used. The most frequently absent mechanism.
• Task-focused feedback — process not person.
• Peer structure — pairing, review, shared practice, since peer support is the strongest predictor of motivation to transfer.
• Routine modification — changing the agenda, template or checklist so the behaviour has a performative home.

The common failure is asking managers to be supportive — an attitude — rather than to perform five specific actions. Specify the manager’s behaviours with the same rigour you specify the target population’s.

4.8 A caution on team instruments

Team change work is saturated with psychometric and stage models that carry more confidence than evidence. Because these are ubiquitous in the rooms you will work in, they warrant explicit treatment.

Tuckman’s forming–storming–norming–performing [C]. Derived in 1965 from a review of 50 articles, categorised by setting into therapy-group studies, T-group (sensitivity training) studies, and natural- or laboratory-group studies — with workplace teams essentially absent. Tuckman himself wrote that this literature “cannot be considered truly representative of small-group developmental processes,” and his own later review noted only about half of studies observed a distinct conflict stage; some groups moved from forming straight to structured working. Groups skip stages, regress and oscillate. The fifth stage (adjourning) was added by Tuckman and Jensen in 1977. Verdict: useful vocabulary for normalising early team friction; not a reliable developmental sequence and not a diagnostic.

MBTI [A against]. Test–retest classification instability is severe — reported figures suggest a large proportion of people receive a different type on retest within weeks — and correlation with job performance is close to zero. The Society for Industrial and Organizational Psychology’s position is that type assessments of this kind should not be used for employee selection. Verdict: acceptable as a conversation-opener if everyone understands it is a conversation-opener. Not acceptable as a basis for selection, team composition, or development decisions. If you need a personality instrument, the Big Five family has far better psychometrics.

Belbin, DiSC and similar [C/D]. Weaker published psychometric evidence than MBTI in most cases, similar commercial ubiquity. Same verdict, more emphatically.

SCARF (Rock) [C]. Status, certainty, autonomy, relatedness, fairness. The individual neural mechanisms behind each domain have support in social neuroscience; the specific five-domain grouping is a practitioner synthesis rather than an empirically derived classification, and the neuroscience framing adds presentational authority rather than predictive power. Verdict: a decent checklist of social threats, oversold as neuroscience. Notice that four of its five domains are recoverable from SDT and organizational justice research without the brain imagery.

Emotional intelligence [B, mixed]. Mattingly and Kraiger’s meta-analysis of 58 studies found a moderate positive effect of formal training on EI scores, robust across gender and measure type, though larger for published studies than dissertations — a publication-bias signal. Separately, the construct faces persistent criticism over overlap with personality and limited incremental validity, with some evidence that trait EI does explain small but significant incremental variance. Verdict: EI training can move EI scores; whether that moves work outcomes beyond what conscientiousness and agreeableness already predict remains contested.

Grit [A against]. Credé, Tynan and Harms: 584 effect sizes, 88 samples, 66,807 individuals, correlation with conscientiousness ρ = .84 — above the conventional threshold for construct redundancy. Perseverance-of-effort has stronger criterion validity than consistency-of-interest and retains some variance beyond conscientiousness, but the authors concluded grit interventions are likely to have weak effects and the construct’s validity is in question. Verdict: a jangle fallacy. You are measuring conscientiousness.

Growth mindset [A, contested]. Sisk et al. (2018), Psychological Science: two meta-analyses — 273 studies / 365,000+ participants on the mindset–achievement relationship, and 43 studies / 57,000+ on interventions; overall effects weak in both. Macnamara and Burgoyne (2023), Psychological Bulletin: a systematic review and meta-analysis of the intervention literature finding effects non-significant once study-design quality and reporting bias were accounted for, concluding apparent effects are “likely attributable to inadequate study design, reporting flaws, and bias.” (Reported study counts differ between sources depending on which outcome subset is analysed; treat any single figure with care.) Tipton and colleagues published a methodological rebuttal arguing that the traditional average-effect approach underestimates genuine heterogeneity across contexts, and that meta-analyses of this kind should follow best practices for examining heterogeneity rather than reporting a pooled mean. Verdict: the strong popular claim is not supported; a weak, context-dependent effect for specific populations under specific conditions remains arguable. Do not build a capability strategy on it — and note that this is a live methodological dispute, not a settled debunking.

4.9 Team level — evidence summary

Claim	Grade	Note
Psychological safety predicts learning, voice, creativity, performance	A	Frazier et al. 2017
Psychological safety can be raised by restructuring manager one-to-ones	B	Field RCT, ~7,000 employees / 1,000 teams
Behaviour spreads via complex contagion needing redundant exposure	B	Centola
Clustered seeding beats influencer seeding for behaviour	C	Directionally supported, context-conditional
Norms interventions can boomerang on high performers	B	Cialdini; injunctive pairing is the fix
Routines have ostensive and performative aspects	B	Feldman & Pentland
Workplace coaching δ≈0.36	A	Jones et al. 2016
Internal coaches outperform external	B	Jones et al. 2016
Multi-faceted interventions beat single techniques	B	Neuman et al. 1989; recurs widely
Tuckman’s stages are a reliable developmental sequence	C	~50% of studies show no distinct storming
MBTI is suitable for selection or development decisions	X	SIOP position; near-zero performance correlation
Grit is distinct from conscientiousness	X	ρ = .84 (Credé et al. 2017)
Growth mindset interventions raise achievement at scale	C	Null in highest-quality subset; contested
Communities of practice are an evidence-based intervention	C	Definitional chaos; weak evaluation

---

Part 5 — Organizations

5.1 The canon, and what each is actually worth

Lewin: force field analysis and “unfreeze–change–refreeze”

Kurt Lewin (1890–1947) is the intellectual founder of the field: field theory, force field analysis, group dynamics, action research, and the finding that group decision changes behaviour more durably than individual persuasion. All of that is real and valuable.

The three-step model attributed to him is largely not his. Cummings, Bridgman and Brown’s 2016 Human Relations paper shows the “changing as three steps” model took its familiar form after Lewin’s death, principally through Schein and French & Bell [A]. Lewin’s 1947 text contains a subheading referring to unfreezing, moving and freezing of group standards — a passing analytic remark within a much richer theory, not a change methodology — and he never wrote “refreezing.”

Why this matters beyond pedantry: the misattribution converted a dynamic, field-theoretic account of behaviour as the resultant of opposing forces in quasi-stationary equilibrium into a static three-box process. Lewin’s actual contribution — that behaviour is held in place by a balance of driving and restraining forces, and that reducing restraining forces is generally more effective than increasing driving forces — is more useful than the model attributed to him, and it anticipates the modern friction and sludge literature by seventy years.

Verdict: the man is essential; the model is an artefact. Force field analysis remains one of the better diagnostic tools available, precisely because it directs attention to what to remove.

Kotter’s eight steps

Establish urgency; form a guiding coalition; create a vision; communicate the vision; empower action; generate short-term wins; consolidate; anchor in the culture.

Appelbaum, Habashy, Malo and Shafiq’s 2012 review in Journal of Management Development is the key assessment. Reviewing fifteen years of literature step by step, they found support for most individual steps but no formal studies covering the entire spectrum and structure of the model — the sequence is not empirically established [A].

Read the finding carefully, because it is misreported in both directions. Appelbaum et al.’s own conclusion is favourable: they found no evidence against the model and judged it a recommendable reference. The sharper critical reading — that much of the supporting material traces back to Kotter’s own work, so in effect Kotter validated Kotter — is a fair inference from what the review found, but it is secondary commentary’s framing, not the authors’ verdict. State both when you make the point; overstating the debunk is its own credibility risk.

A further critique deserves attention: recent work in Journal of Change Management examines “Kotter’s contradiction” — that vilifying the status quo to manufacture urgency undermines the identity and self-worth of the people whose cooperation the change requires, and can derail the effort it was meant to accelerate [C]. This connects directly to the Oreg findings on recipient reactions below.

Honest position: eight sensible propositions, each with some independent support, assembled into a sequence that has never been tested as a sequence — and with a first step that is actively risky as written.

ADKAR and the Prosci methodology

ADKAR — Awareness, Desire, Knowledge, Ability, Reinforcement — is the best of the commercial models, for specific reasons. It is the only mainstream one that locates change at the individual level and treats organizational change as the aggregate of individual transitions. It correctly separates Knowledge (knowing how) from Ability (being able to do it), which is exactly the distinction the transfer literature says is decisive. And it ends on Reinforcement, which the sustainment literature says is decisive.

Prosci’s benchmarking is the largest practitioner dataset available: 25 years, 10,800+ contributors. Its two most-cited findings are that active and visible executive sponsorship has ranked the top contributor to success in every edition, and that initiatives with excellent change management are around seven times more likely to meet objectives than those with poor change management [D].

Handle those numbers carefully. They come from self-report by change practitioners retrospectively rating both the quality of change management and the success of the project — a design in which halo effects and common-method variance are close to unavoidable. The direction is plausible and consistent with the academic literature. The magnitude is not credible as a causal estimate.

Verdict: the most useful commercial framework, at individual level, with genuinely large but methodologically weak supporting data. The sponsorship finding is the most robust element because it replicates independently.

Bridges: transitions

Bridges separated change (external, situational, fast) from transition (internal, psychological, slow), with three phases: ending/loss, the neutral zone, new beginning. No meaningful empirical validation [C], but conceptually valuable and consistent with identity and adjustment literatures. Its practical contribution is real: it legitimises the observation that a change complete on the org chart on Monday is not complete in people’s working lives for months, and that the neutral zone holds both the anxiety and the creativity.

Note what Bridges is not: it is not the Kübler-Ross curve, and conflating them (as much practitioner material does) imports an unevidenced grief-stage sequence into a defensible transition account. See Part 15.

Beer & Nohria: Theory E and Theory O

Economic-value change (top-down, structural, fast, hard) versus organizational-capability change (participative, cultural, slow, soft). Their argument is that the two are usually pursued in isolation and that combining them, in sequence with the tension deliberately held, produces better outcomes [C]. Not tested comparatively, but the typology is analytically sharp and names a genuine strategic choice most programmes make implicitly.

Burke-Litwin

Twelve-variable causal model distinguishing transformational factors (external environment, mission and strategy, leadership, culture) from transactional factors (structure, systems, management practices, work-unit climate, task requirements, motivation, individual needs and values), with defined causal paths [C].

Partially tested and rarely used, which is a shame: it is the only mainstream model specifying causal direction between organizational variables, and therefore the only one making falsifiable claims. Its core assertion — that transformational change requires movement in the transformational variables, and that tinkering with transactional variables produces only incremental change — is a useful diagnostic discipline.

McKinsey 7S

Strategy, structure, systems, shared values, skills, style, staff. A [C/D] checklist with no causal specification and no tested predictions. Virtue: completeness — it stops people changing structure alone. Vice: it implies alignment of seven things is achievable and stable. Use as a coverage check, nothing more.

5.2 The 70% failure myth

Origin: Hammer and Champy, Reengineering the Corporation (1993), observed that “as many as 50 to 70 percent of the organizations that undertake a reengineering effort do not achieve the dramatic results they intended.” That is a descriptive, hedged remark about one intervention type against its own ambitious success criterion.

What happened next: the hedges dropped, “reengineering” became “change,” “do not achieve the dramatic results intended” became “fail,” and the claim became universal. Hughes’ 2011 paper examined the five most prominent published sources — Hammer & Champy, Beer & Nohria, Bain, McKinsey and Kotter — and found each either stated the figure without evidence or cited another source that stated it without evidence [A]. There is no study.

Why it persists: it is commercially useful to everyone selling change services. Consultancies published the figure, then cited their own publications as evidence. Part 7 explains the deeper institutional mechanism.

How to use this professionally: when someone quotes it, the useful response is not aggressive debunking but a question — what would “failure” mean for this initiative, and how would we measure it? That is the question that should have been asked anyway, and it reframes the conversation productively.

Note also that the myth’s replacement is not “change usually succeeds.” We do not know the base rate, because “change initiative” is not a well-defined population and “success” is not consistently measured. The honest position is that the base rate is unknown and probably varies enormously by change type.

5.3 What the evidence actually supports: Stouten, Rousseau & De Cremer

The 2018 Academy of Management Annals review is the best single source in the organizational change literature. It derives ten evidence-based steps [A]. Paraphrased, with commentary:

1. Diagnose the problem and its causes before selecting a solution. The most commonly skipped step. Most change programmes begin with a solution.
2. Assess and address readiness for change, including senior support. Readiness is a specific measurable construct, not a mood.
3. Implement evidence-based interventions appropriate to the diagnosed cause.
4. Communicate a clear, credible vision and rationale, including what will not change.
5. Involve and enable participation by those affected.
6. Enable social support and provide the resources the change requires. Under-resourcing is the most common silent cause of failure.
7. Develop the skills and abilities the new behaviour requires.
8. Reinforce and monitor, using feedback and measurement.
9. Institutionalise through structures, systems, roles and routines.
10. Assess and adapt — evaluate against the diagnosis, iterate.

Note what this contains that Kotter’s does not: diagnosis before solution, resourcing, skill development, and adaptation. Note what it lacks: manufactured urgency, and quick wins as an end in themselves.

5.4 How people actually react to change

Oreg, Vakola and Armenakis (2011) reviewed studies from 1948 to 2007, with 79 meeting inclusion criteria, and built a model of change recipients’ reactions as tridimensional attitudes — affective, cognitive and behavioural — with identified antecedents and consequences [A].

Key implications:

• “Resistance” is not one thing. Someone can approve of a change intellectually, feel threatened by it emotionally, and comply behaviourally. Treating these as a single variable produces incoherent interventions.
• Pre-change antecedents matter as much as change characteristics. Trust in management, history of previous changes, job security and personal disposition all shape reactions before the change is announced. An organization with a history of abandoned initiatives is a materially different implementation environment, and no amount of communication quality compensates.
• Perceived fairness — procedural and distributive — is a consistent predictor. How the decision was made often matters more than what was decided.

Ford, Ford and D’Amelio’s 2008 Academy of Management Review paper, “Resistance to Change: The Rest of the Story,” is the sharpest available reframing [C, conceptual]. Their argument has two parts, and the second is the one usually dropped when the paper is cited:

• Prevailing accounts tell a one-sided story favouring change agents, positioning resistance as an irrational, dysfunctional reaction located “over there” in the recipients.
• Change agents contribute to the occurrence of resistance through their own actions and inactions — broken agreements, communication failures, misrepresentation — and resistance can be a resource for change rather than an obstacle to it.

That second claim is genuinely uncomfortable and genuinely useful: it makes “why are they resisting?” a question about the change agent’s conduct as much as the recipient’s disposition. Recipients who object are frequently the ones who have thought hardest about implementation. Participation associates positively with goal achievement and commitment and negatively with resistance [B].

Part 7 adds a further layer: from a critical-management perspective, treating resistance as pathology to be overcome is itself a political move that serves particular interests. You do not have to accept the full theoretical apparatus to notice that the observation is often correct.

5.5 Culture and climate

Culture is where change programmes go to become unfalsifiable. Some discipline:

Schein’s three levels — artefacts, espoused values, basic underlying assumptions — remains the most useful conceptual model with limited direct empirical validation [C]. Tests of it (e.g. Hogan & Coote on culture, innovation and performance) find partial support for the mediation structure it implies [C].

The Competing Values Framework and OCAI are the most widely used measurement instruments and have real validity problems; criterion validity in particular has been reported as weak [C]. Culture typologies produce satisfying quadrant diagrams and unreliable measurement.

The CIPD’s evidence review on organisational culture and climate makes the most practically important distinction in this area [B]: culture (deep, slow, partly unconscious) versus climate (shared perceptions of policies, practices and procedures, and what gets rewarded and supported). Climate is more proximate to behaviour, more measurable, and far more changeable. Their recommendation — target climate rather than culture — is the single most useful piece of advice in the culture literature.

Restated as a working principle: you cannot change culture directly; you change behaviours and structures, and culture is the lagging record of what you changed. Anything sold as a direct culture intervention is either a climate intervention with grandiose branding, or theatre.

5.6 Organizational learning

A distinct and older literature from change management, frequently conflated with it, and worth separating.

Argyris and Schön: single- and double-loop learning [C]. In single-loop learning, error detection connects only to action strategies; governing variables stay unchanged. In double-loop learning, error detection connects to the norms defining effective performance. The most valuable contribution is the concept of organizational defensive routines — any policy, practice or action that prevents embarrassment or threat to the players involved and, simultaneously, prevents learning. Argyris’ Model O-I learning systems are characterised by defensiveness, self-fulfilling prophecies and escalating error, with feedback loops that make assumptions self-reinforcing.

Evidence position: Argyris and Schön provide detailed descriptions of heuristic efforts and partial successes based on direct observation, but no step-by-step method for transitioning from Model I to Model II. The framework is diagnostically powerful and prescriptively vague. Its practical value is naming a specific, recognisable phenomenon: the meeting in which everybody knows the plan will not work and nobody says so, and in which raising it would be experienced as an attack.

March: exploration and exploitation [B]. Exploitation refines existing capability; exploration pursues new possibilities. March’s central argument is that adaptive processes systematically favour exploitation, because its returns are nearer, more certain and more local — meaning organizations reliably under-invest in exploration and can trap themselves in locally optimal, globally obsolete competence. This is the theoretical basis for the ambidexterity literature and the best available explanation for why capable organizations fail at transitions that look obvious in hindsight.

For change practice, March explains something the change literature usually moralises about: the problem is not that people are conservative, it is that the incentive structure of learning itself is biased. Exploration must be protected structurally, not merely encouraged.

Cohen and Levinthal: absorptive capacity [B]. A firm’s ability to recognise the value of external knowledge, assimilate it, and apply it — and, critically, this ability depends on prior related knowledge, not merely on exposure. Zahra and George’s 2002 reconceptualisation splits it into potential absorptive capacity (acquisition, assimilation) and realised absorptive capacity (transformation, exploitation).

The change implication is sharp and under-appreciated: an organization’s ability to adopt a new practice is a function of what it already knows. Bringing in a practice far from existing capability does not fail because of resistance; it fails because there is nothing for it to attach to. This also explains why “best practice” transplants from very different organizations reliably disappoint.

Szulanski: internal stickiness [B]. 271 observations of 122 best-practice transfers across eight companies. Contrary to conventional wisdom blaming motivational factors, the major barriers were knowledge-related: the recipient’s lack of absorptive capacity, causal ambiguity (nobody can fully specify why the practice works), and an arduous relationship between source and recipient.

This is one of the most important findings in the report for practical purposes, because it is a direct empirical rebuttal of the default diagnosis. When a practice does not spread internally, the first hypotheses should be: they don’t have the prior knowledge to absorb it; nobody can articulate why it works; and the two units don’t have a working relationship — not they don’t want to.

Senge and the learning organization [C/X]. The Fifth Discipline remains influential and evidentially thin. Critiques centre on scarcity of empirical validation, absence of an agreed definition of “learning” in the learning-organization literature, and a systems-level emphasis at the expense of individual mechanisms. There is little empirical evidence connecting learning-organization principles to organizational effectiveness. Verdict: an inspiring synthesis; not a validated model. The underlying components — systems thinking, mental models, shared vision, team learning, personal mastery — are each defensible in isolation with varying support, but the package is not evidence.

5.7 Change fatigue and portfolio load

Practitioner data suggests employees now face roughly 10–14 concurrent or annual enterprise change initiatives, up sharply from the mid-2010s, with high overlap and measurable declines in willingness to support new initiatives [D]. The academic construct — change fatigue, change saturation — has supporting work linking cumulative change exposure to exhaustion, cynicism and reduced commitment [B/C].

The framing that survives scrutiny: fatigue is a capacity problem in the change portfolio, not a resilience deficit in individuals. Structural and unpopular implications:

• The absorptive capacity of a team is finite and small. Two significant behavioural changes at once is often already too many.
• Stopping initiatives is a change intervention. Portfolio pruning creates the capacity that makes remaining initiatives viable. Almost no organization has a mechanism for it.
• Sequencing matters more than content. The same three changes delivered serially and simultaneously produce different outcomes.

Note the double meaning of absorptive capacity here: Cohen and Levinthal’s knowledge-based sense and the attentional sense are distinct constructs that happen to point the same direction. Both say the same operational thing — you can only take on what you have the base to take on.

5.8 Emergent and dialogic approaches

Bushe and Marshak’s distinction between diagnostic OD (organizations as systems to be diagnosed and realigned) and dialogic OD (organizations as meaning-making systems changed by changing the conversations that constitute them) is the main theoretical development in OD in two decades [C].

Dialogic OD’s premises: reality and relationships are socially constructed; organizations are meaning-making systems; language matters; changing conversations is necessary for change; groups self-organise; change emerges from dialogue; the future is created, not predicted; the change agent facilitates rather than prescribes.

Assessment: intellectually serious, methodologically difficult to evaluate, and prone in practice to becoming an excuse for unaccountable facilitation. The strongest defensible claim is a contingency one — planned/diagnostic approaches suit changes with a knowable end state and technical content; emergent/dialogic approaches suit changes where the end state is genuinely unknown and the problem is contested. Most real programmes need both, sequenced.

Appreciative Inquiry belongs to this family with the same assessment: strong theory of positive framing, real practitioner enthusiasm, weak comparative evidence [C].

5.9 Improvement methodologies

Lean and Six Sigma [B, quality-limited]. Systematic reviews of reviews in acute care found the overall quality of evidence poor to fair, with predominantly positive results concentrated in process metrics — length of stay, waiting time, turnaround time — rather than clinical outcomes. Reviews of critical failure factors converge on lack of top-management commitment, poor communication, insufficient training and limited resources. Lean and Six Sigma appear better adapted to settings with linear sequences of events.

Verdict: real and useful for process-flow problems with well-defined steps; oversold for outcome transformation; and its failure factors are exactly the change-management factors in Part 5.3, which suggests the methodology is not the active ingredient.

Agile transformation [C]. Systematic reviews find remarkably little rigorous evidence relative to the popularity of the practice: one review of large-scale agile transformations found almost 90% of included papers were experience reports rather than empirical studies. Identified success factors — management support, customising the model, training and coaching, mindset and alignment — again map onto general change factors rather than anything agile-specific.

Verdict: the practices have plausible mechanisms (short feedback cycles, small batches, team autonomy — all consistent with this report). The transformation industry around them has an evidence base dominated by experience reports.

Toyota Kata [C]. See Part 11. Theoretically the best-fitting practitioner framework in the field, empirically case-based.

5.10 Readiness — the construct worth measuring

“Readiness” is used loosely in practice to mean enthusiasm. Weiner’s 2009 theory in Implementation Science gives it a usable structure [B], and it is the best-specified pre-implementation construct available.

Organizational readiness for change is shared determination to implement the change (change commitment) plus shared belief in the collective ability to do so (change efficacy). Both words matter: it is a shared property of a group, not an average of individual attitudes, and it has two components that are frequently confused and are addressed by completely different interventions.

Weiner identifies the determinants of change efficacy as task demands (what will actually be required), resource availability (what we have), and situational factors (everything else bearing on whether we could do it). Note that all three are assessable in advance and none of them are moods.

The diagnostic value is in the two-by-two:

	High commitment	Low commitment
High efficacy	Proceed. Focus on execution and sustainment.	Rationale, participation, fairness, address distributional concerns. Do not add training.
Low efficacy	Resource it, build capability, sequence early wins, reduce scope. Do not add persuasion.	Both problems. Usually means the change is premature or the portfolio is saturated.

The most common error in practice is treating an efficacy problem as a commitment problem — running communications and engagement campaigns at people who are willing but correctly believe they cannot do it with what they have. That is a reliable way to convert willing people into cynical ones, and it feeds the pre-change history effect Oreg et al. identified.

An organizational readiness for implementing change (ORIC) measure has been developed and psychometrically assessed from Weiner’s theory, making this measurable rather than merely conceptual [B].

5.11 Sensemaking

Weick’s account of sensemaking (1995) is the best available explanation of what people are actually doing during a change, and it is different from what change communications assume [C, theoretically driven].

Its core properties: sensemaking is retrospective (we interpret what has happened, not what is announced), grounded in identity (what this means for who I am), social, ongoing, cue-driven, enactive (acting produces the environment we then interpret), and driven by plausibility rather than accuracy — because at the moment of perception it is usually impossible to tell whether a perception will prove accurate.

Three practical consequences that contradict standard practice:

• People make sense of a change by acting in it, not by being told about it. Announcement-then-explanation assumes prospective, accuracy-driven sensemaking. The evidence points the other way: give people something to do early, and interpretation follows.
• Plausibility beats accuracy. A coherent, plausible account will be adopted over an accurate but fragmented one. This is why rumour outruns official communication, and it is an argument for narrative coherence rather than exhaustive detail.
• Identity mediates everything. A change that threatens a person’s account of who they are and what they are good at will be resisted regardless of its logic — which is the mechanism behind the Kotter urgency critique and behind the Ford et al. reframing.

5.12 Pacing: does change accumulate, or does it happen in bursts?

An empirical question with a substantial answer, largely absent from practitioner discussion.

Punctuated equilibrium holds that organizations experience long convergent periods of incremental adjustment, punctuated by short, discontinuous reorientations that reset the direction for the next convergent period (Tushman and Romanelli; Gersick’s 1991 AMR paper compares the pattern across six domains).

Romanelli and Tushman’s 1994 empirical test in the Academy of Management Journal is the finding that matters [B]:

1. A large majority of organizational transformations were accomplished via rapid and discontinuous change across most or all domains of organizational activity.
2. Small changes in strategies, structures and power distributions did not accumulate to produce fundamental transformations.
3. Major environmental change and CEO succession influenced transformations.

Point 2 is uncomfortable and important. Much of the incrementalist advice in change practice — start small, build momentum, let quick wins accumulate into transformation — is not supported by this evidence. Incremental change appears to produce incremental results; fundamental reorientation appears to be a different kind of event.

How to hold this alongside the rest of the report, which is overwhelmingly incrementalist in its recommendations:

• The two are about different outcomes. Behavioural change in a team, and the embedding of a specific practice, are convergent-period activities where incremental methods are exactly right. Fundamental reorientation of an organization’s strategy, structure and power distribution is a different phenomenon, and the evidence says it does not arrive by accumulation.
• Practically: be honest about which one you are being asked to do. A great deal of disappointment in change work comes from applying convergent-period methods to a reorientation problem, or from selling transformation and delivering — quite properly — incremental improvement.
• The CEO-succession finding also connects to the habit discontinuity hypothesis at organizational scale: leadership transition is a context discontinuity, and it is empirically associated with the changes that actually stick.

5.13 Organizational level — evidence summary

Claim	Grade	Note
Lewin authored unfreeze–change–refreeze	X	Cummings, Bridgman & Brown 2016
Reduce restraining forces before adding driving forces	B	Lewin’s actual contribution
Kotter’s 8 steps validated as a sequence	X	Appelbaum et al. 2012
“70% of change fails”	X	Hughes 2011; no source study
Active visible sponsorship is the top success factor	B/D	Prosci consistently; independently plausible
Excellent change management → 7× objective achievement	D	Self-report, common-method variance
Diagnosis before solution improves outcomes	A	Stouten et al. 2018
Participation reduces resistance, raises commitment	B	Consistent direction across studies
Recipient reactions are tridimensional with pre-change antecedents	A	Oreg et al. 2011
Culture can be changed directly by a culture programme	X	Change climate and behaviour; culture lags
Main barriers to internal practice transfer are motivational	X	Szulanski: knowledge-related barriers dominate
Adoption capacity depends on prior related knowledge	B	Cohen & Levinthal; Zahra & George
Organizations systematically under-invest in exploration	B	March 1991
Learning-organization principles improve effectiveness	C	Little empirical support
Lean/Six Sigma improve process metrics	B	Evidence quality poor to fair; outcomes weaker
Agile transformation has a solid empirical base	C	~90% experience reports
Change fatigue is a portfolio capacity problem	B/D	Practitioner data strong, academic support partial

---

Part 6 — Diffusion, spread and scale

The missing literature. Change management asks “how do we make this happen here?” Diffusion asks “why does this spread there and not here?” — a better question, with better answers, and it has been available since 1962.

6.1 Rogers’ attributes of innovations

Rogers identified five attributes predicting rate of adoption [B]:

• Relative advantage — how much better than the alternative, as perceived by the adopter
• Compatibility — consistency with existing values, experiences and needs
• Complexity — difficulty of understanding and use
• Trialability — can it be experimented with on a limited basis?
• Observability — are the results visible to others?

These five explain a substantial share of variance in diffusion rates — reported ranges across product categories run high, though such figures should be read as indicative rather than precise given heterogeneity of measurement.

The practical power is diagnostic. Most organizational “resistance” is an innovation-attribute problem misdiagnosed as an attitude problem. Before concluding that people are resistant, check:

• Does the new way visibly beat the old way from the user’s position, not from the sponsor’s? (Relative advantage is perceptual and role-specific. A process that saves the organization money and costs an individual an extra hour has negative relative advantage for that individual, and no amount of communication fixes that.)
• Does it fit how work is currently done, or does it require simultaneous change to five other things? (Compatibility.)
• Can someone try it on one case without committing? (Trialability. Almost every corporate rollout is all-or-nothing, which destroys this attribute deliberately.)
• Can people see others doing it and see the results? (Observability. Remote and hybrid work has severely degraded this, which is an underrated cause of adoption difficulty since 2020.)

Note the interaction with complex contagion: observability is what makes redundant social reinforcement possible. If the behaviour is invisible, no amount of clustering helps.

6.2 Greenhalgh et al.: diffusion in service organizations

The 2004 Milbank Quarterly systematic review by Greenhalgh, Robert, Macfarlane, Bate and Kyriakidou is the landmark adaptation of diffusion theory to service organizations [A]. It produced a parsimonious evidence-based model, identified knowledge gaps, and — a contribution in its own right — a transferable methodology for systematically reviewing complex, heterogeneous research evidence.

Its central conceptual contribution for practitioners: diffusion in service organizations is not a single process. It spans a continuum from let it happen (unplanned, informal, emergent) through help it happen (negotiated, influenced) to make it happen (planned, managed, directive). Different innovations and different contexts sit at different points, and applying a make-it-happen approach to something that requires let-it-happen conditions is a recognisable and common failure.

It also emphasises that innovations are reinvented in the course of adoption — adopters modify them — and that this is often necessary for fit rather than a deviation to be suppressed. This connects directly to the fidelity-versus-adaptation question in Part 9.

6.3 NASSS: why technology-enabled change fails

Greenhalgh’s 2017 NASSS framework — Non-adoption, Abandonment, and challenges to Scale-up, Spread and Sustainability — is the most relevant recent framework for any change involving technology [B].

Its structure examines complexity across seven domains: the condition or problem; the technology; the value proposition; the adopter system (staff, patients/users, carers); the organization; the wider system; and embedding and adaptation over time. Its central claim is that failure is predicted by complexity in multiple domains simultaneously — simple in all domains means likely success; complicated is manageable; complex in several domains predicts non-adoption or abandonment.

A scoping review of NASSS use found most studies applied it after implementation to explain what happened, rather than during planning to predict — which is a fair criticism of the field’s discipline rather than of the framework [B]. Complexity assessment tools (NASSS-CAT) were developed in 2020 to support prospective use.

For anyone advising on AI adoption in 2026, NASSS is the most directly applicable framework in existence, and essentially nobody in the commercial market is using it. See Part 10.

6.4 Internal transfer: why good practice doesn’t travel

Szulanski’s finding bears restating because it is so directly contrary to default practice. Across 122 best-practice transfers in eight companies, the dominant barriers were [B]:

• Recipient’s lack of absorptive capacity — no prior related knowledge to attach the practice to.
• Causal ambiguity — nobody can fully specify why the practice produces the result, so what gets transferred is a description of the artefact rather than the mechanism. (Compare Michigan in Part 11: the checklist transferred, the social programme did not.)
• Arduous relationship between source and recipient — transfer requires many exchanges, and a difficult relationship makes those exchanges too costly.

Practical consequences:

• Invest in the relationship between source and recipient units, not just in documentation. Secondments, paired visits and joint working do more than playbooks.
• Attack causal ambiguity explicitly. Ask the source unit not “what do you do?” but “why do you think it works, and what would break it?” Positive deviance methodology is essentially a structured attack on causal ambiguity.
• Check absorptive capacity before transferring. If the receiving unit lacks the base knowledge, the first intervention is capability, not rollout.

6.5 Spread — practical synthesis

Combining Rogers, Greenhalgh, Centola and Szulanski gives a coherent spread strategy that is more evidence-based than anything in commercial change practice:

1. Improve the innovation’s attributes before trying to spread it. Increase relative advantage for the adopter, reduce complexity, create trialability, engineer observability. This is design work, and it is almost always cheaper than persuasion.
2. Reduce causal ambiguity. Know and be able to state why it works and what the active ingredients are, so that adaptation preserves them.
3. Seed dense clusters, not scattered champions, so social reinforcement can occur.
4. Build source–recipient relationships as infrastructure.
5. Assess absorptive capacity and build the base where it is missing.
6. Choose your position on the let-it-happen / help-it-happen / make-it-happen continuum deliberately, and match the effort to it.
7. Expect and manage reinvention. Specify which elements are core (must not change) and which are adaptable, before rollout, rather than adjudicating case by case afterwards.

---

Part 7 — Why change theatre persists

Any rigorously critical review has to explain its own subject matter’s pathology. Why do organizations keep buying, and practitioners keep selling, models with no evidence? Three literatures answer this better than the change literature answers anything about itself.

7.1 Institutional theory: legitimacy over efficacy

Meyer and Rowan (1977) [B]. Many formal organizational structures arise as reflections of rationalised institutional rules. These rules function as myths that organizations incorporate to gain legitimacy, resources, stability and survival prospects — not primarily because they improve performance.

The second half of their argument is the important part for us: because attempting to actually control and coordinate activities according to these adopted structures produces conflict and loss of legitimacy, elements of structure become decoupled from activities and from each other. Integration is avoided, implementation is neglected, and inspection and evaluation are ceremonialised.

That is a precise, forty-nine-year-old description of the change programme that has a governance board, a RAG-rated plan, a communications workstream and a benefits-realisation template, and no observable effect on what anybody does on Tuesday. It is not incompetence. It is a functional response to a legitimacy demand.

DiMaggio and Powell (1983) [B]. Three mechanisms drive organizations in a field toward homogeneity:

• Coercive isomorphism — external pressure: regulation, mandates, expectations of funders, parent companies or clients.
• Mimetic isomorphism — under uncertainty, organizations model themselves on others perceived as successful, because copying is a cheap way to manage uncertainty and to justify decisions afterwards.
• Normative isomorphism — professionalisation: shared training, credentials, professional networks and consultancies propagate a common definition of state-of-the-art.

Change management as a field is a near-perfect case of all three. Certification bodies produce normative pressure; consultancies produce mimetic templates; procurement and audit requirements produce coercive pressure. The homogeneity of change methodologies across unrelated organizations is not evidence that the methodology works. It is evidence of a structured field.

Practical consequence: critique alone will never displace a popular model, because the model is not being selected on efficacy. If you want to displace it, you must supply an alternative that satisfies the same legitimacy function — recognisable, nameable, defensible to a board, and auditable — while being better. This is a design constraint on evidence-based practice, not an objection to it.

7.2 Management fashion

Abrahamson’s work documents the phenomenon empirically rather than anecdotally [B]. Management fashions are rapid, bell-shaped swings in the popularity of management techniques: a long dormancy after invention or rediscovery, then a short-lived, roughly symmetric popularity curve. His analysis of the Quality Circles movement showed rapid growth from 1978, reversal from 1982, and return to pre-popularity levels by 1986.

The fashion-setting process involves management-knowledge entrepreneurs — consultancies, business schools, publishers, gurus — continuously redefining collective beliefs about which techniques represent rational progress.

Two implications worth holding:

• Popularity carries no information about efficacy. It carries information about the fashion cycle’s position.
• The cycle is not purely wasteful. Abrahamson and Fairchild’s later work treats fashion as a collective learning process, in which each wave leaves behind some genuinely useful residue even as the branding is discarded. Total Quality Management left statistical process control; the nudge wave will leave friction reduction and defaults; the agile wave will leave short feedback cycles. Look for the residue, not the brand.

7.3 The critical / power perspective

Critical management studies deals primarily with power, control and resistance [C, theoretically driven]. Its core objection to mainstream change management is worth stating fairly, because it is uncomfortable and frequently correct in part:

• Management thought has historically treated workplace resistance as pathological — deviant, harmful, to be overcome. This framing is not neutral: it positions the change agent’s objectives as legitimate by default and the recipient’s as obstacles.
• Ostensibly technical, neutral change methodologies carry political content. Deciding what counts as “the problem,” who is consulted, what is measured and what is unmentionable are exercises of power presented as method.
• Labour process theory adds that many changes marketed as capability or engagement initiatives function as intensification or control mechanisms, and that employees frequently understand this more accurately than the change team does.

You do not have to adopt the theoretical apparatus to extract the practical value:

• Recipients’ scepticism often encodes accurate information about the change’s real distributional effects. Oreg’s finding that history and trust dominate pre-change reactions is the mainstream version of the same observation.
• If a change genuinely benefits the organization and disadvantages a specific group, no communications strategy resolves that, and pretending otherwise destroys credibility. Naming the trade-off honestly is both more ethical and, per the procedural-fairness evidence, more effective.
• A consultant who can hold this perspective without being captured by it is more useful than one who cannot, because they can anticipate objections that the sponsor cannot see.

7.4 What this means for practice

Three operational conclusions:

1. Expect decoupling and design against it. Assume by default that a new structure will be adopted ceremonially and that the performative layer will not move. The counter is Part 4.5 — put the behaviour inside a recurring routine, and measure the routine.

2. Satisfy the legitimacy function deliberately. Give evidence-based practice a name, a diagram, a sequence and a governance artefact. This is not intellectual dishonesty; it is meeting a real institutional requirement so that the substance can travel. CFIR and RE-AIM already look like this, which is part of why they succeeded in healthcare.

3. Locate yourself in the fashion cycle. When engaging with a hot methodology, ask what the residue will be. Sell the residue.

---

Part 8 — Reinforcement, incentives and consequences

Behaviour analysis is the oldest evidence base in the field and the one most neglected by contemporary change practice, largely for reasons of intellectual fashion rather than evidence.

8.1 Organizational behaviour modification

Stajkovic and Luthans’ meta-analysis of O.B.Mod. field research from 1975–95 (Academy of Management Journal, 40(5), 1122–1149) found a significant main effect on task performance of d = 0.51, roughly a 17% performance increase, with a two-level moderator analysis showing a stronger average effect in manufacturing organizations and moderation by the type of contingent intervention [A].

That manufacturing moderator is worth pausing on: it is consistent with the pattern that appears throughout this report — behavioural techniques work best where the target behaviour is observable, repeated and well-defined. It is a boundary condition, not a disqualification, and it tells you where to apply the method and where to expect less.

The method is unglamorous: identify the specific behaviour, measure its baseline frequency, analyse its antecedents and consequences, intervene (usually with feedback, social recognition and contingent consequences), and evaluate. Notably, the components that carry most of the effect in the literature are performance feedback and social recognition, not money.

Why it fell out of fashion: association with behaviourism, discomfort with the language of contingency and reinforcement applied to adults, and a preference in the OD tradition for meaning and dialogue over consequences. None of these are evidential objections. If you strip the vocabulary, “specify the behaviour, measure it, make sure it is noticed and has consequences” is precisely what the sustainment literature independently concludes.

8.2 Financial incentives

Jenkins, Mitra, Gupta and Shaw’s meta-analysis cumulated 39 studies and 47 relationships [A]:

• Corrected correlation with performance quantity: .34 (approximately a 12% increase).
• Corrected correlation with performance quality: .08 (approximately 3%) — effectively nil.
• Setting and theoretical framework moderated the relationship; task type did not.

This is one of the most practically consequential findings in the report and one of the most ignored. Money reliably increases how much people do and does not reliably increase how well they do it. Any incentive scheme attached to a quality, judgement, safety, care or craft outcome is operating outside its evidence base — and, given the goals-gone-wild findings, is actively risking distortion.

8.3 The crowding-out problem

Deci, Koestner and Ryan’s meta-analysis of 128 studies found that engagement-contingent, completion-contingent and performance-contingent rewards significantly undermined free-choice intrinsic motivation (d = −0.40, −0.36 and −0.28 respectively), as did tangible rewards and expected rewards generally [A]. Their analysis also demonstrated that an earlier meta-analysis claiming the undermining effect was minimal was seriously flawed.

Synthesis of 8.1–8.3, which practitioners rarely hold together:

Consequence type	Effect	Best use
Performance feedback (task-directed)	Strong, cheap	Almost always
Social recognition	Strong, cheap	Almost always
Financial incentive	Quantity yes, quality no	Volume tasks with clear metrics
Tangible expected reward	Undermines intrinsic motivation	Avoid where you want durable self-driven behaviour
Coercion / sanction	Effective short-term, corrosive	Genuine non-negotiables only

The practical rule that falls out: use consequences, but make most of them informational and social rather than material. The evidence for noticing is better than the evidence for paying.

8.4 Behaviour-based safety: a cautionary case

BBS applies behaviour analysis to workplace safety — observation, feedback, positive reinforcement of safe behaviour. It is the most widespread industrial application of the principles in 8.1 and a useful test case.

Evidence: a meta-analysis found BBS interventions associated with a statistically significant reduction in accidents and injuries, with a pooled reduction of roughly 39%, and all included studies showing a post-intervention decrease. But the reviewers judged the quality of the underlying studies poor, and cautioned accordingly [B, quality-limited].

The critique is substantive and worth carrying:

• BBS focuses on individual behaviour and can displace attention from systemic and managerial causes — a “victim blaming” dynamic.
• Incentive-linked BBS programmes can drive under-reporting of injuries, which improves the metric while worsening the reality. This is a direct instance of the goals-gone-wild mechanism.
• Observation-based measurement is itself unreliable.

Proponents respond that a well-run BBS process contains no blame, only identification of what needs to change. The fair synthesis: BBS works to the extent that it changes the system that produces at-risk behaviour, and becomes a blame-displacement mechanism to the extent that it does not. That is a generalisable warning for any behaviour-focused intervention deployed in an environment whose structure produces the behaviour.

---

Part 9 — Implementation science and sustainment

The most rigorous body of change knowledge available, developed where failed implementation is measured in mortality, and directly transferable.

9.1 The four core frameworks

CFIR — Consolidated Framework for Implementation Research [A/B]. Damschroder et al. (2009) organised 39 constructs across five domains: intervention characteristics, outer setting, inner setting, characteristics of individuals, and process. Updated in 2022 following systematic user feedback; the 2025 CFIR User Guide provides a five-step method for conducting implementation research with it.

Its practical value is as a structured pre-mortem. You assess each determinant before implementing, rather than discovering them afterwards. Nothing in commercial change practice does this with comparable coverage.

ERIC — Expert Recommendations for Implementing Change [B]. Powell et al. (2015) produced, through a three-phase modified-Delphi process, a compilation of 73 discrete implementation strategies, each named and defined, subsequently organised into nine clusters by hierarchical cluster analysis with importance and feasibility ratings.

This is the missing link in most change practice: a defined vocabulary of what you can actually do, comparable to BCTTv1 at the individual level. The CFIR–ERIC matching tool maps identified barriers to candidate strategies. Note also the honest caveat from the field: reviews of strategy-selection methods find diversity in recommendations, meaning the matching tool is a structured aid to judgement, not an algorithm.

RE-AIM [A/B]. Glasgow, Vogt and Boles (1999): Reach, Effectiveness, Adoption, Implementation, Maintenance. Its value is making visible the four ways an effective intervention still fails — too few people reached, too few units adopting, poor delivery fidelity, no maintenance. Standard corporate evaluation measures effectiveness only, in the units that adopted well, during the period when someone was watching.

NPT — Normalisation Process Theory [B]. May and Finch (2009) explain how practices become routinely embedded through four generative mechanisms:

• Coherence — does the practice make sense to those doing it?
• Cognitive participation — do key actors commit and build a community around it?
• Collective action — does the practical work actually get done?
• Reflexive monitoring — is it appraised, formally and informally, and reconfigured?

NPT is the most useful available theory for the question “why did this not stick?”, and maps directly onto Feldman and Pentland’s performative aspect.

Workflow these enable: assess determinants (CFIR) → select matched strategies (ERIC) → evaluate across five dimensions (RE-AIM) → explain embedding or its absence (NPT). More rigorous than any commercial change methodology, and free.

9.2 Sustainment

Every popular change model treats “make it stick” as a final step. The evidence says it is the largest and least understood part.

Findings from the healthcare sustainability literature [B]:

• Roughly half of evidence-based interventions are not sustained over time.
• Sustainability depends on factors at multiple levels simultaneously — individual, managerial, financial, leadership, organizational, cultural, political, processual, contextual and temporal. No single lever secures it.
• Most research assesses sustainability only at the end of implementation, so the continuous adjustment process that actually produces sustainment goes unobserved. Longitudinal case work on “the process of sustaining” is the emerging corrective.
• Failure to sustain wastes the original investment and raises resistance to the next initiative — the mechanism by which organizations acquire the change-cynical history that Oreg et al. identify as a pre-change antecedent. Failed change is not neutral; it is a debt.

9.3 Fidelity and adaptation

The question implementation science has begun and organizational change has barely touched: when is adapting an intervention beneficial and when is it fatal?

The framing that works: distinguish core components (the active ingredients, which must be preserved) from the adaptable periphery (delivery, sequencing, language, scheduling — which should be adapted for fit). Greenhalgh’s diffusion review makes the same point through reinvention: adopters modify innovations, often necessarily.

The operational discipline that follows:

• Specify core versus adaptable before rollout, not during. If you cannot say which elements are core, you have a causal-ambiguity problem (Szulanski) and should solve that first.
• Log adaptations. Undocumented adaptation is the difference between “we learned something” and “we have no idea what we implemented.”
• Measure fidelity separately from outcomes, so that a null result can be attributed correctly.

9.4 Safety culture and high-reliability organizations

Relevant because safety-critical industries have the most mature behaviour-change practice in the commercial world.

Weick and Sutcliffe’s five principles of collective mindfulness — preoccupation with failure, reluctance to simplify interpretations, sensitivity to operations, commitment to resilience, deference to expertise [B/C] — describe how high-reliability organizations sustain low accident rates under demanding conditions. The evidence position is honest: the principles are well-observed descriptively, and the evidence that implementing them produces reliability is much thinner. A US Department of Veterans Affairs evidence brief on HRO principle implementation found the literature limited and heterogeneous [B].

The most transferable specific finding: in more effective HROs, attending to failure means seeing the reality of danger in a near miss; in less effective ones, it means seeing the reality of safety — “nothing happened, so we’re fine.” That is a behavioural distinction, observable in how near-miss reports are discussed, and it generalises to any domain with rare high-consequence failures.

Hold Perrow’s Normal Accident Theory alongside it: in tightly coupled, interactively complex systems, catastrophic failure may be structurally inevitable rather than culturally preventable. Holding both is more honest than adopting either.

9.5 De-implementation: the discipline of stopping

The literature on starting things is enormous; the literature on stopping them is small, recent, and disproportionately useful — particularly given the change-fatigue findings in Part 5.7, where stopping is the intervention that creates the capacity for everything else.

The construct. De-implementation is the deliberate process of reducing or stopping practices that are ineffective, harmful or low-value. It is not the absence of implementation; it is its own activity with its own barriers, and it has developed principally in healthcare around low-value care.

Frameworks [B]. Norton and colleagues’ de-implementation framework sets out factors to consider when planning a de-implementation initiative. The GROW framework — Get Ready, Remove/Reduce/Replace, Optimize, Watch — provides a systems-level sequence for phasing out an ineffective practice and sustaining an evidence-based replacement. Note the Replace step: the evidence suggests stopping is easier when something takes the stopped thing’s place, which is consistent with everything in Part 2 about cue–response structures. A vacated cue tends to get refilled.

What works [B]. An overview of systematic reviews published in Implementation Science (2024) synthesised de-implementation effectiveness across low-value care. Strategies falling under adapt and tailor to context, develop stakeholder interrelationships, and change infrastructure and workflow produced consistent reductions in low-value practice. These are ERIC strategy clusters, and they are the same structural and relational levers that appear throughout this report — not persuasion, not education alone.

Why it is hard, honestly stated. Decisions about what to stop are not purely evidential: political, economic and relational interests are attached to every existing practice. Someone’s role, budget, expertise or status is usually embedded in the thing being stopped, and cognitive biases plus organizational pressure reinforce established routines. This is the mirror image of Part 7’s institutional analysis — practices are retained for legitimacy reasons just as they are adopted for them.

Transfer to organizational change. The healthcare framing translates directly:

• Run a stop list alongside every change portfolio, with the same governance as the start list. Almost no organization does this, which is why portfolios only accumulate.
• Apply Replace, not just Remove — identify what occupies the freed cue, time slot or meeting.
• Expect the political layer and name it rather than treating stopping as a purely rational exercise.
• Measure de-implementation with the same rigour as implementation: reach (how many units actually stopped), fidelity (did it stop or just go underground), maintenance (is it still stopped at twelve months). Practices that are formally stopped and informally continued are the decoupling pattern in reverse.

---

Part 10 — Technology and AI adoption

Included because it is the live case in 2026, because it is the most evidence-rich behaviour-change problem currently available, and because almost every organization is getting it wrong in ways this report predicts.

10.1 The technology acceptance literature

TAM and UTAUT [B]. Venkatesh et al.’s Unified Theory of Acceptance and Use of Technology (2003) integrated eight prior models into four constructs — performance expectancy, effort expectancy, social influence, facilitating conditions — moderated by gender, age, experience and voluntariness. In the original longitudinal field study across four organizations it explained approximately 70% of variance in behavioural intention and about 50% in actual use, against 17–41% for predecessor models.

That is an unusually high figure for social science and deserves scrutiny rather than celebration. Caveats:

• The gap between the intention figure (70%) and the use figure (50%) is itself the point: intention is not use. This is the intention–action gap of Part 2.3 reappearing at organizational scale.
• UTAUT has been criticised for limited applicability outside organizational and working environments, and for reduced predictive performance in non-US settings, suggesting cultural moderation.
• Much of the subsequent literature consists of adding variables to UTAUT in new contexts, which inflates R² without adding theoretical content.

Notice the mapping: performance expectancy is Rogers’ relative advantage; effort expectancy is complexity inverted; social influence is norms and social proof; facilitating conditions is opportunity in COM-B terms. UTAUT is not a separate theory so much as the same mechanisms specified for a technology context. That is reassuring rather than disappointing — convergence across independent literatures is the strongest signal available in a field this weak.

10.2 What the AI trials actually show

The generative-AI period has produced unusually good experimental evidence, which will not last — take advantage of it.

Noy and Zhang (2023), Science [A]. 453 college-educated professionals, occupation-specific incentivised writing tasks, randomised access to ChatGPT. Average time taken fell by 40% and output quality rose by 18%. Participants with weaker skills benefited most, and inequality between workers decreased — a compression effect that recurs across this literature.

The adoption finding in that paper is as important as the productivity finding and is almost never quoted: workers exposed to ChatGPT during the experiment were twice as likely to report using it in their real job two weeks later, and 1.6 times as likely two months later. A single structured exposure produced measurable, decaying real-world adoption. That is trialability (Rogers) and mastery experience (Bandura) operating exactly as predicted — and the decay curve is the argument for spacing and reinforcement rather than one-off enablement.

Dell’Acqua et al. (2023), “Navigating the Jagged Technological Frontier” [A]. 758 BCG consultants, randomised. On 18 realistic tasks inside the AI capability frontier, consultants with GPT-4 completed 12.2% more tasks, 25.1% faster, with significantly higher quality. On a complex task deliberately selected to sit outside the frontier, subjects using AI were 19 percentage points less likely to produce correct solutions.

Software development field experiments [A]. Combined randomised evidence across Microsoft, Accenture and a Fortune 100 company covering 4,867 developers found roughly a 26% increase in completed tasks.

Three findings to carry:

1. Real, large productivity effects exist inside the capability frontier.
2. The frontier is jagged and invisible. Tasks that look similar in difficulty to a human sit on opposite sides of it, and users cannot reliably tell which is which. The failure mode is not that AI is unhelpful; it is that AI is confidently unhelpful on the tasks where the user cannot detect it.
3. Effects are largest for lower-skilled performers, which has significant implications for capability strategy, equity, and — less comfortably — for the value of mid-tier expertise.

10.3 The organizational adoption failure

MIT’s 2025 “State of AI in Business” analysis [C/D]. Based on 52 executive interviews, surveys of 153 leaders and analysis of 300 public deployments: approximately 95% of GenAI pilots delivered no measurable P&L impact. Only about 5% of custom tools crossed the pilot-to-production boundary. Generic chatbots reached high adoption for trivial tasks and stalled where workflows required context. The report’s central explanation is a learning gap — most systems do not retain feedback, adapt to context or improve over time.

Treat the 95% figure with the same scepticism as the 70% change-failure myth: it comes from a non-peer-reviewed report with a modest and non-random sample, “no measurable P&L impact” is a demanding and ambiguous criterion, and the figure travelled far faster than its methodology. It is directionally informative, not a fact.

The shadow AI finding is more interesting than the headline. The same analysis reported that while only around 40% of companies had official LLM subscriptions, roughly 90% of surveyed workers reported daily use of personal AI tools for work tasks. Broader 2026 workforce data shows similar patterns of high individual use alongside patchy organizational integration, with wide variation by role, sector and seniority, and with training, workflow integration and clear use cases identified as the binding constraints rather than access.

This is decoupling in the exact Meyer and Rowan sense: the formal structure (governed, procured, policy-compliant AI programme) has separated from the actual activity (people using whatever works). Recognising it as decoupling rather than as a compliance problem changes the response entirely.

10.4 What this report predicts about AI adoption

Every finding in Parts 2–9 applies, and the failures are predictable:

• Relative advantage is role-specific (Rogers). AI that saves the organization money while adding review burden to a specific role has negative relative advantage for that role. This is the single most common design error.
• Observability is low. AI use is mostly invisible — it happens inside someone’s window. Complex contagion cannot operate without observability, which is why adoption stalls in dispersed teams and spreads in co-located ones. Engineering observability is probably the highest-leverage cheap intervention available.
• Trialability is often destroyed by all-or-nothing enterprise rollouts and by policies that prohibit experimentation before approval.
• Absorptive capacity binds (Cohen & Levinthal; Szulanski). Teams without prior related knowledge cannot absorb the practice, and “prompt training” does not supply the domain judgement needed to detect the jagged frontier.
• Causal ambiguity is severe. High performers frequently cannot articulate what they do that works, so transfer fails.
• Capability, not awareness, is the deficit (COM-B). Almost all corporate AI enablement is education and persuasion; the evidence points at training, modelling and enablement.
• The jagged frontier is a judgement skill, acquired through practice with feedback in the specific domain — which is behaviour modelling and deliberate practice, not a webinar.
• NASSS predicts the failures directly: complexity in the technology, adopter system, organization and wider system domains simultaneously, which is the definition of the current situation.

The practical design that follows — and this is directly usable in consulting work:

1. Make use visible: shared prompt libraries, worked examples in team channels, demos in existing meetings, AI-assisted work labelled as such.
2. Create trialability: sanctioned low-stakes sandboxes with clear boundaries, before policy lockdown.
3. Teach the frontier, not the tool: paired practice on real domain tasks with immediate correctness feedback, so people learn where it fails.
4. Target specific behaviours in specific roles with genuine role-level relative advantage, rather than organization-wide enablement.
5. Treat shadow use as diagnostic data, not as a violation — it tells you exactly where relative advantage is real.
6. Instrument fidelity and maintenance, not licences activated. Seats provisioned is a reach metric masquerading as an outcome metric.

---

Part 11 — Cases, and what actually happened

Case evidence in this field is heavily contaminated by retrospective narrative construction. Each case below is given in the standard telling and the more defensible reading.

11.1 Michigan Keystone ICU — the best case in the literature

Standard telling: Peter Pronovost’s five-item checklist for central line insertion eliminated bloodstream infections across Michigan’s ICUs. Checklists save lives.

What happened: From 2003 the Michigan Health and Hospital Association ran a statewide programme across ICUs. The checklist covered hand hygiene, chlorhexidine skin prep, maximal barrier precautions, avoiding the femoral site, and daily review of line necessity. Median infection rate fell from 2.7 per 1,000 catheter-days at baseline to 0 within three months, with a 66% reduction sustained at 16–18 months, an estimated 1,500 lives and $175m saved, and continued decline through 2013 [A].

The more defensible reading: the checklist was not the intervention. Dixon-Woods, Bosk, Aveling, Goeschel and Pronovost’s “Explaining Michigan” (Milbank Quarterly, 2011) — a collaboration between two of the programme’s own leaders and three social scientists, developing an ex post theory using six theory-derived insights — describes the checklist as something close to a McGuffin: the visible object around which a much larger social programme was organised. The programme included statewide networking between units, comparative measurement infrastructure, explicit reframing of infections as preventable harm rather than inevitable complication, senior executive engagement per unit, and — decisively — empowering nurses to stop the procedure when sterile technique was breached. That last item is a status and authority intervention wearing a checklist’s clothing.

The confirming evidence is negative, twice over:

• Keystone Surgery applied the same checklist-plus-safety-programme approach to general surgery in Michigan hospitals. Reames et al. (JAMA Surgery, 2015) compared participating with non-participating hospitals and found no differences in surgical site infections or 30-day mortality [A]. The investigators’ own explanation is instructive: many sites lacked the infrastructure to collect and feed performance data back to frontline staff. The measurement loop did not travel with the checklist.
• Matching Michigan, the England-wide replication, produced far weaker results, and Dixon-Woods and colleagues’ ethnographic study of it (Implementation Science, 2013) traced this to the programme being experienced as top-down surveillance rather than clinician-owned improvement, with much weaker peer networks between units [B].

Same intervention artefact, three different social programmes, three different outcomes.

The transferable lesson: artefacts do not produce behaviour change; the social programme that gives an artefact meaning produces behaviour change. When an intervention is copied, the artefact travels and the social programme does not — which is Szulanski’s causal ambiguity, demonstrated at national scale. This is the single most important case in the literature for anyone selling or buying a change methodology.

11.2 Alcoa and the “keystone habit”

Standard telling: Paul O’Neill became CEO of Alcoa in 1987, announced that his sole priority was worker safety, and by focusing on this one “keystone habit” transformed the company — profits at record highs within a year, five-fold profit growth and a $27bn market-capitalisation increase by his retirement. Popularised by Charles Duhigg in The Power of Habit.

What is solid: the safety results. Alcoa’s lost-workday injury rate fell from around 1.86 per 100 workers at the start of O’Neill’s tenure to roughly 0.2, against a US average around 5 at the time. O’Neill’s personal commitment was genuine, publicly stated from his first speech as CEO, and sustained across his tenure (1987 to 1999). Some facilities went years without a lost-workday injury. That is a large, durable, well-documented behaviour change.

The more defensible reading of the causal story:

• Confounds. O’Neill made numerous other changes — capital discipline, portfolio decisions, management changes — during a period of favourable aluminium markets. Attributing a $27bn market-cap change to a safety focus is narrative, not analysis.
• Mechanism ambiguity. The plausible mechanism is not “keystone habit” in Duhigg’s habit-loop sense. It is that safety was a topic on which nobody could argue against the goal, which forced the organization to build cross-functional communication infrastructure, empowered frontline reporting, and demanded root-cause analysis — capabilities that then generalised. That is an organizational learning mechanism (Argyris’ double-loop; Weick’s preoccupation with failure), not a habit mechanism.
• Selection. We hear about Alcoa because it worked. There is no comparison set of firms that adopted a single unifying operational priority and did not transform.

The transferable lesson, stated honestly: choosing a single goal that is morally unarguable, measured relentlessly, escalated instantly, and personally owned by the CEO for over a decade builds infrastructure and habits of attention that generalise. That is worth knowing. “Find your keystone habit” is not a method.

11.3 GE Work-Out

Standard telling: Jack Welch’s Work-Out sessions removed bureaucracy by bringing employees together to challenge processes, with managers required to decide on proposals on the spot.

Defensible reading: the mechanism is well-specified and consistent with the evidence — participation by those who do the work, rapid decision (removing the delay that kills participation), and visible authority transfer. Documented improvements exist, but the evidence is company-reported and inseparable from a decade of simultaneous restructuring, delayering, Six Sigma deployment and portfolio management [C/D]. GE’s subsequent trajectory is also a caution against reading any of its practices as validated.

Transferable lesson: the “decide in the room” rule is the active ingredient and is cheap to copy. Most participative processes fail not because people won’t participate but because nothing visibly happens afterwards — a reinforcement failure that trains people not to participate next time. In Part 8’s terms, participation was placed on an extinction schedule.

11.4 Toyota Kata

Standard telling: Mike Rother’s research (2004–2009) identified that Toyota’s advantage lies not in its tools but in two repeated behavioural routines — the Improvement Kata (challenge, current condition, target condition, experiment toward it) and the Coaching Kata (a five-question routine by which managers develop that pattern in others).

Defensible reading: the strongest theoretical fit of any practitioner framework to the evidence in this report. Kata combines deliberate practice, coaching, routine modification and daily cadence — it targets the performative aspect of routines directly, uses short cycles with immediate feedback, and makes the manager a coach rather than a solution-provider. Empirically the evidence is largely case-based and practitioner-reported [C]; there is no controlled comparative evidence that Kata outperforms other structured improvement approaches.

Transferable lesson: the design principle worth stealing regardless of evidence grade is that behaviour change is taught as a repeated routine with a coach, not as a concept in a workshop. The five coaching questions are a scripted intervention with a fixed cadence. Scripting and cadence are what make it survive contact with busy managers.

11.5 Positive deviance in Vietnam

Standard telling: facing child malnutrition, Jerry and Monique Sternin found families whose children were well-nourished despite equal poverty, identified their uncommon practices, and had the community discover and practise these behaviours themselves rather than being taught.

Defensible reading: case results are well documented and the approach has spread to more than 60 countries [B for spread; C for comparative effectiveness]. The design satisfies an unusual number of evidence-supported mechanisms at once — in-group social proof, self-discovery supporting autonomy, practice rather than instruction, locally feasible solutions by construction, diffusion through dense village networks, and a direct assault on causal ambiguity. Weaknesses: positive deviants may differ in unmeasured ways; the approach requires that a solution already exists locally; and the evidence is overwhelmingly case-based.

Transferable lesson: in most organizations some teams already achieve the target outcome with the same resources. Finding and codifying what they actually do — with them, not about them — is a better first move than importing an external best practice, and it is almost never the first move taken.

11.6 Government nudge units

Standard telling: behavioural insights teams produced large returns from small interventions — tax letters, pension auto-enrolment, organ donation defaults.

Defensible reading: the strongest results come from a narrow class of interventions — defaults and structural friction changes in high-volume administrative systems. Auto-enrolment works because it changes the structure of the decision, not because it persuades. The broader “nudge” category does not survive publication-bias correction as a general effect [A]. There is also a documented in-house replication problem: large field partnerships have found much smaller average effects than the published literature implies.

Transferable lesson: the durable finding underneath nudge is friction, now formalised as sludge auditing — systematically cataloguing and costing unjustified frictions such as complex forms, hidden terms, waiting time and jargon, and deciding when and how to reduce them. The OECD has published international guidance on sludge audits. This is Lewin’s restraining forces in modern dress, and it is the most reliably useful idea from behavioural science for organizational application.

11.7 The pattern across cases

Read together, the cases produce a consistent finding that contradicts how change is usually sold:

The nameable artefact is never the mechanism. The checklist, the keystone habit, the kata card, the nudge — each is the visible object around which a social and structural programme is organised. Programmes that copy the artefact fail. Programmes that reconstruct the social and structural conditions succeed, whether or not they use the same artefact.

Three converging explanations, from three different literatures, for why this happens:

• Causal ambiguity (Szulanski): nobody can fully specify what makes it work, so the specifiable part gets copied.
• Institutional isomorphism (DiMaggio & Powell): copying a visible practice from a successful organization is a legitimate, defensible action under uncertainty regardless of whether it works.
• Ostensive/performative (Feldman & Pentland): the artefact is ostensive and portable; the performance is local and is not.

This should make you sceptical of every methodology purchase, including the ones you might sell.

---

Part 12 — The mechanisms that recur

Strip away the frameworks and a small set of mechanisms appears repeatedly across all levels and all bodies of evidence. This is the layer worth thinking in, because it survives the obsolescence of any particular model.

1. Cue and context. Behaviour is triggered by environment. Stable cues build automaticity; disrupted contexts open windows. Evidence: [A/B].

2. Friction. The number of steps between intention and action is a stronger determinant than the strength of the intention. Reducing friction on the desired behaviour and adding it to the undesired one is the most reliable structural intervention available. Evidence: [A] for defaults, [B] generally.

3. Capability. Knowing is not being able. Skill requires demonstration, practice, feedback and spacing. Most change programmes deliver information and call it capability. Evidence: [A].

4. Absorptive capacity. New practice attaches to prior related knowledge; without a base, adoption fails for reasons that look like resistance. Evidence: [B].

5. Social proof from credible peers. People adopt what people like them visibly do. Requires redundancy and in-group credibility — which is why external experts and senior leaders are poor sources of behavioural proof and good sources of legitimacy. Evidence: [B].

6. Observability. Behaviour that cannot be seen cannot be socially reinforced, cannot be imitated, and cannot diffuse. Under-recognised, and severely degraded by dispersed work. Evidence: [B].

7. Visible authority commitment. Ranks top in every practitioner dataset and is corroborated academically. Its function is informational — it signals that the change will outlast the quarter and that resources will follow. Evidence: [B/D].

8. Participation in design. Supports autonomy, reduces resistance, surfaces feasibility information, creates commitment. Effective even when the decision itself is not negotiable, provided the how genuinely is. Evidence: [B].

9. Feedback loops, task-directed. Short-cycle, process-focused, actionable. Default corporate implementations — annual, person-focused, comparative — are the versions the evidence says can harm. Evidence: [A].

10. Consequences. Behaviour that is noticed and has consequences persists; behaviour that isn’t, doesn’t. Prefer informational and social consequences to material ones. Evidence: [A].

11. Routine embedding. The behaviour must live inside something that recurs — an agenda, template, checklist, handover, review. Behaviours without a recurring host decay. Evidence: [B].

12. Finite attention. Change tolerance is limited. Adding a change without subtracting one degrades everything in flight. Evidence: [B/D].

A useful test for any proposed intervention: name which of the twelve it operates on, and how you would know it did. Communications campaigns typically operate on none of them and are the most commonly funded intervention in change programmes.

---

Part 13 — An integrated process

Synthesises Stouten et al.’s ten evidence-based steps, COM-B/BCW at individual level, Rogers/Greenhalgh for spread, CFIR/ERIC/RE-AIM/NPT at organizational level, and the transfer literature at the seam. Deliberately diagnostic-heavy, because that is what the evidence says gets skipped.

Phase 0 — Specify the behaviour

Nothing proceeds until this is done, and it is skipped in most programmes.

• Who exactly (named role, not “the organization”)
• does what exactly (observable action, not “collaborates better”)
• when and where (the cue and context)
• instead of what (the current behaviour being displaced — every new behaviour displaces something, and the something is usually invisible)
• how would we know (observable evidence, not survey response)

If a programme cannot fill this in for at least three specific behaviours, it does not yet have a change; it has an aspiration.

Two further questions, equally important:

Is behaviour change actually the right intervention? If the deficit is structural — wrong tools, wrong incentives, impossible workload, contradictory policy — behaviour change is the expensive wrong answer, and per the BBS critique it becomes a blame-displacement mechanism. Behaviour change programmes are frequently commissioned to avoid confronting structural problems.

Whose interests does this serve, and who loses? Ask it explicitly and early. If there is a genuine distributional cost to a specific group, naming it honestly is both more ethical and — per the procedural-fairness evidence — more effective than a communications strategy that pretends otherwise.

Phase 1 — Diagnose

Individual level (COM-B). For each target behaviour, is the deficit Capability (physical/psychological), Opportunity (physical/social), or Motivation (reflective/automatic)? Then check the answer against evidence rather than assumption — most organizations diagnose motivation when the evidence points at opportunity or capability. Szulanski’s finding is the standing warning here.

Innovation attributes (Rogers). From the adopter’s position, not the sponsor’s:

• Relative advantage — for this role?
• Compatibility — with how work is actually done?
• Complexity — how many other things must change simultaneously?
• Trialability — can someone try it on one case without committing?
• Observability — can people see others doing it and see results?

Weakness in any of these is a design problem, not an attitude problem, and design problems are cheaper to fix than attitudes.

Team level. Who models it now? Who has status among the target group? What routine could host the behaviour? What does the manager currently review? What is the peer norm and is it accurately perceived? Is there a positive deviant team?

Organizational level (CFIR domains).

• Intervention: relative advantage as perceived by those doing it, complexity, adaptability, evidence strength, cost.
• Outer setting: external pressures, regulation, client demands, sector norms.
• Inner setting: structures, culture and climate, implementation readiness, available resources, communication networks, tension for change, compatibility.
• Individuals: capability, opportunity, motivation of key roles — sponsors, implementation leads, opinion leaders, target adopters.
• Process: planning, engaging, executing, reflecting and evaluating.

Absorptive capacity check. Does the receiving unit have prior related knowledge? If not, capability building precedes rollout, and the rollout timeline is wrong.

Causal ambiguity check. Can anyone state why this works and what would break it? If not, you cannot specify core versus adaptable components, and adaptation during rollout will destroy the active ingredient without anyone noticing.

Portfolio check. What else lands on these people in the next two quarters? What could be stopped?

History check. What happened to the last three changes affecting this group? Oreg et al. establish that this shapes reactions before you say anything.

Readiness assessment (Weiner). Distinguish change commitment (shared determination to implement) from change efficacy (shared belief in collective capability), and assess efficacy against its three determinants — task demands, resource availability, situational factors. Both components are required and they are addressed by opposite interventions: low efficacy responds to resourcing, capability building, scope reduction and early wins; low commitment responds to rationale, participation and fairness. Adding persuasion to an efficacy problem, or training to a commitment problem, is the most common diagnostic error at this step. The ORIC instrument operationalises this if you want a measure rather than a judgement.

Phase 2 — Design

Select intervention functions from the Behaviour Change Wheel matched to the diagnosed deficit, then specify techniques — BCTTv1 at individual level, ERIC at organizational level. Name them. A change plan that cannot list its techniques is a communications plan.

Design rules with the strongest evidence:

• Bundle. Multi-faceted interventions outperform single techniques, consistently, across the OD and implementation literatures.
• Reduce friction first. Cheapest, fastest, least dependent on contested effect sizes. Run a sludge audit on the target behaviour: count the steps.
• Improve the innovation’s attributes before trying to spread it. Usually cheaper than persuasion.
• Build capability by demonstration and practice, not explanation. Behaviour modelling for interpersonal skills; backward chaining for procedures; spacing and retrieval throughout.
• Attach behaviours to existing routines. Do not create new meetings; modify existing agendas. New meetings are the first thing cut under pressure.
• Use implementation intentions explicitly. Have people write “when X, I will Y.” Ten minutes, one of the best effect sizes in the literature.
• Engineer observability. If the behaviour is invisible, make it visible before expecting it to spread.
• Seed in dense clusters — whole intact teams, not scattered volunteers.
• Design consequences. Mostly informational and social. Notice the behaviour. Avoid tangible expected rewards on anything you want people to care about.
• Design the sustainment mechanism now. Which recurring event hosts this in month nine, and who notices if it stops?
• Design the manager’s specific actions, not the manager’s attitude.
• Specify core versus adaptable components before rollout.

Explicitly design what stops. If nothing is being removed, the change is being added to a saturated system. Use the de-implementation discipline: Get Ready, Remove/Reduce/Replace, Optimize, Watch — and identify what will occupy the freed cue, time slot or meeting, because a vacated cue gets refilled.

Decide whether this is convergent or reorientation work. Incremental methods embed practice; they do not produce fundamental transformation. If the brief is reorientation, say so, and price and scope accordingly.

Choose your position on the let-it-happen / help-it-happen / make-it-happen continuum, and match the governance and effort to it. Applying make-it-happen machinery to something that needs emergent adoption is a recognisable failure pattern.

Phase 3 — Prepare the ground

• Secure active, visible sponsorship with specified behaviours: what will the sponsor personally do, how often, visible to whom. “Support” is not a plan.
• Communicate rationale, and what will not change. Loss drives most reaction; naming continuity reduces it. Avoid manufactured urgency in the form that denigrates current practice and the people who built it — both the identity-threat critique of Kotter and the recipient-reactions evidence point against it.
• Participate genuinely on the how. State clearly what is fixed and what is open. False participation is more damaging than none, because it teaches people that consultation is theatre — and, per Part 8, places participation on an extinction schedule.
• Resource it. Under-resourcing is a silent, common and fatal cause. If the change requires two hours a week of a team’s time, that time must come from somewhere named.
• Build source–recipient relationships where practice is being transferred between units.

Phase 4 — Implement

• Wave-based rollout with a real comparison group. You will roll out in waves anyway; make it a stepped wedge and get causal evidence for free.
• Short feedback cycles, task-focused, weekly not quarterly.
• Track fidelity, not just outcomes. If it did not work, you must be able to distinguish “the intervention failed” from “the intervention was not delivered.” This distinction is missing from most evaluations and is the difference between abandoning a good idea and fixing a delivery problem.
• Log adaptations with reasons.
• Expect and use objection. Recipients who object hold implementation information. Route it somewhere that can act on it and be seen to.
• Protect the pilot’s honesty. Pilots staffed with enthusiasts, extra resource and executive attention tell you nothing about rollout. Pilot under realistic conditions or do not claim the result generalises.

Phase 5 — Embed and sustain

Where most of the work is, and where most programmes have already declared victory.

• Routine hosting: the behaviour lives in an agenda item, a template field, a checklist step, a handover, a standing review.
• Role assignment: someone owns it after the programme closes, with authority and time.
• Measurement continuity: the indicator survives the programme’s reporting period.
• Reinforcement: it is noticed when done and when not done. Noticing is the cheapest reinforcement and the most frequently absent.
• Onboarding integration: new joiners learn the new behaviour as the only behaviour. This is how a change becomes invisible, which is the terminal state you want.
• Re-measure at 6 and 12 months, expecting decay, with a defined response to it.

NPT check at 6 months. Does it make sense to people (coherence)? Do they commit to it (cognitive participation)? Does the work actually get done (collective action)? Is it appraised and adjusted (reflexive monitoring)? A “no” on any of the four tells you precisely where embedding failed.

Decoupling check. Is the structure being maintained ceremonially while the practice has reverted? The test is Part 4.5: show me the last three times this actually happened.

Phase 6 — Evaluate honestly

Use RE-AIM — the five things that must all hold:

• Reach: what proportion of the intended population was actually touched?
• Effectiveness: did it change the target behaviour and the outcome, against a comparison?
• Adoption: what proportion of units and staff took it up, and how did adopters differ from non-adopters?
• Implementation: was it delivered as designed? What adaptations were made and why?
• Maintenance: is it still happening at 6 and 12 months?

Report the failures. An evaluation that reports only success is a marketing document, and in a consulting relationship it destroys the credibility that makes the second engagement possible.

---

Part 14 — Measurement

14.1 What to measure

Leading indicators — behaviour-proximate, weekly or fortnightly:

• Frequency of the target behaviour (observed or system-logged, not self-reported)
• Opportunity to perform (was the situation encountered at all?)
• Manager enabling actions performed (count of specified actions)
• Peer interactions about the behaviour
• Friction metrics: steps, time or clicks to perform the behaviour

Fidelity indicators:

• Was the intervention delivered as designed?
• What adaptations were made, by whom and why? (Adaptation is not automatically failure; undocumented adaptation is.)

Lagging indicators — quarterly and beyond:

• The operational outcome the behaviour was supposed to move
• Maintenance at 6 and 12 months
• Reach and adoption across units

Diagnostic instruments worth using:

• COM-B self-assessment for the target population, pre and post
• LTSI for training-linked change (16 validated factors)
• Readiness — commitment and efficacy — pre-implementation
• Psychological safety at team level, with within-group agreement statistics
• CFIR-structured determinant assessment for anything organization-wide

14.2 The counterfactual

The most valuable and most neglected measurement practice: know what would have happened anyway.

Realistic options, in descending rigour:

1. Stepped-wedge rollout — all units eventually receive the intervention, order randomised. Ethically comfortable, operationally natural, genuinely causal.
2. Matched comparison units — pair similar teams, roll out to one first.
3. Interrupted time series — enough pre-period data points to establish a trend, then look for a level or slope change.
4. Pre/post with documented external context — weakest, but honest if the confounds are named.

Almost every organizational rollout could be option 1 at near-zero marginal cost. The obstacle is not methodological; it is that nobody asks in the planning meeting.

14.3 Measurement traps specific to this field

• Common-method inflation. Blume et al. found transfer relationships systematically inflated when predictor and outcome come from the same source in the same context. If the person who attended the training also rates whether they use it, your number is inflated by an unknown amount.
• Self-report of behaviour. People are poor at reporting their own behaviour frequency, and worse at reporting the sources of their own learning (see 70-20-10).
• Reporting-period artefacts. Effects measured while a programme is running and being watched include the attention effect. Measure after the spotlight moves.
• Metric capture. Any metric attached to consequences will be optimised, including in ways you did not intend. BBS under-reporting is the canonical case; goals-gone-wild is the general theory.
• Survivorship in case selection. Success Case Method deliberately samples both extremes for this reason. Most internal case studies sample only one.

---

Part 15 — The folklore table

Widely repeated in change and L&D work; not supported by the evidence. Precise knowledge here is a professional differentiator — delivered with tact, since many people have built careers on some of these.

Claim	Status	What’s actually true
It takes 21 days to form a habit	False	Median 66 days, range 18–254 (Lally 2010); the 21 comes from a 1960 plastic-surgery observation
70% of change initiatives fail	Unsupported	Traced to a hedged 1993 remark about reengineering; no source study (Hughes 2011)
Lewin gave us unfreeze–change–refreeze	False	Assembled after his death (Cummings et al. 2016). His actual contribution is force field analysis
Kotter’s 8 steps are research-validated	False	No study validates the model or its sequence (Appelbaum et al. 2012)
The Kübler-Ross curve describes organizational change	False	Bonanno’s longitudinal bereavement research found no evidence for the five stages even in grief; roughly two-thirds of bereaved people show a resilient trajectory rather than staged distress. Application to organizational change is a double extrapolation. Use Bridges if you need a transition model
70-20-10 is an evidence-based ratio	Unsupported	Retrospective self-report by ~200 successful executives; never a prescription (Clardy 2018)
Learning styles improve outcomes when matched	False	No adequate supporting evidence (Pashler et al.); repeatedly falsified
We remember 10% of what we read, 90% of what we do	Fabricated	No traceable source study
Adults learn by fundamentally different mechanisms (andragogy)	Weak	Assumptions not empirically derived; evidence inconclusive; Knowles later described it as assumptions, not theory
Tuckman’s forming–storming–norming–performing is a reliable sequence	Weak	Derived from therapy-group literature; ~50% of studies show no distinct storming stage
MBTI is suitable for selection, team design or development decisions	False	Severe test-retest classification instability; near-zero performance correlation; SIOP advises against for selection
Grit is a distinct predictor beyond conscientiousness	False	ρ = .84 with conscientiousness across 66,807 individuals (Credé et al. 2017)
Growth mindset interventions reliably raise achievement	Contested/weak	Weak in Sisk et al. 2018; null in the highest-quality subset (Macnamara & Burgoyne 2023); heterogeneity argument contests this
SCARF is a neuroscience-derived model	Overstated	Practitioner synthesis; the five-domain grouping is not empirically derived. Mostly recoverable from SDT and justice research
Willpower is a depletable resource (ego depletion)	False as formulated	d = 0.06 in preregistered multi-lab replication
Stage-matched interventions beat unmatched (Transtheoretical Model)	False	Cochrane: no difference in direct comparisons across 41 trials
People naturally resist change	Oversimplified	Reactions are tridimensional and driven by fairness, trust, history and self-interest (Oreg et al. 2011)
Feedback improves performance	Partly false	Average d = 0.41 but harms in over a third of cases (Kluger & DeNisi 1996)
Multi-source (360) feedback improves outcomes	Weak	Associated with smaller coaching effects (Jones et al. 2016)
Money motivates performance	Half true	Correlates .34 with quantity, .08 with quality (Jenkins et al. 1998); tangible expected rewards undermine intrinsic motivation
Nudges reliably produce behaviour change at scale	Weak	Effect collapses toward zero after publication-bias correction. Defaults and friction changes survive
You change culture with a culture programme	False	Change behaviour, climate and structures; culture lags
Find your influencers to spread change	Misleading	Behaviour needs redundant exposure; seed dense clusters (Centola)
The main barrier to internal best-practice transfer is motivation	False	Knowledge-related barriers dominate: absorptive capacity, causal ambiguity, arduous relationships (Szulanski 1996)
Communication is the main change lever	False	Necessary, never sufficient; addresses none of the twelve mechanisms directly
The checklist saved the lives	Incomplete	The social programme around it did; the same method produced no effect in Keystone Surgery
Deliberate practice explains expertise	Contested	<1% of performance variance in professions (Macnamara et al. 2014); disputed by Ericsson
Quick wins build momentum into transformation	Unsupported	Small changes in strategy, structure and power did not accumulate into fundamental transformation (Romanelli & Tushman 1994). Quick wins may still help embed a specific practice
Psychological safety can’t really be engineered	Outdated	A field RCT across ~7,000 employees raised it by restructuring manager one-to-ones
Resistance is located in the recipients	False	Change agents contribute to resistance through their own actions and inactions; resistance can be a resource (Ford, Ford & D’Amelio 2008)
Readiness means enthusiasm	Imprecise	Two distinct components — change commitment and change efficacy — requiring opposite interventions (Weiner 2009)
The evidence base generalises across cultures	Unverified	~70–80% of organizational samples from US/Europe (~11% of world population)
Gamification works through points and leaderboards	False	Effects are moderated by game fiction and social interaction, not reward mechanics (Sailer & Homner 2020)
Popular adoption of a method indicates it works	False	Management fashion follows bell-shaped popularity curves independent of efficacy (Abrahamson)
95% of AI pilots fail	Treat as the new 70%	Non-peer-reviewed, modest sample, demanding criterion; directionally informative, not a fact

---

Part 16 — Gaps in the literature

For the academic track: places where the field is thin and the questions are answerable.

1. The team level is under-theorised. A mature individual literature, a mature organizational literature, and a comparatively empty middle. Specifically: how does a team-level behavioural norm establish? What is the minimum number of visible adopters, and does it vary by behaviour type and team size? Centola’s work suggests the question is tractable experimentally.

2. Sustainment mechanisms are described but not tested. The sustainability literature reliably lists factors and rarely tests interventions. A randomised comparison of specific sustainment strategies — routine hosting versus periodic re-training versus measurement continuity versus role assignment — does not exist and is entirely feasible in a multi-site organization.

3. Absorptive capacity, in the attentional sense, is unquantified. How many concurrent behavioural changes can a team absorb before marginal returns turn negative? Practitioner data suggests the number is small. Nobody has estimated it properly, and it would be one of the most practically consequential numbers in the field.

4. The corporate/implementation-science translation gap is itself a research object. CFIR, NPT, NASSS and RE-AIM have been validated almost exclusively in health and public sector settings. Their performance in commercial organizational change is untested, and testing it is a legitimate and publishable programme with an obvious practitioner audience.

5. Fidelity and adaptation. When is adaptation beneficial (fitting local context) and when is it fatal (dropping the active ingredient)? Implementation science has begun this; organizational change has not touched it, and it is the question underneath every methodology-rollout failure.

6. Failure documentation and base rates. Publication and reporting bias in this field is extreme. There is no trial registry equivalent for organizational change, and consequently no way to estimate base rates — which is precisely why the 70% myth survived unchallenged for decades. A registry of pre-registered organizational change initiatives with pre-specified success criteria would be a significant contribution and is technically trivial.

7. Longitudinal individual behaviour in work settings. Lally-style automaticity tracking has essentially never been done for workplace behaviours, which differ from health habits in frequency, complexity and social dependence. The 66-day figure is quoted in corporate settings constantly with no evidence it applies.

8. Whether change management as a discipline adds value net of confounds. The 7× figure is self-reported by practitioners rating their own work. A serious independent test — matched projects, third-party outcome assessment, blinded rating — has not been done and would be uncomfortable for everyone selling the service.

9. Observability as a manipulable variable. The diffusion literature identifies it; nobody has run the experiment of deliberately engineering observability for an organizational behaviour and measuring adoption against a control. Given dispersed work, this is both timely and cheap to test.

10. AI adoption as behaviour change. The productivity RCTs are excellent; the adoption research is almost entirely survey-based and vendor-adjacent. The open questions — what predicts durable individual AI practice, how the jagged frontier is learned, whether shadow use converts to sanctioned use or is suppressed by it — are answerable now and will be historically unrepeatable once the technology stabilises.

11. The decoupling hypothesis, tested directly. Institutional theory predicts that change programmes adopted under isomorphic pressure will show high structural adoption and low performative change. This is testable by measuring ostensive adoption (documents, governance, training completion) against performative adoption (observed behaviour) across a portfolio, and predicting the gap from the strength of external legitimacy pressure. Nobody has done it, and it would be a genuinely important paper.

12. Cross-cultural replication of change interventions. Given that 70–80% of the evidence base comes from ~11% of the world’s population, and that OD’s core practices encode culturally specific assumptions about agency and authority, direct replication of well-established interventions in non-WEIRD organizational settings is both straightforward and under-supplied. The implementation-intentions and behaviour-modelling findings are the obvious candidates: strong, simple, and never seriously tested outside Western white-collar populations.

13. De-implementation outside healthcare. The construct, frameworks and effectiveness evidence are almost entirely healthcare-derived. Whether the same strategy clusters — contextual tailoring, stakeholder relationships, infrastructure and workflow change — reduce low-value practice in commercial organizations is untested and directly answerable, and it is the highest-value unanswered question for anyone working on change-portfolio saturation.

14. The convergent/reorientation boundary. Romanelli and Tushman established that small changes do not accumulate into transformation. Nobody has established where the boundary is — how large or how coordinated a change must be to constitute reorientation rather than convergence — which is precisely the judgement practitioners are asked to make.

---

Part 17 — Applying this to yourself

Reading the individual literature honestly, the operations that survive scrutiny for a single person are unglamorous.

Design the context, not the resolve. Most self-change failure is environmental. Where the behaviour happens, what is visible, what is within reach, and what the default is will beat willpower — and the willpower construct itself did not survive replication. Before trying harder, remove a step.

Use implementation intentions literally. Write “when [specific cue], I will [specific action].” Not a resolution — a conditional. The highest-return ten minutes in the literature (d ≈ 0.65).

Expect months, tolerate misses. Median 66 days for simple daily behaviours, enormous variance, and missing a day does not break the curve. Streak-based systems are motivationally effective and psychologically brittle; the evidence does not require perfection.

Exploit discontinuities deliberately. Job changes, moves, new projects, new clients, new tools — windows when context-bound behaviour is already loose. Stack intended changes onto transitions you are already going through rather than launching them from a stable state. This is the single most under-used finding for personal application.

To break something, disrupt the cue before you try to resist it. Cue disruption > environmental re-engineering > vigilant monitoring, in that order of durability and inverse order of effort.

Learning goals before performance goals on anything you do not yet know how to do. “Acquire and use three specific techniques,” not an outcome number, until the strategies exist.

Space and retrieve. For anything you want to retain — a literature, a language, a domain — spaced retrieval beats rereading by a wide margin, and the optimal gap grows with the retention interval you care about. Build the review schedule when you build the note, not later.

Get task-focused feedback and refuse the other kind. Ask “what specifically would make this better” rather than “what did you think.” The former directs attention to the task; the latter to you, which is the documented failure mode.

Watch someone competent do the thing, not explain it. Vicarious mastery is second only to your own mastery experience in building self-efficacy, and it is far cheaper. Observe the performance; the explanation is the ostensive account and is usually wrong about what matters.

Practise scripted, not conceptual. The Kata insight applies at n = 1: a short scripted routine on a fixed cadence beats an understood principle applied when remembered.

Watch for your own causal ambiguity. When something you do works, write down why you think it works and what would break it, while you still remember. This is what makes your practice transferable — to clients, to collaborators, and to yourself in eighteen months.

---

Part 18 — Using this commercially

Notes on deployment, given the consulting-practice purpose. These are judgements, not findings.

Lead with diagnosis, not methodology. The evidence’s clearest message is that diagnosis-before-solution is the most-skipped and highest-value step. It is also commercially advantageous: a diagnostic engagement is fast, low-risk for the client, differentiating, and generates the specification for everything that follows. Most competitors arrive with a methodology; arriving with a diagnostic protocol is a different position.

The three-question opener. Almost any change conversation can be productively opened with: What specific behaviour, by whom, is supposed to be different? — How would we know? — What happened to the last three changes here? These map to Phase 0 and the history check, and they surface more useful information in ten minutes than a discovery workshop.

Satisfy the legitimacy function on purpose. Per Part 7, evidence-based practice will not displace a fashionable model unless it also looks like a method: named, diagrammed, sequenced, auditable. Package the Part 13 process as a method. This is not dishonesty; it is meeting a real institutional requirement so the substance can travel.

Use the folklore knowledge carefully. Correcting someone’s cherished model in front of their colleagues buys you a point and costs you the room. Better patterns: ask what the claim would predict and whether that has been observed; offer the correction as a curiosity (“there’s a strange history to that number”); or simply build on the useful residue without contesting the framing. Save the direct correction for cases where the false belief is about to cause a costly decision.

Sell the seam, not the event. The transfer evidence says the value is in the weeks after the intervention, in the team, through managers and peers. Most L&D procurement buys the event. Repositioning the offer around the seam is both better-evidenced and commercially defensible, because it is where the client’s money is currently being wasted.

Price the counterfactual as a feature. Offering a stepped-wedge rollout with a real comparison group is nearly free operationally and is a strong differentiator: it means the client will actually know whether it worked. Few competitors will offer this because few want the answer.

Be honest about effect sizes. The evidence supports moderate effects from well-designed interventions, not transformation. Practitioners who promise transformation are competing on a claim you should not match. The counter-position — “here is what the evidence says this will move, and how we will know” — is a narrower promise and a much more defensible one, and it survives the second engagement.

Watch your own decoupling. The same institutional pressures act on consultants. If the deliverable is a document, a framework and a governance structure, and nothing in the client’s Tuesday changes, you have produced ceremony. The Part 4.5 test applies to your own work: show me the last three times this actually happened.

---

Part 19 — Reading list, tiered

Read first — highest value per hour

1. Stouten, Rousseau & De Cremer (2018), “Successful Organizational Change,” Academy of Management Annals 12(2). The best single organizational change source. Start here.
2. Michie, van Stralen & West (2011), “The Behaviour Change Wheel,” Implementation Science 6:42. The best individual-level diagnostic.
3. Blume, Ford, Baldwin & Huang (2010), “Transfer of Training: A Meta-Analytic Review,” Journal of Management 36(4). The transfer problem, quantified.
4. Kluger & DeNisi (1996), “The Effects of Feedback Interventions on Performance,” Psychological Bulletin 119(2). Still the most ignored important finding in the field.
5. Szulanski (1996), “Exploring Internal Stickiness,” Strategic Management Journal 17(S2). Short, and it reverses the default diagnosis.
6. Hughes (2011), “Do 70 Per Cent of All Organizational Change Initiatives Really Fail?” Journal of Change Management 11(4). Short, and inoculating.

Read second — depth

1. Greenhalgh, Robert, Macfarlane, Bate & Kyriakidou (2004), “Diffusion of Innovations in Service Organizations,” Milbank Quarterly 82(4).
2. Cummings, Bridgman & Brown (2016), “Unfreezing change as three steps,” Human Relations 69(1).
3. Oreg, Vakola & Armenakis (2011), “Change Recipients’ Reactions to Organizational Change,” Journal of Applied Behavioral Science 47(4).
4. Frazier et al. (2017), “Psychological Safety: A Meta-Analytic Review and Extension,” Personnel Psychology 70(1).
5. Lacerenza et al. (2017), “Leadership Training Design, Delivery, and Implementation,” Journal of Applied Psychology 102(12).
6. Taylor, Russ-Eft & Chan (2005), “A Meta-Analytic Review of Behavior Modeling Training,” Journal of Applied Psychology 90(4).
7. Feldman & Pentland (2003), “Reconceptualizing Organizational Routines,” Administrative Science Quarterly 48(1).
8. Damschroder et al. (2009) CFIR, plus the 2022 update and the 2025 CFIR User Guide.
9. Powell et al. (2015), ERIC — the 73 implementation strategies, Implementation Science 10:21.
10. Greenhalgh et al. (2017), NASSS framework, Journal of Medical Internet Research 19(11) e367.
11. Weiner (2009), “A theory of organizational readiness for change,” Implementation Science 4:67. Short, and it fixes the vaguest word in the practitioner vocabulary.
12. Ford, Ford & D’Amelio (2008), “Resistance to Change: The Rest of the Story,” Academy of Management Review 33(2).
13. Romanelli & Tushman (1994), “Organizational Transformation as Punctuated Equilibrium: An Empirical Test,” Academy of Management Journal 37(5). The finding that small changes don’t accumulate.
14. Castro, Englmaier & Guadalupe, “Fostering Psychological Safety in Teams: Evidence from an RCT.” The intervention evidence that didn’t exist five years ago.

Read third — orientation and critique

1. Centola, How Behavior Spreads: The Science of Complex Contagions (Princeton, 2018).
2. Michie et al. (2013), BCT Taxonomy v1, Annals of Behavioral Medicine 46(1).
3. Ordóñez, Schweitzer, Galinsky & Bazerman (2009), “Goals Gone Wild,” Academy of Management Perspectives 23(1) — with Locke & Latham’s reply.
4. Meyer & Rowan (1977), “Institutionalized Organizations,” American Journal of Sociology 83(2); and DiMaggio & Powell (1983), “The Iron Cage Revisited,” American Sociological Review 48(2). The explanation for the field’s own pathology.
5. Abrahamson (1996), “Management Fashion,” Academy of Management Review 21(1), 254–285.
6. West (2005), “Time for a change: putting the Transtheoretical Model to rest,” Addiction 100(8).
7. Clardy (2018), “70-20-10 and the Dominance of Informal Learning,” Human Resource Development Review 17(2).
8. Maier et al. (2022), “No evidence for nudging after adjusting for publication bias,” PNAS 119(31) — with Mertens et al.’s original and the published rejoinders.
9. Dixon-Woods et al. (2011), “Explaining Michigan,” Milbank Quarterly 89(2); with Dixon-Woods et al. (2013), “Explaining Matching Michigan,” Implementation Science 8:70, and Reames et al. (2015) on Keystone Surgery. The best natural experiment in the literature.
10. Dell’Acqua et al. (2023/2025), “Navigating the Jagged Technological Frontier,” Organization Science; with Noy & Zhang (2023), Science 381(6654).
11. Weick, Sensemaking in Organizations (1995); and Weick & Sutcliffe, Managing the Unexpected — with Perrow’s Normal Accidents as counterweight.
12. Implementation Science (2024), overview of systematic reviews on de-implementation of low-value practices.
13. CIPD evidence reviews (organisational culture and climate; psychological safety, 2024; behaviour change; learning). The most useful practitioner-facing synthesis of academic evidence available, and free.

Popular books, used correctly

• Clear, Atomic Habits — excellent synthesis, weak sourcing. Use for language, never for citation.
• Duhigg, The Power of Habit — good on the habit loop, unreliable on the case narratives (see Alcoa).
• Heath & Heath, Switch — the Rider/Elephant/Path framing popularises the capability/motivation/context triad and maps cleanly onto COM-B. Cases are cherry-picked.
• Rother, Toyota Kata — the best practitioner book on the mechanics of repeated behavioural practice, evidence grade notwithstanding.
• Edmondson, The Fearless Organization — the construct’s originator, closer to the evidence than most popularisations.
• Wood, Good Habits, Bad Habits — the best popular treatment of the context-cue account, by the researcher who built it.

---

References

Abrahamson, E. (1996). Management fashion. Academy of Management Review, 21(1), 254–285. See also Abrahamson & Fairchild (1999), Management fashion: lifecycles, triggers, and collective learning processes. Administrative Science Quarterly, 44(4), 708–740.

Appelbaum, S. H., Habashy, S., Malo, J-L. & Shafiq, H. (2012). Back to the future: revisiting Kotter’s 1996 change model. Journal of Management Development, 31(8), 764–782.

Argyris, C. & Schön, D. (1978/1996). Organizational Learning and Organizational Learning II. Addison-Wesley.

Baldwin, T. T. & Ford, J. K. (1988). Transfer of training: A review and directions for future research. Personnel Psychology, 41(1), 63–105.

Blume, B. D., Ford, J. K., Baldwin, T. T. & Huang, J. L. (2010). Transfer of training: A meta-analytic review. Journal of Management, 36(4), 1065–1105.

Brinkerhoff, R. O. (2005). The Success Case Method: A strategic evaluation approach to increasing the value and effect of training. Advances in Developing Human Resources, 7(1), 86–101.

Bushe, G. R. & Marshak, R. J. (2015). Dialogic Organization Development. Berrett-Koehler.

Bonanno, G. A., Wortman, C. B. et al. (2002; 2004). Resilience to loss and chronic grief: A prospective study from preloss to 18 months postloss. Journal of Personality and Social Psychology, 83(5); and Prospective patterns of resilience and maladjustment during widowhood. Psychology and Aging, 19(2).

Castro, S., Englmaier, F. & Guadalupe, M. Fostering psychological safety in teams: Evidence from an RCT. HEC Paris working paper.

Centola, D. (2010). The spread of behavior in an online social network experiment. Science, 329(5996), 1194–1197. And Centola (2018), How Behavior Spreads. Princeton University Press.

CIPD (2024). Psychological safety: An evidence review. Chartered Institute of Personnel and Development.

Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T. & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354–380.

Clardy, A. (2018). 70-20-10 and the dominance of informal learning: A fact in search of evidence. Human Resource Development Review, 17(2), 153–178.

Cohen, W. M. & Levinthal, D. A. (1990). Absorptive capacity: A new perspective on learning and innovation. Administrative Science Quarterly, 35(1), 128–152.

Credé, M., Tynan, M. C. & Harms, P. D. (2017). Much ado about grit: A meta-analytic synthesis of the grit literature. Journal of Personality and Social Psychology, 113(3), 492–511.

Cummings, S., Bridgman, T. & Brown, K. G. (2016). Unfreezing change as three steps: Rethinking Kurt Lewin’s legacy for change management. Human Relations, 69(1), 33–60.

Damschroder, L. J. et al. (2009). Fostering implementation of health services research findings into practice: a consolidated framework. Implementation Science, 4:50. Updated: Damschroder, Reardon, Opra Widerquist & Lowery (2022), Implementation Science, 17:75. CFIR User Guide (2025).

Deci, E. L., Koestner, R. & Ryan, R. M. (1999). A meta-analytic review of experiments examining the effects of extrinsic rewards on intrinsic motivation. Psychological Bulletin, 125(6), 627–668.

Deci, E. L., Olafsen, A. H. & Ryan, R. M. (2017). Self-determination theory in work organizations: The state of a science. Annual Review of Organizational Psychology and Organizational Behavior, 4, 19–43.

Dell’Acqua, F., McFowland, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. & Lakhani, K. (2023/2025). Navigating the jagged technological frontier. Organization Science (HBS Working Paper 24-013).

DiMaggio, P. J. & Powell, W. W. (1983). The iron cage revisited: Institutional isomorphism and collective rationality in organizational fields. American Sociological Review, 48(2), 147–160.

Dixon-Woods, M., Bosk, C. L., Aveling, E. L., Goeschel, C. A. & Pronovost, P. J. (2011). Explaining Michigan: Developing an ex post theory of a quality improvement program. Milbank Quarterly, 89(2), 167–205.

Dixon-Woods, M. et al. (2013). Explaining Matching Michigan: An ethnographic study of a patient safety program. Implementation Science, 8:70.

Feldman, M. S. & Pentland, B. T. (2003). Reconceptualizing organizational routines as a source of flexibility and change. Administrative Science Quarterly, 48(1), 94–118.

Ford, J. D., Ford, L. W. & D’Amelio, A. (2008). Resistance to change: The rest of the story. Academy of Management Review, 33(2), 362–377.

Frazier, M. L., Fainshmidt, S., Klinger, R. L., Pezeshkan, A. & Vracheva, V. (2017). Psychological safety: A meta-analytic review and extension. Personnel Psychology, 70(1), 113–165.

Gersick, C. J. G. (1991). Revolutionary change theories: A multilevel exploration of the punctuated equilibrium paradigm. Academy of Management Review, 16(1), 10–36.

Glasgow, R. E., Vogt, T. M. & Boles, S. M. (1999). Evaluating the public health impact of health promotion interventions: the RE-AIM framework. American Journal of Public Health, 89(9), 1322–1327.

Gollwitzer, P. M. & Sheeran, P. (2006). Implementation intentions and goal achievement: A meta-analysis of effects and processes. Advances in Experimental Social Psychology, 38, 69–119.

Greenhalgh, T., Robert, G., Macfarlane, F., Bate, P. & Kyriakidou, O. (2004). Diffusion of innovations in service organizations: Systematic review and recommendations. Milbank Quarterly, 82(4), 581–629.

Greenhalgh, T., Wherton, J., Papoutsi, C., Lynch, J., Hughes, G., A’Court, C. et al. (2017). Beyond adoption: a new framework for theorizing and evaluating nonadoption, abandonment, and challenges to the scale-up, spread, and sustainability of health and care technologies (NASSS). Journal of Medical Internet Research, 19(11), e367. See also the NASSS-CAT complexity assessment tools (2020).

Grimshaw, J. et al. / Norton, W. E. et al. — de-implementation frameworks; and Implementation Science (2024), Effectiveness of de-implementation of low-value healthcare practices: an overview of systematic reviews, 19:1384. GROW framework (Get Ready, Remove/Reduce/Replace, Optimize, Watch).

Holton, E. F., Bates, R. A. & Ruona, W. E. A. (2000); Bates, Holton & Hatala (2012). Learning Transfer System Inventory: development and revised factorial replication and validation. Human Resource Development Quarterly / Human Resource Development International, 15(5).

Hughes, M. (2011). Do 70 per cent of all organizational change initiatives really fail? Journal of Change Management, 11(4), 451–464.

Jenkins, G. D., Mitra, A., Gupta, N. & Shaw, J. D. (1998). Are financial incentives related to performance? A meta-analytic review of empirical research. Journal of Applied Psychology, 83(5), 777–787.

Jones, R. J., Woods, S. A. & Guillaume, Y. R. F. (2016). The effectiveness of workplace coaching: A meta-analysis of learning and performance outcomes from coaching. Journal of Occupational and Organizational Psychology, 89(2), 249–277.

Kluger, A. N. & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin, 119(2), 254–284.

Lacerenza, C. N., Reyes, D. L., Marlow, S. L., Joseph, D. L. & Salas, E. (2017). Leadership training design, delivery, and implementation: A meta-analysis. Journal of Applied Psychology, 102(12), 1686–1718.

Lally, P., van Jaarsveld, C. H. M., Potts, H. W. W. & Wardle, J. (2010). How are habits formed: Modelling habit formation in the real world. European Journal of Social Psychology, 40(6), 998–1009.

Locke, E. A. & Latham, G. P. (1990; 2002). A Theory of Goal Setting and Task Performance; Building a practically useful theory of goal setting. American Psychologist, 57(9), 705–717.

Lundahl, B. W., Kunz, C., Brownell, C., Tollefson, D. & Burke, B. L. (2010). A meta-analysis of motivational interviewing: Twenty-five years of empirical studies. Research on Social Work Practice, 20(2), 137–160.

Macnamara, B. N., Hambrick, D. Z. & Oswald, F. L. (2014). Deliberate practice and performance in music, games, sports, education, and professions: A meta-analysis. Psychological Science, 25(8), 1608–1618.

Macnamara, B. N. & Burgoyne, A. P. (2023). Do growth mindset interventions impact students’ academic achievement? A systematic review and meta-analysis. Psychological Bulletin, 149(3–4), 133–173.

Maier, M., Bartoš, F., Stanley, T. D., Shanks, D. R., Harris, A. J. L. & Wagenmakers, E-J. (2022). No evidence for nudging after adjusting for publication bias. PNAS, 119(31).

March, J. G. (1991). Exploration and exploitation in organizational learning. Organization Science, 2(1), 71–87.

Mattingly, V. & Kraiger, K. (2019). Can emotional intelligence be trained? A meta-analytical investigation. Human Resource Management Review, 29(2), 140–155.

May, C. & Finch, T. (2009). Implementing, embedding, and integrating practices: An outline of Normalization Process Theory. Sociology, 43(3), 535–554.

Mertens, S., Herberz, M., Hahnel, U. J. J. & Brosch, T. (2022). The effectiveness of nudging: A meta-analysis of choice architecture interventions across behavioral domains. PNAS, 119(1).

Meyer, J. W. & Rowan, B. (1977). Institutionalized organizations: Formal structure as myth and ceremony. American Journal of Sociology, 83(2), 340–363.

Michie, S., van Stralen, M. M. & West, R. (2011). The behaviour change wheel: A new method for characterising and designing behaviour change interventions. Implementation Science, 6:42.

Michie, S. et al. (2013). The behavior change technique taxonomy (v1) of 93 hierarchically clustered techniques. Annals of Behavioral Medicine, 46(1), 81–95.

Neuman, G. A., Edwards, J. E. & Raju, N. S. (1989). Organizational development interventions: A meta-analysis of their effects on satisfaction and other attitudes. Personnel Psychology, 42(3), 461–489.

Noy, S. & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.

Ordóñez, L. D., Schweitzer, M. E., Galinsky, A. D. & Bazerman, M. H. (2009). Goals gone wild: The systematic side effects of overprescribing goal setting. Academy of Management Perspectives, 23(1), 6–16.

Oreg, S., Vakola, M. & Armenakis, A. (2011). Change recipients’ reactions to organizational change: A 60-year review of quantitative studies. Journal of Applied Behavioral Science, 47(4), 461–524.

Pashler, H., McDaniel, M., Rohrer, D. & Bjork, R. (2008). Learning styles: Concepts and evidence. Psychological Science in the Public Interest, 9(3), 105–119.

Powell, B. J. et al. (2015). A refined compilation of implementation strategies: results from the Expert Recommendations for Implementing Change (ERIC) project. Implementation Science, 10:21.

Pronovost, P. et al. (2006). An intervention to decrease catheter-related bloodstream infections in the ICU. New England Journal of Medicine, 355(26), 2725–2732.

Reames, B. N., Krell, R. W., Campbell, D. A. et al. (2015). A checklist-based intervention to improve surgical outcomes in Michigan: evaluation of the Keystone Surgery program. JAMA Surgery, 150(3), 208–215.

Rogers, E. M. (2003). Diffusion of Innovations (5th ed.). Free Press.

Romanelli, E. & Tushman, M. L. (1994). Organizational transformation as punctuated equilibrium: An empirical test. Academy of Management Journal, 37(5), 1141–1166.

Rother, M. (2010). Toyota Kata. McGraw-Hill.

Sailer, M. & Homner, L. (2020). The gamification of learning: A meta-analysis. Educational Psychology Review, 32, 77–112.

Schein, E. H. (2010). Organizational Culture and Leadership (4th ed.). Jossey-Bass.

Senge, P. (1990). The Fifth Discipline. Doubleday.

Sheeran, P., Listrom, O. & Gollwitzer, P. M. (2024). The when and how of planning: Meta-analysis of the scope and components of implementation intentions in 642 tests.

Sisk, V. F., Burgoyne, A. P., Sun, J., Butler, J. L. & Macnamara, B. N. (2018). To what extent and under which circumstances are growth mind-sets important to academic achievement? Two meta-analyses. Psychological Science, 29(4), 549–571.

Stajkovic, A. D. & Luthans, F. (1997). A meta-analysis of the effects of organizational behavior modification on task performance, 1975–95. Academy of Management Journal, 40(5), 1122–1149.

Stajkovic, A. D. & Luthans, F. (1998). Self-efficacy and work-related performance: A meta-analysis. Psychological Bulletin, 124(2), 240–261.

Stouten, J., Rousseau, D. M. & De Cremer, D. (2018). Successful organizational change: Integrating the management practice and scholarly literatures. Academy of Management Annals, 12(2), 752–788.

Sunstein, C. R. (2019/2021). Sludge Audits / Sludge: What Stops Us from Getting Things Done. MIT Press. See also OECD (2024), Fixing Frictions: ‘Sludge Audits’ Around the World.

Szulanski, G. (1996). Exploring internal stickiness: Impediments to the transfer of best practice within the firm. Strategic Management Journal, 17(S2), 27–43.

Taylor, P. J., Russ-Eft, D. F. & Chan, D. W. L. (2005). A meta-analytic review of behavior modeling training. Journal of Applied Psychology, 90(4), 692–709.

Thalheimer, W. (2018). The Learning-Transfer Evaluation Model (LTEM). Work-Learning Research.

Tuckman, B. W. (1965). Developmental sequence in small groups. Psychological Bulletin, 63(6), 384–399.

Venkatesh, V., Morris, M. G., Davis, G. B. & Davis, F. D. (2003). User acceptance of information technology: Toward a unified view. MIS Quarterly, 27(3), 425–478.

Verplanken, B. & Wood, W. (2006). Interventions to break and create consumer habits. Journal of Public Policy & Marketing, 25(1), 90–103.

Weick, K. E. (1995). Sensemaking in Organizations. Sage.

Weick, K. E. & Sutcliffe, K. M. (2007). Managing the Unexpected (2nd ed.). Jossey-Bass.

Weiner, B. J. (2009). A theory of organizational readiness for change. Implementation Science, 4:67. See also Shea, Jacobs, Esserman, Bruce & Weiner (2014), Organizational readiness for implementing change: a psychometric assessment of a new measure (ORIC), Implementation Science, 9:7.

“WEIRD” sampling in organizational research: see analyses of sample diversity finding ~70–80% of samples drawn from the US and Europe, including The “WEIRDEST” Organizations in the World? Assessing the Lack of Sample Diversity in Organizational Research and related work on going beyond WEIRD samples in organizational research.

Wenger, E. (1998). Communities of Practice. Cambridge University Press.

West, R. (2005). Time for a change: putting the Transtheoretical (Stages of Change) Model to rest. Addiction, 100(8), 1036–1039.

Wood, W. & Neal, D. T. (2007). A new look at habits and the habit–goal interface. Psychological Review, 114(4), 843–863. See also Wood & Neal (2016) on habit change interventions.

Zahra, S. A. & George, G. (2002). Absorptive capacity: A review, reconceptualization, and extension. Academy of Management Review, 27(2), 185–203.

---

Appendix A — Verification log

Included because a report that grades other people’s evidence should be transparent about its own. Roughly 60 verification searches were run for v2.

A.1 Directly verified in preparing this document

Key figures, study designs and citations checked against primary or authoritative secondary sources during research:

Claim / citation	Verified detail
Lally et al. 2010	96 participants, 84-day reporting, median 66 days, range 18–254, ~half not automatic at 84 days
Gollwitzer & Sheeran 2006	d = 0.65, 94 independent tests, 8,000+ participants; 2024 update, 642 tests
Kluger & DeNisi 1996	607 effect sizes, 23,663 observations, d = 0.41, >1/3 (≈38%) reduced performance
Stajkovic & Luthans 1998	r = .38, 114 studies, 157 effect sizes, 21,616 participants
Stajkovic & Luthans 1997 (O.B.Mod.)	d = .51, ≈17% performance increase, 1975–95 field research
Jenkins et al. 1998	39 studies, 47 relationships; quantity r = .34 (~12%), quality r = .08 (~3%)
Deci, Koestner & Ryan 1999	128 studies; d = −0.40 / −0.36 / −0.28 by reward contingency; Psych Bulletin 125(6), 627–668
Blume et al. 2010	89 studies; open vs closed skill moderation; same-source inflation
Taylor, Russ-Eft & Chan 2005	117 studies; >1.0 SD vs controls; knowledge decays, skill/behaviour stable or increasing
Lacerenza et al. 2017	335 studies; δ = .63/.73/.82/.72; 15 moderators
Jones, Woods & Guillaume 2016	δ = 0.36 overall; 0.28/0.51/1.24; internal > external coaches; multisource feedback reduces effect
Frazier et al. 2017	136 samples, 22,000+ individuals, ~5,000 groups
Cepeda et al. 2006	839 assessments, 317 experiments, 184 articles; optimal ISI scales with retention interval
Macnamara et al. 2014	26/21/18/4/<1 per cent variance by domain
Neuman, Edwards & Raju 1989	126 studies; multi-faceted > single technique
Oreg, Vakola & Armenakis 2011	1948–2007, 79 studies meeting criteria; tridimensional attitudes
Cummings, Bridgman & Brown 2016	Three-step model post-dates Lewin; “refreezing” not his term
Appelbaum et al. 2012	J. Management Development 31(8), 764–782; support for individual steps, none for whole model; authors’ own conclusion favourable
Hughes 2011	Five sources examined; each asserted or cited without evidence
Stouten, Rousseau & De Cremer 2018	AMA Annals 12, 752–788; ten evidence-based steps
Mertens et al. 2022 / Maier et al. 2022	d = 0.43 pooled; no evidence after bias adjustment; 2025 second-order d = 0.27 → 0.004
Clardy 2018	HRDR 17, 153–178; five literature traditions examined
Pashler et al. 2008	Preferences yes, matching benefit no
Szulanski 1996	271 observations, 122 transfers, 8 companies; knowledge-related barriers dominate
Cohen & Levinthal 1990	ASQ 35(1), 128–152; Zahra & George 2002 four-capability reconceptualisation
March 1991	Organization Science 2, 71–87
Meyer & Rowan 1977	AJS 83, 340–363; decoupling argument
DiMaggio & Powell 1983	Coercive, mimetic, normative isomorphism
Abrahamson	Bell-shaped fashion curves; Quality Circles 1978–86 case
CFIR	39 constructs, 5 domains (2009); 2022 update Implementation Science 17:75; 2025 User Guide
ERIC	73 discrete strategies, 9 clusters, modified-Delphi; Implementation Science 10:21
RE-AIM	Glasgow, Vogt & Boles 1999, AJPH 89(9), 1322–1327
NPT	May & Finch four constructs
Greenhalgh et al. 2004	Milbank Quarterly 82, 581–629
Behaviour Change Wheel	9 intervention functions and 7 policy categories, individually confirmed
BCTTv1	93 techniques, 16 clusters
Pronovost / Michigan	2.7 → 0 per 1,000 catheter-days; 66% at 16–18 months; ~1,500 lives, $175m
Reames et al. 2015	JAMA Surgery 150(3), 208–215; no difference in outcomes
Dixon-Woods et al. 2011	Milbank Quarterly 89(2), 167–205; ex post theory, six theory-derived insights
West 2005	Addiction; arbitrary boundaries, non-sequential transitions
Cochrane stage-based smoking cessation	41 trials, 33,000+ smokers; 4 direct comparisons show no difference
Lundahl et al. 2010	119 studies; g = 0.28 vs weak comparisons, 0.09 vs specific treatments
Credé et al. 2017	584 effect sizes, 88 samples, 66,807 individuals; ρ = .84 with conscientiousness
Sisk et al. 2018 / Macnamara & Burgoyne 2023	273 + 43 studies; 63 intervention studies, null in highest-quality subset
Mattingly & Kraiger 2019	58 studies; moderate positive effect on EI scores; published > dissertation effect
Sailer & Homner 2020	g = 0.49 / 0.36 / 0.25; game fiction and social interaction as moderators
Ego depletion replications	k = 36, N = 3,531, d = 0.06; separate 23-lab, N = 2,141 replication negligible
Venkatesh et al. 2003	~70% variance in intention, ~50% in use; 17–41% for predecessors
Noy & Zhang 2023	453 professionals; 37% faster, 18% higher quality; weaker performers gain most
Dell’Acqua et al.	758 BCG consultants; +12.2% tasks, 25.1% faster inside frontier; −19 percentage points correct outside it
Software-development AI RCTs	4,867 developers across three settings; ≈26% more completed tasks
MIT 2025 GenAI report	52 interviews, 153 leader surveys, 300 deployments; ~95% no measurable P&L impact; ~90% personal-tool use vs ~40% official subscriptions
Holton/Bates LTSI	16 factors; 5,990 participants, 17 countries; factor structure replicated
BBS meta-analysis	~39% pooled reduction; reviewers judged study quality poor
O’Neill / Alcoa	CEO from 1987; lost-workday rate ~1.86 → ~0.2 per 100 workers vs US average ~5
Tuckman 1965	Psych Bulletin 63, 384–399; review of 50 articles across therapy-group, T-group and natural/laboratory settings; adjourning added 1977 with Jensen
Noy & Zhang 2023	Science 381(6654), 187–192; time −40%, quality +18%; 2× real-job use at 2 weeks, 1.6× at 2 months
Greenhalgh et al. 2017 (NASSS)	JMIR 19(11), e367; seven domains classified simple / complicated / complex; NASSS-CAT tools 2020
Venkatesh et al. 2003	MIS Quarterly 27(3), 425–478; eight models integrated
Sisk et al. 2018	Psychological Science 29, 549–571
Abrahamson 1996	Academy of Management Review 21, 254–285
Ford, Ford & D’Amelio 2008	AMR 33(2), 362–377; change agents contribute to resistance; resistance as a resource
Stajkovic & Luthans 1997	AMJ 40(5), 1122–1149; stronger effect in manufacturing organizations
Weiner 2009	Implementation Science 4:67; commitment + efficacy; determinants task demands, resources, situational factors; ORIC measure
Romanelli & Tushman 1994	Transformations rapid and discontinuous; small changes did not accumulate; environmental change and CEO succession as influences
Psychological safety RCT	Castro, Englmaier & Guadalupe; ~7,000 employees, 1,000 teams; one-to-one content and frequency as conduit; low-touch email/PDF intervention
De-implementation	Implementation Science 2024 overview of systematic reviews; effective clusters = adapt/tailor to context, stakeholder interrelationships, infrastructure and workflow change; GROW and Norton frameworks
WEIRD sampling	~70–80% of organizational samples from US/Europe; ~11% of world population
Bonanno	No evidence for five stages; ~65% resilient trajectory in longitudinal bereavement research

A.2 Cited from established background knowledge, not independently re-verified here

Standard, widely reproduced citations not re-checked against source. Treat page numbers and minor details as needing confirmation before formal academic use:

Rogers (2003) edition and pagination; Argyris & Schön publication details; Senge (1990); Bridges’ transitions; Beer & Nohria Theory E/O; Burke-Litwin; McKinsey 7S; Schein (2010) edition; Perrow, Normal Accidents; Revans and action learning; Wenger (1998); Weick (1995) pagination; Thalheimer LTEM (2018); Brinkerhoff (2005) pagination; Zahra & George (2002) pagination; May & Finch (2009) pagination; DiMaggio & Powell (1983) pagination; Jenkins et al. (1998) pagination; Lundahl et al. (2010) pagination; Macnamara & Burgoyne (2023) volume/pagination; Gersick (1991) pagination; Romanelli & Tushman (1994) pagination; Michie et al. BCW book (2014); Cialdini focus theory primary sources; Armenakis/Holt readiness instruments; Hogan & Coote culture test; Norton et al. de-implementation framework details; ORIC psychometric paper details.

A.2b Corrections made during the final verification sweep

Recorded because they are the reason to do the sweep:

1. Psychological safety was wrongly described as intervention-poor. A field RCT (Castro, Englmaier & Guadalupe; ~7,000 employees, 1,000 teams) demonstrates it can be raised through the content and frequency of manager one-to-ones. Corrected in Part 0, §4.2, the team summary table, and the folklore table. This was the most consequential error in the draft, because it changes a recommendation from “create the conditions and hope” to a specific, cheap, testable intervention.
2. Noy & Zhang time saving. Draft said “37% faster”; the published Science abstract reports average time taken decreased by 40%. Corrected, and the paper’s adoption findings (2× real-job use at two weeks, 1.6× at two months) added — they had been omitted entirely and are arguably the more useful result.
3. Tuckman’s source literature. Draft said “dominated by therapy groups”; the 1965 review covered 50 articles across therapy-group, T-group and natural/laboratory settings. Corrected to be accurate rather than merely dismissive.
4. Macnamara & Burgoyne study count. Draft stated 63 studies; sources report different counts depending on outcome subset. Replaced with a description of the analysis plus an explicit note that counts differ, and the Tipton et al. methodological rebuttal was upgraded from a passing mention to a stated position — this is a live dispute, not a closed debunking.
5. Ford, Ford & D’Amelio under-cited. The draft used only the “resistance contains information” half of their argument. The sharper half — that change agents produce resistance through their own actions and inactions — has been added with the full citation.

A.3 Claims deliberately softened or flagged

• Rogers’ five attributes explaining 49–87% of variance. Figure appears in secondary sources; heterogeneity of measurement across categories makes a precise range unreliable. Reported in the text as indicative.
• Prosci “7× more likely.” Retained but graded [D] with the common-method-variance objection stated.
• MIT “95% of AI pilots fail.” Retained with an explicit warning that it should be treated as the new 70% myth — non-peer-reviewed, modest sample, ambiguous criterion.
• “Kotter validated Kotter.” Attributed to secondary critical commentary, not to Appelbaum et al., whose own conclusion was favourable.
• Centola’s network prescriptions. Mechanism graded [B], specific prescriptions [C], with the 2025 conditional findings noted.
• Alcoa financial outcomes. Safety figures reported as solid; financial causal attribution explicitly rejected as narrative.
• Positive deviance. Spread [B], comparative effectiveness [C].
• Change fatigue initiative counts (10–14). Practitioner-sourced, graded [D].

---

Appendix B — What changed from v1

Corrections

1. Appelbaum et al. 2012 mischaracterised. v1 presented the review as a debunking and attributed “Kotter validated Kotter” to the authors. The review’s own conclusion is favourable to Kotter. Corrected, with both readings now stated.
2. Michigan case incomplete. v1 cited Dixon-Woods loosely and omitted the England replication. Added “Explaining Matching Michigan” (2013) and full detail on Keystone Surgery, giving three outcomes from one artefact.
3. File date. v1 was dated 1 August; this version is correctly dated.
4. Kübler-Ross conflation. v1 discussed Bridges without distinguishing it from the change curve. Now separated, with the curve added to the folklore table.

Major additions

• Part 3 — learning separated into its own part; added andragogy, gamification, LTEM, Success Case Method, LTSI, Phillips ROI critique.
• Part 6 — diffusion, spread and scale (Rogers, Greenhalgh 2004, NASSS, Szulanski). A significant gap in v1.
• Part 7 — why change theatre persists (Meyer & Rowan, DiMaggio & Powell, Abrahamson, critical management studies). The explanation for the field’s own pathology, entirely absent from v1.
• Part 8 — reinforcement, incentives and consequences (O.B.Mod., Jenkins, Deci/Koestner/Ryan, behaviour-based safety). A whole evidence base missing from v1.
• Part 10 — technology and AI adoption (TAM/UTAUT, the AI RCTs, the jagged frontier, MIT’s pilot-failure figure, shadow AI as decoupling).
• Part 18 — commercial deployment notes.
• Section 2.3 — breaking habits as a distinct problem; 2.10 transtheoretical model critique; 2.11 motivational interviewing; ego-depletion replication status.
• Section 4.8 — team instrument critiques (Tuckman, MBTI, Belbin/DiSC, SCARF, EI, grit, growth mindset).
• Section 5.6 — organizational learning (Argyris, March, Cohen & Levinthal, Szulanski, Senge).
• Section 5.9 — Lean/Six Sigma and Agile evidence.
• Section 9.3 — fidelity and adaptation.
• Mechanisms expanded from ten to twelve (added absorptive capacity, observability, consequences).
• Folklore table expanded from 18 to 32 entries.
• Gaps expanded from 8 to 11.
• Appendices A and B added.

Added in the final verification sweep

• §1.4 — whose behaviour was actually studied: WEIRD sampling, the Hofstede double-bind, and how to handle generalisability honestly.
• §5.10 — readiness properly specified (Weiner): commitment versus efficacy, the two-by-two, and the most common diagnostic error at that step.
• §5.11 — sensemaking (Weick): why people interpret change by acting in it, and why plausibility beats accuracy.
• §5.12 — pacing and punctuated equilibrium: the finding that small changes do not accumulate into transformation, and how to hold it alongside an otherwise incrementalist report.
• §9.5 — de-implementation: frameworks, what actually works, and why stopping is politically harder than starting.
• §4.2 — the psychological safety RCT.
• Part 0 expanded from 20 to 23 findings; Part 16 gaps from 11 to 14; folklore table from 32 to 37 entries; five corrections logged in Appendix A.2b.

The structural change that matters most: v1 treated change as a three-level stack (individual, team, organization). v2 treats it as a stack plus four cross-cutting problems — how change spreads (Part 6), why it is adopted ceremonially (Part 7), what sustains it (Parts 8–9), and how it is paced and stopped (§5.12, §9.5). Those explain most of the variance the level-based view cannot.

Two things I would still change with more time. First, the report remains heavier on diagnosis than on delivery detail; a companion set of worked protocols (a COM-B diagnostic script, a manager one-to-one restructure, a sludge audit template, a stepped-wedge design note) would make it operational rather than merely correct. Second, several key sources are practitioner or working-paper grade — the psychological safety RCT, the MIT AI figure, Prosci’s benchmarking — and those grades should be revisited as peer-reviewed versions appear.

---

Evidence grades reflect the state of the literature as reviewed on 10 August 2026. Where a source is graded [D], it is included because it represents the largest available dataset in its domain, not because its methods are strong. Where a claim is graded [X], the underlying idea may still be useful heuristically — the grade concerns the evidence, not the utility.