CFIR · ERIC · RE-AIM · NPT: a deep dive, the evidence on what strategies achieve, a translation layer, and a working toolkit

Companion to Learning and Behaviour Change: A Critical Review of the Literature (v2) and to Learning to Fly the Machine That Flies (v2)

About this version. This is version 2, completed on 24 October 2026.

• Re-checked. I re-checked every load-bearing claim in version 1 against its source and corrected 33 of them (Appendix D). One correction withdraws a widely repeated claim: the evidence does not support the idea that bundling more implementation strategies produces bigger effects.
• New. What implementation strategies actually achieve, with effect sizes (§3.7). The wider toolkit beyond the four core frameworks (Part 13). The evidence from schools, universities, health professions education and workplaces (Part 14). A bridge to the aviation report and to AI rollouts (Part 15).

Version 1, from 10 August 2026, stays online unchanged: read version 1.

Contexts: corporate L&D and capability building · higher education · AI capability adoption Units of implementation: a learning programme · a new practice or behaviour · an L&D operating model · a learning technology

---

Bottom line

Learning and development evaluates whether people learned. It almost never evaluates whether the learning was implemented:

• delivered as designed
• to the people who needed it
• taken up by the units and managers meant to carry it
• still alive a year later

Implementation science exists to see that second half. Four frameworks do most of the work:

• CFIR diagnoses what will help or hinder.
• ERIC names what you can do about it.
• RE-AIM reports reach, adoption, fidelity and maintenance against honest denominators.
• NPT explains whether a practice is becoming normal work.

Version 2 adds the evidence v1 lacked. Implementation strategies have modest, measurable effects. They are typically worth single-digit percentage points, they vary widely from trial to trial, and they are not reliably larger when bundled. So the value of these frameworks lies in diagnosis, specification and measurement, not in the size of the menu.

Version 2 also corrects v1 where it was wrong. It maps the wider toolkit, sets out what already exists in schools, universities and workplaces, and connects the whole apparatus to the design kit in the aviation report.

What changed in v2

• Verified. Every load-bearing claim was re-checked. About thirty were corrected or re-scoped. Among them:
  • the CFIR 2.0 counts
  • the full ERIC cluster membership
  • the matching-tool figures
  • RE-AIM’s maintenance convention
  • the NPT 2022 update
  • NoMAD’s figures
  • the Kirk review’s numbers
  • the claim that multifaceted interventions outperform single ones, which the evidence does not support
  • an unsourced critique of NPT, which is removed

  Appendix D lists them all.

• New. §3.7, what strategies actually achieve. Part 13, the wider toolkit. Part 14, what exists in education and workplace learning. Part 15, the bridge to aviation and AI.
• Linked. Every load-bearing citation links to its source.

How to read this document

Four frameworks from implementation science, treated properly: origins, every domain and construct, what an assessment actually looks like, the instruments, the protocols, the limitations, and a translation layer for learning contexts.

Terminology convention. Original construct names are kept throughout, so this stays citable and you can read the source literature without translation loss. Each carries an L&D reading alongside it. Where the healthcare vocabulary is actively misleading in a learning context, that is flagged rather than silently fixed.

Evidence grades carry over from the companion report: [A] meta-analysis or replicated experimental evidence · [B] multiple studies, reasonable methods · [C] single study, conceptual, or contested · [D] practitioner or self-interested source · [X] unsupported.

Verification tags (new in v2) say how a source was checked:

• [VERIFIED]: full text or official page read
• [ABSTRACT]: abstract only
• [SECONDARY]: reported by another source
• [UNVERIFIED]: cited but not checked

v1 claims were re-checked and carry no tag. Where they changed, Appendix D says how.

A note on what these frameworks are and are not. None of them is an intervention. None will tell you an effect size (§3.7 reports what the strategies achieve). They are structured attention — they make you look at things you would otherwise skip, in an order that reduces the chance of expensive mistakes. That is a smaller claim than the field sometimes makes for them, and it is still more than any commercial change methodology offers.

Navigation.

• Part 0 is the map.
• Parts 2–5 are the frameworks in depth. §3.7 gives the effect sizes.
• Part 7 is the learning translation, and contains the single most important idea in this document.
• Part 9 is the instruments.
• Part 11 is the honest account of limitations.
• Parts 13–15 are new in v2: the wider toolkit, the evidence from education and workplaces, and the aviation bridge.
• Part 16 is the reading list.
• Appendices A and B are the full strategy and construct lists.
• Appendix C is the verification log; Appendix D lists what changed from v1.

---

Part 0 — The map

0.1 What each one is for

Framework	Type	Answers	Use it when
CFIR	Determinant framework	What will help or hinder this?	Before you design. Diagnosing context.
ERIC	Strategy compilation	What can we actually do about it?	After diagnosis. Selecting and naming interventions.
NPT	Action theory	Is this becoming normal work — and if not, where is it failing?	During and after. Explaining embedding or its absence.
RE-AIM	Evaluation framework	Did it reach, work, get adopted, get delivered, and last?	Planning and evaluation. Reporting honestly.

They are not competitors. They occupy different slots in the same workflow, and the workflow is the point:

CFIR tells you what you’re up against → ERIC gives you a named menu of responses → NPT tells you whether the practice is becoming routine → RE-AIM tells you whether any of it mattered, at what reach, and whether it survived.

0.2 The one thing to take from this document

Learning and development almost universally evaluates the innovation and almost never evaluates the implementation.

• Did they learn it? → innovation effectiveness.
• Did the programme actually get delivered as designed, to the people intended, by the people meant to deliver it, and is it still running in a year? → implementation.

Kirkpatrick, LTEM, ROI methodologies and most L&D evaluation practice sit entirely in the first column. The second column is where programmes die, and it is invisible in the standard toolkit. Everything else in this document follows from that gap.

0.3 Fast facts

CFIR	Damschroder et al. 2009; major update 2022; Outcomes Addendum 2022; User Guide 2025. Five domains (one with two subdomains); 48 constructs and 19 subconstructs in the updated version.
ERIC	Powell et al. 2015. 73 discrete strategies from a three-round modified Delphi. Waltz et al. 2015 grouped them into 9 clusters by concept mapping. CFIR–ERIC matching tool, Waltz et al. 2019.
RE-AIM	Glasgow, Vogt & Boles 1999, AJPH 89(9). Five dimensions across two levels. Cited in more than 2,800 publications by 2019 (Glasgow et al. 2019). PRISM (Feldstein & Glasgow 2008) adds four contextual domains.
NPT	May & Finch 2009, Sociology 43(3). Four constructs, each with four sub-constructs. NoMAD instrument: 23 items, validated 2018.

---

Part 1 — The landscape these frameworks sit in

1.1 Five kinds of thing, routinely confused

Implementation science distinguishes types of theoretical approach that practitioners tend to blur. The five-way classification below is Nilsen’s (2015), the standard map of the field. Getting this right prevents the most common misuse — expecting a framework to do a job it is not built for.

Process models describe the stages of translating research into practice. They are sequences. (Knowledge-to-Action, the Quality Implementation Framework, EPIS.)

Determinant frameworks specify classes of factors that influence implementation outcomes. They are checklists of what to assess. They describe rather than explain, and they do not predict. CFIR is here.

Classic theories are imported from other disciplines — diffusion of innovations, theory of planned behaviour, organizational learning theory.

Implementation theories are developed specifically to explain implementation, with stated mechanisms. NPT is here, and this is why NPT behaves differently from the other three: it makes claims about how embedding happens, not just what is associated with it.

Evaluation frameworks specify what to measure to judge success. RE-AIM is here.

ERIC is none of these. It is a compilation — a controlled vocabulary of actions. It is closer to a taxonomy than a theory, and it should be read as the implementation-level equivalent of the BCT Taxonomy v1 at individual level.

The practical consequence: a determinant framework will never tell you what to do (that’s ERIC), an evaluation framework will never tell you why something failed (that’s NPT), and none of them will tell you whether your intervention content is any good (that’s your own evidence base — see the companion report).

1.2 Implementation outcomes vs innovation outcomes

The distinction the CFIR Outcomes Addendum (2022) exists to make, and the one that matters most for learning.

Proctor et al.’s taxonomy (2011) proposed eight conceptually distinct implementation outcomes [B]:

Outcome	Definition	Learning translation
Acceptability	Perception that the innovation is agreeable, palatable, satisfactory	Do managers and learners find the practice reasonable?
Adoption	Intention, initial decision, or action to try it	How many teams/faculties/units actually took it up?
Appropriateness	Perceived fit, relevance, compatibility for a setting or problem	Does this fit how this group actually works?
Feasibility	Extent to which it can actually be carried out	Can a manager really do this in the time available?
Fidelity	Delivered as intended	Was the programme delivered as designed? Were the active ingredients present?
Implementation cost	Cost of the delivery effort	True cost including manager and learner time, not just vendor fees
Penetration	Integration within a setting and its subsystems	What proportion of eligible people/teams are covered?
Sustainability	Maintained and institutionalised	Is it still happening at 12 months without the programme team?

A scoping review of ten years of implementation-outcomes research found acceptability, appropriateness, feasibility and adoption dominate early exploration and preparation phases, while cost, fidelity, penetration and sustainability appear more in later implementation and sustainment phases [B] — which is a useful sequencing hint in itself.

The CFIR Outcomes Addendum adds three distinctions worth carrying:

1. Anticipated vs actual implementation outcomes. Anticipated outcomes are perceptions or measures of the likelihood of future success — appropriate before or early in implementation. Actual outcomes are retrospective measures of what happened. Most L&D “evaluation” collects anticipated outcomes (intention to apply) and reports them as actual ones.
2. Implementation vs reach vs innovation effectiveness. Three different things, routinely merged into one satisfaction number.
3. Implementation determinants vs innovation effectiveness determinants. CFIR constructs primarily influence implementation outcomes, not whether the innovation itself works. If your programme content is weak, no amount of implementation excellence rescues it — and CFIR will not tell you the content is weak.

For L&D this is the whole game. A leadership programme can have excellent learning outcomes (innovation effectiveness), 8% reach, 30% adoption across business units, poor delivery fidelity in the units that did adopt, and zero sustainability — and be reported internally as a success, because only the first was measured.

---

Part 2 — CFIR

2.1 What it is and where it came from

The Consolidated Framework for Implementation Research was published by Damschroder and colleagues in Implementation Science in 2009 (4:50). It was a consolidation: the authors reviewed existing implementation theories and synthesised overlapping constructs into a single organised framework. That origin explains both its strength — comprehensiveness — and its weakness: it is a superset of other people’s constructs rather than a theory with mechanisms.

The original version specified 39 constructs across five domains. Following a decade of use, Damschroder, Reardon, Opra Widerquist and Lowery published an updated version in 2022 (Implementation Science 17:75), based on a systematic literature review plus a survey of authors who had published CFIR studies. The update revised, added, removed and relocated constructs; centred innovation recipients; and added determinants of equity. Constructs map back to the original for longitudinal consistency.

Two companion publications matter:

• CFIR Outcomes Addendum (2022) — the conceptual distinctions in §1.2.
• CFIR User Guide (2025) — a five-step guide for conducting implementation research with the framework.

2.2 The five domains

The updated framework is organised as concentric circles: the Innovation sits at the centre, surrounded by the Individuals who do the work, within an Inner Setting, within an Outer Setting, with the Implementation Process cutting across.

A structural note that trips people up: the Innovation domain in the updated version is stated as “the thing being implemented, e.g., a new clinical treatment, educational program, or city service.” Note that educational programmes are explicitly named. The framework was built with settings like schools and city services in mind, not only clinics.

Domain I — Innovation

The thing being implemented. Eight constructs. This domain is essentially Rogers’ attributes of innovations, formalised and extended.

Construct	Definition (the degree to which:)	L&D reading
Innovation Source	The group that developed and/or visibly sponsored use of the innovation is reputable, credible, trustable	Who is this programme from? A respected internal faculty? An unknown vendor? A mandate from a function with poor credibility?
Innovation Evidence-Base	The innovation has robust evidence supporting its effectiveness	Does the learning design rest on anything? (Most doesn’t — see companion report Part 3.)
Innovation Relative Advantage	It is better than other available innovations or current practice	Better for the person being asked to do it, not for the sponsor. The most common fatal error.
Innovation Adaptability	It can be modified, tailored, or refined to fit local context	Can a business unit or department adjust it without breaking it?
Innovation Trialability	It can be tested or piloted on a small scale and undone	Note “and undone.” Reversibility is a distinct property and almost never designed for.
Innovation Complexity	It is complicated — scope, number and nature of connections and steps	How many other things must change for this to work?
Innovation Design	It is well designed and packaged — assembled, bundled, presented	Materials, platform, look and feel, how it arrives. Underrated.
Innovation Cost	Purchase and operating costs are affordable	Including the operating cost in manager and learner time.

Domain II — Outer Setting

The setting in which the Inner Setting exists — sector, system, region, market. Seven constructs, one with three subconstructs.

Construct	Definition	L&D reading
Critical Incidents	Large-scale/unanticipated events disrupt implementation or delivery	Restructure, funding cut, acquisition, sector crisis, a new regulator
Local Attitudes	Sociocultural values and beliefs encourage the Outer Setting to support it	Sector norms about development, professional identity, attitudes to formal learning
Local Conditions	Economic, environmental, political, technological conditions enable support	Labour market, funding climate, technology maturity
Partnerships & Connections	The Inner Setting is networked with external entities	Awarding bodies, universities, professional bodies, provider networks
Policies & Laws	Legislation, regulation, professional guidelines, accreditation standards support it	Directly relevant to standards-driven training and HE quality regimes
Financing	External funding is available	Grants, levies, subsidies, sector funding schemes
External Pressure	External pressures drive implementation	Umbrella construct for the three below
↳ Societal Pressure	Mass media, advocacy groups, social movements	Public expectations, e.g. around AI, safety, inclusion
↳ Market Pressure	Competing with and/or imitating peer entities	“Everyone else has an AI academy.” Straight isomorphism — see companion report Part 7
↳ Performance-Measurement Pressure	Quality or benchmarking metrics or established goals	League tables, NSS/TEF-type regimes, benchmarking, audit

Domain III — Inner Setting

The setting in which the innovation is implemented — the organization, business unit, faculty, school, team. Eleven constructs, several with subconstructs. Crucially, the framework distinguishes constructs A–D, which exist regardless of the innovation, from E–K, which are specific to it.

Persistent characteristics (A–D):

Construct	Definition	L&D reading
Structural Characteristics	Infrastructure supports functional performance	Umbrella for the three below
↳ Physical Infrastructure	Layout, configuration of space, tangible features	Rooms, labs, simulators, workshop space, campus geography
↳ Information Technology Infrastructure	Telecoms, electronic documentation, data storage/management/reporting	LMS/LXP, SSO, data plumbing, whether anyone can actually get a report
↳ Work Infrastructure	Organization of tasks and responsibilities within and between individuals and teams; staffing levels	Shift patterns, teaching loads, utilisation targets, whether there is any slack at all
Relational Connections	High-quality formal and informal relationships, networks and teams within and across boundaries	Whether L&D has real relationships with the line, or only a service catalogue
Communications	High-quality formal and informal information-sharing within and across boundaries	Does anything travel between units?
Culture	Shared values, beliefs and norms	Umbrella for the four below
↳ Human Equality-Centeredness	Shared beliefs about the inherent equal worth of all human beings	Added in the 2022 equity update
↳ Recipient-Centeredness	Norms around caring for and addressing the needs of recipients	Is the learner’s actual situation considered, or is the catalogue the point?
↳ Deliverer-Centeredness	Norms around caring for and addressing the needs of deliverers	Are trainers, facilitators and managers supported, or used?
↳ Learning-Centeredness	Norms around psychological safety, continual improvement, and using data to inform practice	The single most important construct in CFIR for L&D purposes. See §2.7.

Innovation-specific characteristics (E–K):

Construct	Definition	L&D reading
Tension for Change	The current situation is intolerable and needs to change	Is there felt pain, or is this a nice-to-have? Determines everything downstream
Compatibility	Fits with workflows, systems and processes	Does the practice fit the actual working week?
Relative Priority	Implementing it is important compared to other initiatives	The portfolio question. Usually the real answer to “why didn’t it land”
Incentive Systems	Tangible/intangible incentives, rewards, disincentives, punishments support it	What happens to someone who does it? What happens to someone who doesn’t?
Mission Alignment	It is in line with the overarching purpose or goals of the Inner Setting	Does this connect to what the unit is actually for?
Available Resources	Resources are available to implement and deliver	Umbrella for the three below
↳ Funding	Funding is available
↳ Space	Physical space is available
↳ Materials & Equipment	Supplies are available
Access to Knowledge & Information	Guidance and/or training is accessible to implement and deliver	Note: this is about supporting the implementers, not the learners

Domain IV — Individuals

Two subdomains. This is the biggest structural change in the 2022 update, and it is a significant improvement.

Roles subdomain — nine roles:

Role	Definition	L&D reading
High-level Leaders	High authority: key decision-makers, executives, directors	Exec sponsor, dean, C-suite
Mid-level Leaders	Moderate authority: supervised by a high-level leader, supervise others	Line managers, heads of department. Usually the binding constraint
Opinion Leaders	Informal influence on attitudes and behaviours of others	The person people actually ask. Rarely on the org chart
Implementation Facilitators	Subject matter experts who assist, coach or support implementation	Internal coaches, learning partners, educational developers
Implementation Leads	Individuals who lead efforts to implement	The programme lead
Implementation Team Members	Collaborate with and support the leads, ideally including deliverers and recipients	Note the “ideally” — recipient representation on the team is a design recommendation
Other Implementation Support	Individuals who support leads and/or team members	PMO, comms, IT, HR ops
Innovation Deliverers	Directly or indirectly delivering the innovation	Trainers, facilitators, faculty, and — critically — line managers
Innovation Recipients	Directly or indirectly receiving the innovation	Learners; and indirectly, their teams and customers

Characteristics subdomain — four constructs, applied to each role:

Construct	Definition	Note
Need	The individual has deficits related to survival, well-being or personal fulfilment which the innovation addresses
Capability	Interpersonal competence, knowledge and skills to fulfil their Role	This is COM-B
Opportunity	Availability, scope and power to fulfil their Role
Motivation	Committed to fulfilling their Role

This is a significant and under-appreciated feature. CFIR 2.0 embeds COM-B inside the Individuals domain and applies it per role. The 2022 paper says so directly: the characteristics are based on the COM-B system “or role-specific theories” (Damschroder et al. 2022). You do not run one motivation assessment; you ask whether the mid-level leader has the capability, opportunity and motivation to fulfil the mid-level leader role in this implementation, and separately whether the deliverer does, and separately whether the recipient does. That is a far sharper diagnostic than “stakeholder engagement,” and it links CFIR directly to the individual-level behaviour change evidence in the companion report.

Domain V — Implementation Process

The activities and strategies used to implement. Nine constructs, definitions framed as “the degree to which individuals…”:

Construct	Definition	L&D reading
Teaming	Join together, intentionally coordinating and collaborating on interdependent tasks	Is there a real implementation team, or a project manager and a steering group?
Assessing Needs	Collect information about priorities, preferences and needs of people	Umbrella for the two below
↳ Innovation Deliverers	…of deliverers, to guide implementation and delivery	Almost never done in L&D. Nobody asks the managers what they need
↳ Innovation Recipients	…of recipients, to guide implementation and delivery	The learning needs analysis — but about implementation, not content
Assessing Context	Collect information to identify and appraise barriers and facilitators	This is the CFIR assessment itself
Planning	Identify roles and responsibilities, outline steps and milestones, define goals and success measures in advance
Tailoring Strategies	Choose and operationalize implementation strategies to address barriers, leverage facilitators, fit context	This is where ERIC enters
Engaging	Attract and encourage participation	Umbrella for the two below
↳ Innovation Deliverers	Attract and encourage deliverers to serve on the team and/or deliver
↳ Innovation Recipients	Attract and encourage recipients to serve on the team and/or participate	Note: participation on the team, not just in the programme
Doing	Implement in small steps, tests or cycles of change to trial and cumulatively optimize delivery	PDSA logic
Reflecting & Evaluating	Collect and discuss quantitative and qualitative information about success	Umbrella for the two below
↳ Implementation	…about the success of implementation	The missing half of L&D evaluation
↳ Innovation	…about the success of the innovation	The half L&D does
Adapting	Modify the innovation and/or the Inner Setting for optimal fit and integration	Note “and/or the Inner Setting” — you are permitted to change the organization, not only the programme

2.3 What a CFIR assessment actually looks like

The gap between “we used CFIR” and a real CFIR assessment is large. Here is the real shape.

Step 1 — Define the innovation precisely. One sentence. If you cannot say what is being implemented in one sentence, stop. In learning contexts the most common failure is defining the innovation as “the programme” when what you actually need adopted is a practice (see Part 7).

Step 2 — Define the Inner Setting and Outer Setting explicitly. The framework notes there may be multiple Inner Settings and multiple levels (unit, classroom, team) and multiple Outer Settings. Write down which level you are assessing. Assessments that float between “the organization” and “the team” produce mush.

Step 3 — Identify who has power or influence over implementation outcomes. The CFIR guidance is explicit that data should be collected from individuals with power and/or influence over implementation outcomes. Use the Roles subdomain as your sampling frame: high-level leaders, mid-level leaders, opinion leaders, deliverers, recipients. In practice: 10–20 semi-structured interviews for a single-site assessment; 5–8 per site for a multi-site comparison.

Step 4 — Select constructs and justify the selection. You are not expected to assess all 48. Kirk et al.’s systematic review found that construct cherry-picking without justification was a major weakness in published use. Choose deliberately, write down why, and record which constructs you excluded.

Step 5 — Build the interview guide from construct definitions. Each construct definition is written as “the degree to which…”, which converts directly into a question. Instrument in Part 9.

Step 6 — Code and rate. The classic CFIR practice is to rate each construct for valence (positive/negative influence) and strength (weak/strong), typically on a −2 to +2 scale, with 0 for neutral or mixed and blank for absent. Ratings are assigned per site, by at least two coders, with disagreements resolved by discussion.

Step 7 — Compare high and low performers. The most powerful CFIR design is comparative: rate constructs across sites that achieved different implementation outcomes, and identify which constructs discriminate. This is what the Friendship Bench work in Zimbabwe did — RE-AIM to classify performance, CFIR to explain the difference, CFIR–ERIC to generate targeted strategies. Without an implementation outcome to compare against, you have a description, not a diagnosis. Fifteen of the 26 studies in Kirk’s review had this problem.

Step 8 — Feed the discriminating barriers into ERIC. Part 3.

2.4 Rating scheme

Rating	Meaning
+2	Construct is a strong positive influence; clear, consistent evidence across informants
+1	Weak positive influence; some supporting evidence or mixed with positive lean
0	Neutral, or evidence is genuinely mixed with no lean
−1	Weak negative influence
−2	Strong negative influence; clear, consistent evidence
(blank)	Not rated: the construct was not asked about, or the answers did not correspond to it. Record separately from 0, because “not covered” and “mentioned as neutral” mean different things

Source (v2): the −2 to +2 scheme comes from Damschroder & Lowery (2013) and is restated in the 2025 CFIR User Guide [VERIFIED]. In the 2013 rules, a +2 or −2 needs at least two interviewees giving explicit examples.

Record for each rating: the construct, the rating, at least one verbatim quote, and the number of informants who raised it. That triple is what makes the assessment defensible when someone challenges it.

2.5 Limitations

It is a determinant framework, so it describes rather than explains [A]. CFIR tells you what is associated with implementation difficulty. It does not specify mechanisms, does not weight constructs, and does not predict. Two barriers rated −2 are not necessarily equally important, and CFIR gives you no way to say which matters more. Judgement fills the gap, and judgement is not evenly distributed.

Quality of use in the published literature is poor. Kirk et al.’s 2016 systematic review of 26 studies found only 11 (42 per cent) assessed implementation outcomes, which fundamentally limits the ability to identify determinants — you cannot say what caused a difference if you have not measured the difference. They also found gaps in depth of use, wide variation in which constructs were selected, and an absence of justification for those selections, limiting comparability. Their four recommendations remain the best quality checklist: apply across multiple phases; be transparent about construct selection; identify and assess implementation outcomes; integrate the framework into data collection, analysis and interpretation rather than bolting it on at write-up.

It is heavy. Forty-eight constructs is a lot of attention. Pragmatic adaptations exist — the “CFIR simplified” work in a learning health system is one published example — and are legitimate, provided the simplification is documented.

Cultural transferability is not established. A systematic review evaluating CFIR for use in low- and middle-income countries found the need for optimisation. The framework’s assumptions about organizational structure, roles and data availability are not universal — relevant given the WEIRD sampling problem discussed in the companion report.

It says nothing about whether your innovation is any good. The Outcomes Addendum is explicit that CFIR constructs primarily influence implementation outcomes, not innovation effectiveness. A perfectly implemented weak programme is a well-executed waste.

2.6 The equity dimension

The 2022 update deliberately added determinants of equity and centred innovation recipients. Two constructs carry most of this: Human Equality-Centeredness in the Inner Setting Culture subconstructs, and the elevated treatment of Innovation Recipients in both the Roles subdomain and the Assessing Needs / Engaging process constructs.

For learning this is directly usable and rarely used. The question “which recipients are being systematically underserved by how this is being implemented?” maps onto Reach in RE-AIM and to representativeness, and it is the question that distinguishes a capability programme that works from one that works for the people who were already doing fine.

2.7 The construct L&D should notice first

Inner Setting → Culture → Learning-Centeredness: “There are shared values, beliefs, and norms around psychological safety, continual improvement, and using data to inform practice.”

Three observations:

1. Implementation science has independently arrived at psychological safety as a contextual determinant of implementation success, not merely as a team-performance variable. That converges with the meta-analytic evidence and with the large field experiment discussed in the companion report, which is still a working paper (Castro, Englmaier & Guadalupe 2022).
2. It is framed as a persistent characteristic of the Inner Setting (constructs A–D), meaning it exists before your programme arrives and will shape it.
3. For an L&D practitioner this is the strongest available argument that psychological safety work is not a soft add-on but a precondition for any capability initiative landing — and the CFIR construct gives you a defensible place to say so in a diagnostic report.

---

Part 3 — ERIC

3.1 What it is and where it came from

The Expert Recommendations for Implementing Change project addressed a specific problem: implementation strategies were being described in the literature with inconsistent, overlapping and poorly described labels, making it impossible to compare studies or accumulate knowledge.

Powell and colleagues convened a broad group of implementation and clinical experts in a three-phase, modified-Delphi process that refined and expanded an earlier list into 73 discrete implementation strategies, each with a name and a definition (Implementation Science, 2015, 10:21).

Waltz and colleagues then used concept mapping to characterise relationships among the 73 strategies and to rate their importance and feasibility, producing nine clusters via hierarchical cluster analysis (Implementation Science, 2015, 10:109). The algorithm produced 13 clusters; the team reduced them to nine by judgement, merging two to form “Develop stakeholder interrelationships” and moving three strategies: Start a dissemination organization (into Change infrastructure), Organize clinician implementation team meetings (into Develop stakeholder interrelationships) and Remind clinicians (into Support clinicians). The final nine are therefore an expert-curated grouping, not a raw statistical result.

The full 73 with definitions and L&D translations are in Appendix A. This section covers the clusters, the matching tool, and how to use the compilation without misusing it.

3.2 The nine clusters

The table gives the full published membership of each cluster (Waltz et al. 2015, Table 1), with the experts’ mean importance and feasibility ratings for each cluster.

#	Cluster (strategies)	Members (Powell 2015 numbers)	Expert ratings: importance / feasibility (1–5)	L&D reading
1	Use evaluative and iterative strategies (10)	Assess for readiness and identify barriers and facilitators (4); Audit and provide feedback (5); Purposely reexamine the implementation (56); Develop and implement tools for quality monitoring (26); Develop and organize quality monitoring systems (27); Develop a formal implementation blueprint (23); Conduct local needs assessment (18); Stage implementation scale up (61); Obtain and use patients/consumers and family feedback (46); Conduct cyclical small tests of change (14)	4.19 / 4.01 (highest on both)	The diagnostic and improvement engine. Almost entirely absent from L&D practice
2	Provide interactive assistance (4)	Facilitation (33); Provide local technical assistance (54); Provide clinical supervision (53); Centralize technical assistance (8)	3.67 / 3.29	The “someone helps you do it” cluster. Maps to coaching
3	Adapt and tailor to context (4)	Tailor strategies (63); Promote adaptability (51); Use data experts (67); Use data warehousing techniques (68)	3.59 / 3.30	Core-function specification; local fit
4	Develop stakeholder interrelationships (17)	Identify and prepare champions (35); Organize clinician implementation team meetings (48); Recruit, designate, and train for leadership (57); Inform local opinion leaders (38); Build a coalition (6); Obtain formal commitments (47); Identify early adopters (36); Conduct local consensus discussions (17); Capture and share local knowledge (7); Use advisory boards and workgroups (64); Use an implementation advisor (65); Model and simulate change (45); Visit other sites (72); Involve executive boards (40); Develop an implementation glossary (25); Develop academic partnerships (24); Promote network weaving (52)	3.47 / 3.64	The largest cluster, and the one the evidence in the companion report most strongly supports
5	Train and educate stakeholders (11)	Conduct ongoing training (19); Provide ongoing consultation (55); Develop educational materials (29); Make training dynamic (43); Distribute educational materials (31); Use train-the-trainer strategies (71); Conduct educational meetings (15); Conduct educational outreach visits (16); Create a learning collaborative (20); Shadow other experts (60); Work with educational institutions (73)	3.43 / 3.93	Where L&D already lives: one cluster of nine, 11 strategies of 73
6	Support clinicians (5)	Facilitate relay of clinical data to providers (32); Remind clinicians (58); Develop resource sharing agreements (30); Revise professional roles (59); Create new clinical teams (21)	3.23 / 3.06	Read as “support deliverers.” Job redesign, prompts, real-time data
7	Engage consumers (5)	Involve patients/consumers and family members (41); Intervene with patients/consumers to enhance uptake and adherence (39); Prepare patients/consumers to be active participants (50); Increase demand (37); Use mass media (69)	3.25 / 2.95	Read as “engage learners and their managers as active participants, not audiences”
8	Utilize financial strategies (9)	Fund and contract for the clinical innovation (34); Access new funding (1); Place innovation on fee for service lists/formularies (49); Alter incentive/allowance structures (2); Make billing easier (42); Alter patient/consumer fees (3); Use other payment schemes (70); Develop disincentives (28); Use capitated payments (66)	2.86 / 2.09	Least transferable cluster; healthcare-payment specific. But “alter incentive structures” and “develop disincentives” transfer directly
9	Change infrastructure (8)	Mandate change (44); Change record systems (12); Change physical structure and equipment (11); Create or change credentialing and/or licensure standards (22); Change service sites (13); Change accreditation or membership requirements (9); Start a dissemination organization (62); Change liability laws (10)	2.40 / 2.01 (lowest on both)	The structural cluster. My judgement is that the most durable interventions live here; Waltz’s experts rated it least important and least feasible, and that divergence is itself worth discussing with a client

The distribution is the finding. Training and education is one cluster of nine, holding 11 of the 73 strategies (15 per cent). Experts rated it high on importance and the most feasible cluster after the evaluative one, which is exactly why plans default to it. If your implementation plan consists entirely of cluster 5, you are using a seventh of the available strategies, and the transfer evidence in the companion report says it is not the most powerful seventh.

3.3 The strategies L&D most reliably under-uses

From the full 73, these are the ones with the strongest fit to learning problems and the weakest presence in actual L&D practice:

Strategy	ERIC definition (abbreviated)	Why it matters for learning
Assess for readiness and identify barriers and facilitators	Assess an organization’s degree of readiness, barriers that may impede, strengths that can be used	The step before design. Weiner’s commitment/efficacy distinction operationalised
Conduct cyclical small tests of change	Implement changes in a cyclical fashion using small tests before going system-wide, with systematic measurement	Trialability, engineered
Develop a formal implementation blueprint	Aim/purpose, scope of change, timeframe and milestones, performance measures — used and updated over time	Distinct from a project plan: it is about implementation, not delivery
Audit and provide feedback	Collect and summarise performance data over a period and give it to practitioners and administrators	Task-directed feedback at system level. The largest evidence base of any strategy, with a modest average effect (§3.7)
Identify and prepare champions	Identify and prepare individuals who dedicate themselves to supporting and driving an implementation, overcoming indifference or resistance	Note “prepare” — champions are usually named and abandoned
Identify early adopters	Identify early adopters at the local site to learn from their experiences	Positive deviance by another name
Capture and share local knowledge	Capture how implementers made something work in their setting, then share it with other sites	Direct attack on causal ambiguity (Szulanski)
Promote adaptability	Identify how the innovation can be tailored to local needs and clarify which elements must be maintained to preserve fidelity	The core/adaptable specification, in one strategy
Promote network weaving	Build on existing high-quality working relationships and networks to promote information sharing and shared vision	Complex contagion, operationalised
Shadow other experts	Provide ways for key individuals to directly observe experienced people using the practice change	Vicarious mastery experience — the second strongest self-efficacy source
Organize implementer team meetings	Develop and support teams who are implementing, with protected time to reflect, share lessons and support each other’s learning	Note “protected time.” The detail that makes it real
Revise professional roles	Shift and revise roles among practitioners, redesign job characteristics	Structural. Almost never on an L&D plan
Change record systems	Change record systems to allow better assessment of implementation or outcomes	If you cannot see the behaviour, you cannot manage it
Stage implementation scale up	Phase efforts starting with small pilots or demonstration projects, gradually moving to system-wide rollout	The stepped wedge, in strategy form
Alter incentive/allowance structures	Work to incentivize adoption and implementation	Implementation trials are inconclusive (§3.7); the motivation literature argues for informational and social before material
Make training dynamic	Vary information delivery methods to cater to different learning styles and work contexts, and shape training to be interactive	⚠ See §3.6 — this strategy contains a documented myth

3.4 The CFIR–ERIC Matching Tool

Waltz and colleagues (2019) built the bridge between diagnosis and action.

How it was made. 435 implementation experts were invited and 169 (39 per cent) took part. Each selected and ranked the top seven ERIC strategies they believed would best address a CFIR barrier, covering about six barriers each (Waltz et al. 2019). The tool reports endorsement levels:

• Level 1 — more than 50% of experts ranked the strategy in their top seven for that barrier.
• Level 2 — 20 to 49.9 per cent did.

Consensus was rare: only 39 of 1,832 endorsed barrier–strategy pairs (2.1 per cent) reached Level 1, and 21 of the 39 barriers had no Level 1 strategy at all.

How you use it. Select the CFIR barriers your assessment identified; the tool returns matched ERIC strategies with endorsement levels.

How to use it responsibly, because the caveats are substantial:

• It is expert opinion, not effectiveness data. Waltz et al. themselves advise caution due to lack of consensus across experts.
• Each CFIR barrier had an average of 47 different strategies endorsed by at least one expert. That is not a recommendation engine; it is a structured brainstorm with weighting. Note also that the tool was built on the 39 constructs of the original CFIR, not CFIR 2.0.
• A single strategy may address multiple barriers depending on how it is operationalised — which means the operationalisation carries the weight, not the strategy name.
• A 2023 mixed-methods study in the Veterans Health Administration compared the tool’s recommendations with real-world strategy effectiveness data (Yakovchenko et al. 2023). Sites did not use the recommended strategies more often than others. But 70 per cent of the top-20 recommended strategies were positively associated with implementation outcomes, against 48 per cent of the strategies sites actually used. That is modest support for the tool’s validity, not proof of effectiveness.
• A 2026 scoping review of 53 studies found the commonest reported challenges were context-specific adaptation (24.5 per cent) and subjective bias from relying on expert consensus (17.0 per cent). Only 15.7 per cent of studies validated, and 7.8 per cent evaluated, the strategies the tool generated (Meng et al. 2026) [B].

Verdict: use it to broaden your candidate set and to force explicit reasoning from barrier to action. Do not use it as an oracle, and never present its output as evidence of effectiveness.

3.5 Specifying a strategy properly

Naming an ERIC strategy is the beginning, not the end. Proctor, Powell and McMillen (2013) recommend specifying seven dimensions. (The formal reporting standard for implementation studies is StaRI; Pinnock et al. 2017.) Without these, “we did facilitation” is uninterpretable and unrepeatable.

Element	Question	Learning example
Actor	Who enacts the strategy?	Learning partner, not “L&D”
Action	What precisely do they do?	Run a 30-minute structured pre-brief with each line manager
Action target	What/whom is it aimed at, and which determinant?	Mid-level leaders; targets Relative Priority and Opportunity
Temporality	When, in what order?	Two weeks before cohort start, before invitations go out
Dose	How much, how often, how long?	Once per manager, 30 minutes, plus one 15-minute follow-up at week 4
Implementation outcome affected	Which outcome should move?	Fidelity of manager enabling actions; adoption at team level
Justification	Why this, on what basis?	Transfer literature: manager pre-brief and opportunity to perform

Adopt this table as a standing requirement in any implementation plan you write. It is the single highest-leverage documentation discipline available, it takes ten minutes per strategy, and it makes an evaluation interpretable a year later when nobody remembers what was done.

3.6 Limitations

It is a vocabulary, not evidence. The 73 strategies were derived by expert consensus for conceptual distinctness, not selected for effectiveness. Importance and feasibility ratings reflect expert belief. Effectiveness evidence for individual strategies exists in varying quality elsewhere and is not part of ERIC.

It is healthcare-shaped. Cluster 8 (financial strategies) is largely untransferable outside healthcare payment systems. Several strategies use “clinician,” “patient/consumer,” “clinical innovation” in their definitions. Translation is required and should be documented rather than assumed.

One strategy embeds a debunked idea. “Make training dynamic” is defined as varying delivery methods “to cater to different learning styles and work contexts.” The learning-styles matching hypothesis has no adequate supporting evidence [X] — see the companion report. The rest of the strategy (vary methods, make it interactive, fit the work context) is defensible on other grounds. Use the strategy, reject the stated rationale, and be able to explain why. This is a useful reminder that implementation science is not immune to the folklore problem.

Strategies are frequently under-specified in practice, which is why §3.5 exists. Reviews of implementation strategy reporting often find specification inadequate.

Bundling is under-theorised, and more is not reliably better. Real implementations use multiple strategies, and which combinations work is largely unknown. The common assumption that more components mean a bigger effect is not supported. An overview of 25 systematic reviews found no compelling evidence that multifaceted interventions outperform single-component ones (Squires et al. 2014). An earlier review of 235 studies found no consistent relationship between the number of components and the size of effect; formal meta-regression was not feasible (Grimshaw et al. 2004) [B]. Choose strategies because each answers a diagnosed barrier, not to make the bundle look thorough. (The companion review’s statement to the contrary rests on an organisation-development attitudes meta-analysis and is flagged for correction in its next revision.)

3.7 What the strategies actually achieve

ERIC names 73 strategies. The evidence on what they achieve is concentrated on about a dozen, almost all tested on clinicians’ professional practice. It says three things:

1. Effects are small. Typically 3–11 percentage points in the share of professionals doing the desired thing.
2. They vary enormously. A typical trial of any one strategy has a fair chance of finding nothing.
3. Adding components does not reliably add effect.

That should change both what you promise a client and how you design. A strategy is a bet with a modest expected value. Diagnosis (CFIR), specification (§3.5) and measurement (Part 4) therefore matter more than the size of the menu.

How to read the numbers. Most Cochrane reviews of these strategies (the EPOC group’s) report a median adjusted risk difference: the percentage-point change in the share of professionals doing the desired thing, adjusted for baseline. The interquartile range (IQR) is the spread across trials, not a confidence interval. Effects in different units (odds ratios, standardised mean differences) are not comparable across rows.

Strategy (ERIC #)	Best current evidence	Typical effect	Certainty	What makes it work better	Read for learning work
Audit and provide feedback (5)	Cochrane 2026, 292 trials	Median +2.7 points (IQR 0.0–8.6); mean +6.2 (95% CI 4.1–8.2)	Moderate	Low baseline; feedback on the recipient’s own practice; comparison with top peers or a benchmark; a local champion who already has a relationship with the recipient; an action plan with specific advice. Repeated delivery predicted smaller effects, reversing the 2012 review	The largest evidence base of any strategy. The L&D bottleneck is having a behavioural measure to feed back at all
Conduct educational meetings (15)	Cochrane 2021, 215 studies	Median +4.0 (IQR 0.3–13.0) as the main component	Moderate (vs nothing)	Interactive vs didactic: unknown (very low certainty)	Workshops move practice a little. The evidence cannot rank formats
Conduct educational outreach visits (16)	Cochrane 2007, 69 studies	Median +5.6 (IQR 3.0–9.0)	Not graded	No tested factor explained the variation	A learning partner working with the team in situ. Consistent small effects; not updated since 2007
Inform local opinion leaders (38)	Cochrane 2019, 24 studies	Median +10.8 (IQR 3.5–14.6)	Moderate	Unclear which identification method matters	The largest verified median. Opinion leaders are nominated by peers, not appointed
Remind clinicians (58)	Cochrane 2009 (on-screen) and 2017 (paper)	+4.2 (IQR 0.8–18.8) and +6.8 (IQR 3.8–17.5). Single-component reminders +11.0 vs multi-component +4.0	Moderate (process)	Discrete, binary actions	Prompts at the point of action work for discrete steps, not judgement
Decision support in the workflow (12, 58)	Kwan et al. 2020, 122 trials	+5.8 points (95% CI 4.0–7.6) on processes; patient outcomes about zero	Not graded	Larger when users must record a reason for not following the advice, and when the action can be completed inside the tool	The nearest model for prompts inside AI-assisted work: friction and accountability features, not information
Distribute educational materials (31)	Cochrane 2020, 84 studies	Median +4 (IQR 1–9) vs nothing	Low to moderate	Effect uncertain when added to other strategies	Cheap; similar in size to meetings and feedback. A playbook is support, not a change strategy
Tailor strategies (63)	Cochrane 2015, OR shown (OR 1.56, 1.27–1.93) and 2026 update	“Probably a slight improvement” (2026 authors; numbers not re-verified)	Moderate (as stated by the 2026 authors)	—	This is the CFIR→ERIC step, and its measured advantage is slight
Facilitation (33)	Baskerville et al. 2012, 23 studies; FIRE trial; VA study	OR 2.76 (SMD 0.56) on guideline adoption	Not appraised. Publication bias (p = .003); FIRE null at 24 months; VA: reach and adoption up, downstream access not	Tailored delivery, more contact time, fewer practices per facilitator	Skilled internal coaching. The largest pooled effect on the thinnest certainty, and in practice a bundle of feedback, goal setting and tailoring
Create a learning collaborative (20)	Wells et al. 2018, 64 studies	83% report improvement on at least one measure; no pooled effect	Low	Straightforward practices with a clear evidence–practice gap	Cohort programmes with shared measures. Screen for “clear gap, simple practice” before building one
Alter incentive structures (2, 28, 70)	Cochrane, Scott et al. 2011, 7 studies	Insufficient evidence either way	—	—	The case against material incentives for judgement-type work comes from the motivation literature (companion review, Part 8), not from implementation trials
School implementation packages	Cochrane 2024, 39 trials	SMD 0.95 (0.71–1.19) on schools implementing health policies	Moderate; 26 of 39 trials at high risk of bias	All multi-strategy, so no strategy can be isolated	The nearest education-sector evidence. The outcome is organisational adoption, not teacher learning
De-implementation (cross-cutting)	Kien et al. 2024, 46 reviews	Most consistent reductions from tailoring, relationship and infrastructure/workflow strategies	Low to moderate (AMSTAR 2)	—	Retiring legacy programmes and practices looks more tractable through workflow and structure than through more education

No verified effect size exists for Identify and prepare champions (35) or Use train-the-trainer strategies (71). The one quantitative signal on champions is indirect: in the 2026 feedback review, feedback involving a local champion with an existing relationship to the recipient worked better.

Three rules for setting expectations

1. Expect single-digit percentage-point changes from any one strategy, with a wide spread. Plan evaluations that can detect effects of that size. Cluster trials in this field are typically powered to detect only 10–20-point gains (Grimshaw et al. 2012), so most internal L&D evaluations cannot see a realistic effect at all.
2. Bundle by diagnosis, not by count. Adding components does not reliably add effect (Squires et al. 2014; Grimshaw et al. 2004). Single-component reminders outperformed multi-component ones. Some pairings do look additive: feedback with co-interventions, and meetings with other strategies.
3. Treat every “why it works” as a hypothesis. A systematic review of 46 studies found none that established a mechanism of an implementation strategy by accepted criteria (Lewis et al. 2020). Name the proximal behaviour each strategy is supposed to move, and measure that.

The tension worth stating to clients. The 2026 feedback review finds feedback on an individual’s own practice more effective than team-level feedback. This document recommends team-level practice data (Appendix A, #5) because in most organisations individual practice data is experienced as surveillance, and surveillance destroys the psychological safety that capability work depends on (§2.7). Both positions have evidence behind them. Choose deliberately, and if you choose individual feedback, make it private, developmental and decoupled from performance judgement. That last condition is the aviation lesson in the companion bridge (Part 15).

The transfer caveat. Nearly all of these trials measure whether clinicians do the recommended thing at work. That is closer to L&D’s real outcome, behaviour in the job, than most L&D evaluation gets, so the evidence transfers better than its healthcare vocabulary suggests. But no review reports a corporate L&D, higher-education or teacher-learning outcome for any ERIC strategy. Every “read for learning work” note above is my judgement.

---

Part 4 — RE-AIM (and PRISM)

4.1 What it is and where it came from

Glasgow, Vogt and Boles published RE-AIM in the American Journal of Public Health in 1999 (89(9), 1322–1327). Its original purpose was consistent reporting — specifically, to force attention onto external validity, the dimensions that determine whether an effective intervention has any real-world impact. By 2019 it had been cited in more than 2,800 publications (Glasgow et al. 2019), applied across domains as varied as aging, cancer screening, physical activity, health policy, eHealth, worksite health promotion and quality improvement.

Its founding insight is the one L&D most needs: an intervention’s public health impact is a function of all five dimensions, not just effectiveness. A highly effective programme that reaches 5% of the target population, is adopted by 20% of settings, is delivered with poor fidelity and is discontinued after a year has close to zero impact — and, evaluated on effectiveness alone, looks like a success.

4.2 The five dimensions

Note that dimensions operate at two levels — individual and setting — and that conflating them is the most common error.

The official definitions have been extended since 1999 (D’Lima et al. 2022 quotes the current wording). Reach and Adoption now ask for the reasons why or why not, Effectiveness for heterogeneity across subgroups, and Implementation for adaptations made. The definitions below are the classic ones, read with those additions.

Reach — individual level

Definition: the absolute number, proportion, and representativeness of individuals who are willing to participate in a given initiative, intervention or programme.

Three components, and most reports give only the first:

• Absolute number — how many people
• Proportion — number ÷ eligible population. The denominator is the hard part and where most L&D reporting quietly fails
• Representativeness — how do participants differ from eligible non-participants? On seniority, function, tenure, prior performance, location, contract type

Learning translation: not “500 people attended.” Rather: 500 of 3,200 eligible (16%), skewed toward head-office functions, higher-performing units, and people who were already engaged.

Effectiveness (or Efficacy) — individual level

Definition: the impact of an intervention on important outcomes, including potential negative effects, quality of life, and economic outcomes.

Note the explicit inclusion of negative effects. RE-AIM asks you to report harms — time cost, opportunity cost, disengagement, unintended consequences. Standard learning evaluation does not have a place to put those.

Learning translation: the learning and behaviour outcome, measured against a comparison, plus what it cost people and what went wrong.

Adoption — setting level

Definition: the absolute number, proportion and representativeness of settings and intervention agents (the people who deliver the programme) who are willing to initiate it.

Two distinct things: which units took it up, and which deliverers agreed to deliver it. Representativeness applies again.

Learning translation: how many business units, faculties, departments or sites ran it? How many managers actually did the manager part? How do adopting units differ from non-adopting ones — and is the difference the reason it worked?

Implementation — setting level (and individual)

Definition: at the setting level, the intervention agents’ fidelity to the elements of the protocol, including consistency of delivery as intended and the time and cost of the intervention. At the individual level, it refers to clients’ use of the intervention strategies.

That second half matters and is almost always dropped: implementation at individual level means whether the recipient actually used the thing.

Learning translation: was the programme delivered as designed, by whom, with what adaptations, at what true cost — and did learners actually use the techniques?

Maintenance — both levels

Definition: at setting level, the extent to which a programme or policy becomes institutionalised or part of routine organizational practices and policies. At individual level, the long-term effects on outcomes 6 or more months after the most recent intervention contact in the framework’s original convention.

Current official guidance says the maintenance period should be tailored to the programme rather than fixed. Choosing a period deliberately, writing it down before launch, and making it at least six months is still a higher bar than almost anything in L&D evaluation.

Learning translation: is the practice still happening at six and twelve months? Is the programme still running without the original team? Has it entered onboarding, the standard operating rhythm, or the promotion criteria?

4.3 The calculations

The framework becomes rigorous the moment you write the denominators down.

Dimension	Metric	Formula / method
Reach	Participation rate	Participants ÷ eligible individuals
	Representativeness	Compare participants vs eligible non-participants on 4–6 pre-specified characteristics
Effectiveness	Primary outcome	Change vs comparison group; report effect size and CI, not just significance
	Negative effects	Pre-specified list, actively collected
Adoption	Setting adoption rate	Adopting settings ÷ eligible settings
	Deliverer adoption rate	Deliverers who delivered ÷ deliverers approached
	Representativeness	Compare adopting vs non-adopting settings
Implementation	Fidelity	Protocol elements delivered ÷ elements specified, per delivery
	Adaptations	Count, type, reason, whether core or peripheral
	Cost	Delivery cost + deliverer time + recipient time
Maintenance (setting)	Continuation rate	Settings still delivering at 6/12 months ÷ settings that adopted
Maintenance (individual)	Sustained outcome	Outcome at ≥6 months post last contact

The composite question RE-AIM forces: if reach is 16%, adoption is 30%, fidelity is 60% and maintenance is 40%, then the proportion of the eligible population receiving a faithfully delivered, sustained intervention is small — and that number, not the effect size, is your actual impact. Computing it is uncomfortable and clarifying.

4.4 PRISM

RE-AIM tells you what happened. It does not tell you why. The Practical, Robust Implementation and Sustainability Model (Feldstein & Glasgow, 2008) was developed as a contextually expanded version of RE-AIM to fill that gap, and has over 200 published applications (Glasgow, Trinkley, Ford & Rabin 2024). The 2024 consolidation relabels its domains; the current labels are used below.

PRISM has two components: four contextual domains plus the RE-AIM outcomes.

PRISM domain	Covers	L&D reading
Recipient perspectives on the intervention and implementation strategies	How the intervention and the way it is rolled out look from organisational and individual recipients’ positions	The programme as the sponsor sees it vs as the learner and manager see it. The gap is usually the story
Recipient characteristics	Characteristics, barriers and needs of organisational and individual recipients	Learners, their managers and the units they sit in
Implementation and sustainability infrastructure	The structures and resources supporting delivery and continuation	Who owns it after go-live; measurement continuity; budget line
External environment	Regulation, market, competition, funding, policy	Sector standards, accreditation, labour market, funding regime

When to use PRISM instead of plain RE-AIM: whenever you need to explain results rather than just report them, and whenever you are planning rather than evaluating. In practice PRISM overlaps substantially with CFIR — if you are already running a CFIR assessment, you do not also need PRISM’s contextual domains, and combining RE-AIM outcomes with CFIR determinants is the cleaner pairing.

4.5 Limitations

Dimensions are reported unevenly, and the pattern is consistent. Systematic reviews of RE-AIM use find maintenance the least frequently reported dimension, and within dimensions the under-reported elements are consistently: representativeness of individuals and settings, maintenance of effects, costs, and sustained implementation [B]. In the most recent review (D’Lima et al. 2022; 157 studies), maintenance was still the least evaluated dimension. In a sub-sample, setting-level maintenance was evaluated more often (60 per cent) than individual-level maintenance (20 per cent). Six individual-level criteria and one setting-level criterion were reported less than 20 per cent of the time. Qualitative methods were used to explain implementation results in two-thirds of studies. Only 8 of 157 studies used RE-AIM for planning rather than evaluation alone.

The lesson for practice is the inverse of the pattern: the dimensions people skip are the ones that carry the information. Reach without representativeness is a headcount. Effectiveness without maintenance is a snapshot. Adoption without denominators is marketing.

It is descriptive, not explanatory. RE-AIM tells you maintenance was 40%; it will not tell you why. Pair with NPT or CFIR.

The dimensions are not independent. Programmes optimised for reach often sacrifice fidelity; programmes optimised for effectiveness often sacrifice reach by selecting motivated participants. RE-AIM makes the trade-off visible but does not resolve it.

Applying it retrospectively is weak. Reviews note that pragmatic and planning-phase use is where the value is, and that retrospective RE-AIM scoring of a completed programme mostly documents what was not collected.

4.6 RE-AIM as a replacement for Kirkpatrick

For an L&D audience this is the operational headline.

Kirkpatrick	RE-AIM equivalent	What RE-AIM adds
L1 Reaction	(nothing)	RE-AIM has no satisfaction dimension
L2 Learning	Effectiveness (individual)	Requires a comparison and asks for negative effects
L3 Behaviour	Effectiveness + Implementation (individual use)	Separates “did they change” from “did they use the strategies”
L4 Results	Effectiveness (organizational outcomes)	Same
—	Reach	Who was not reached, and how do they differ
—	Adoption	Which units and deliverers took it up
—	Implementation (fidelity, adaptation, cost)	Was it delivered as designed, and what did it really cost
—	Maintenance	Is it still happening at six months

Three of RE-AIM’s five dimensions (Reach, Adoption, Maintenance) and the setting-level half of a fourth (Implementation: fidelity, adaptation, cost) have no Kirkpatrick equivalent. That is the size of the blind spot.

---

Part 5 — Normalisation Process Theory

5.1 What it is and why it behaves differently

NPT was developed by Carl May, Tracy Finch and colleagues (Sociology, 2009, 43(3), 535–554; Implementation Science, 2009, 4:29). It is best read as an explanatory theory of implementation work, not a checklist of determinants. Its developers now describe it as three things at once: a classificatory framework, a process model and an explanatory theory (May, Finch & Rapley 2026).

The framework’s own statement of what it is:

NPT is an Action Theory, which means that it is concerned with explaining what people do rather than their attitudes or beliefs.

Each of the four constructs represents a generative mechanism of social action — a kind of work that people do as a new practice becomes (or fails to become) normal. The question NPT answers is: what work is required for this practice to become routinely embedded, and is that work happening?

This makes it the right tool for the question L&D asks most often and answers worst: why didn’t it stick?

5.2 The four constructs and their sixteen sub-constructs

The 2022 update. The NPT coding manual (May et al. 2022) recasts the theory as 12 primary constructs in context–mechanism–outcome form:

• Contexts (what implementers work with): strategic intentions, adaptive execution, negotiating capacity, reframing organisational logics.
• Mechanisms: the four constructs below.
• Outcomes: intervention performance, relational restructuring, normative restructuring, sustainment.

The 16 sub-constructs that follow become optional secondary codes. For learning work the additions matter. “Reframing organisational logics” is where a capability programme changes what a unit believes good work is. “Normative restructuring” is the outcome that tells you a practice has become how things are done here, not just what people do. Definitions below are paraphrased from the official toolkit, and four are condensed (differentiation, communal specification, skill set workability, contextual integration). Use the official wording for coding.

The sub-construct definitions below are paraphrases of the official toolkit wording (normalization-process-theory.northumbria.ac.uk). Four depart noticeably from it: Differentiation, Communal specification, Skill set workability and Contextual integration. When citing, quote the toolkit, not this table.

Coherence — sense-making work

The question: what is the work?

Sub-construct	Definition	Diagnostic question for learning
Differentiation	The practice is defined by its differences from other practices	Can people say how this differs from what they did before?
Communal specification	The practice is understandable in relation to its distinctive features and perceived suitability — a shared understanding of aims and purpose	Do the team collectively agree what this is for?
Individual specification	Actors individually understand their specific tasks and responsibilities	Does each person know what they specifically now do differently?
Internalization	The value, benefits and importance of the practice are understandable	Do people grasp why it is worth the effort?

Failure signature: people can describe the programme but not the practice; each person has a different account of what it is for; individuals cannot state their own new task.

Cognitive Participation — relational work

The question: who does the work?

Sub-construct	Definition	Diagnostic question for learning
Initiation	Work that brings the practice forth — key people driving it	Is there anyone actually driving this, or did it arrive and stop?
Enrolment	Actors working together and organizing themselves to participate	Have people organised themselves around it, or are they complying individually?
Legitimation	Ensuring actors that it is right for them to be involved	Do people believe it is legitimate for someone in their role to be doing this?
Activation	Collectively defining the actions and procedures needed to sustain it	Has the group defined how they will keep it going?

Failure signature: the champion left and nothing happened; people are unsure whether this is “their job”; no collective agreement on how it continues.

Collective Action — operational work

The question: how does the work get done?

Sub-construct	Definition	Diagnostic question for learning
Interactional workability	The interactional work people do with each other, with artefacts and with other elements of the practice when operationalising it in everyday settings	Does it actually work in the flow of the day, with real people and real tools?
Relational integration	The knowledge work people do to build accountability and maintain confidence in the practice and in each other	Do people trust the practice and each other enough to rely on it?
Skill set workability	How workload and training requirements are affected; whether the practice is allocated to people with the right skills	Is this being done by people who can do it, and does the workload work?
Contextual integration	Issues of resource allocation, infrastructure, policy — the practice’s fit with organizational resources and rules	Do the systems, policies and budgets support it or fight it?

Failure signature: it works in the workshop and not on Tuesday; it depends on one skilled person; the expenses policy or the timetable makes it impossible.

Reflexive Monitoring — appraisal work

The question: how is the work understood and judged?

Sub-construct	Definition	Diagnostic question for learning
Systematization	Participants seek to determine how effective and useful the practice is for them and others, involving the work of collecting information in a variety of ways — formal or highly informal	Is anyone collecting information about how it is going? Note that anecdote collected around common themes counts
Communal appraisal	Participants work together, in formal collaboratives or informal groups, to evaluate the worth of the practice, drawing on experiential and systematized information	Does the group ever actually ask each other “is this working?” — and whose evidence counts when they disagree?
Individual appraisal	Participants work experientially as individuals to appraise the practice’s effects on them and on their context	Do individuals judge it worth it given everything else they carry?
Reconfiguration	Appraisal work by individuals or groups leads to attempts to redefine procedures, modify practices, or change the innovation itself	When people learn it isn’t working, does anything actually change?

The official worked example of individual appraisal is worth transplanting directly into learning: a falls-prevention programme that complicates an already demanding workload “may well have a low value attributed to it in practice irrespective of its effects on falls.” Substitute any capability initiative. People appraise a practice against their whole load, not against its own merits — which is the same finding as Relative Priority in CFIR and finite attention in the companion report, arriving from a third direction.

Failure signature: no data, no forum, no mechanism to change anything — so the practice quietly reverts and nobody notices.

5.3 NoMAD — the measurement instrument

The Normalization MeAsure Development questionnaire operationalises NPT as a survey [B].

Structure — 23 items:

• 3 general normalisation items scored 0–10, covering how familiar the practice feels, how far it currently feels like a normal part of work, and how far it will become one
• 20 items across four subscales: Coherence (4), Cognitive Participation (4), Collective Action (7), Reflexive Monitoring (5)

Validation (Finch et al., 2018, BMC Medical Research Methodology): 831 questionnaires returned, of which 413 were complete (522 answered at least one construct item); mean site response rate 39%. Confirmatory factor analysis, using item parcels, showed acceptable fit (CFI = 0.95, TLI = 0.93, RMSEA = 0.08, SRMR = 0.03), though the RMSEA sits above the authors’ own target of below 0.07. Internal consistency: Coherence α = 0.71; Cognitive Participation α = 0.81; Collective Action α = 0.78; Reflexive Monitoring α = 0.65. The overall 20-item normalisation scale α = 0.89.

Two things to note. First, Reflexive Monitoring has the weakest internal consistency (0.65) — treat that subscale’s scores with more caution than the others. Second, subsequent psychometric work examined NoMAD alongside ORIC, the organizational readiness measure, in 318 staff across 12 organisations in nine countries (Batterham et al. 2024). It found them related but distinct, which supports the pairing ORIC before, NoMAD during and after. It also found test–retest reliability suboptimal for both, so read single administrations as conversation starters rather than verdicts.

A learning-adapted NoMAD-style item set is in Part 9.

5.4 What NPT is good and bad at

Good at: explaining why a practice did or did not become routine; distinguishing four genuinely different failure modes that all look like “poor engagement” from the outside; giving a vocabulary for the work implementation requires, which makes invisible labour visible.

Bad at: predicting; assessing context before you start (use CFIR); measuring outcomes (use RE-AIM). It used to be weak at telling you what to do about a deficit. That gap is closing: May et al. (2025) derive an NPT-based taxonomy of 24 general and 96 micro-strategies from 63 papers, though ERIC remains the better-established menu.

5.5 Limitations

Emphasis on agency. A systematic review of 108 studies using NPT in feasibility studies and process evaluations (May et al. 2018) found about one in ten critiqued the theory, mostly its terminology. Two criticised its emphasis on agency and one its normative focus [B]. The theory foregrounds what people do, which is its strength, but it can under-weight structural constraint — the situations where people’s agency is genuinely limited by resources, hierarchy or policy. Contextual integration partially covers this, but as one sub-construct among sixteen.

Power and equity. Implementation science generally has been criticised for failing to account for inequity, and NPT is open to that challenge. The theory’s developers have acknowledged this and work is underway to address it, including integration with equity-focused design approaches. For now, treat NPT as blind to who benefits and who bears the cost of the new practice, and supply that analysis yourself.

Normative framing. “Normalisation” carries an implicit assumption that embedding is the goal. Sometimes the right answer is that a practice should be abandoned. NPT will describe the failure to normalise; it will not tell you the failure was correct. Pair with the de-implementation literature.

---

Part 6 — Using the four together

6.1 The lifecycle map

Phase	Question	Primary framework	Supporting
Explore	Should we do this at all? Is the innovation any good?	— (your own evidence base)	CFIR Innovation domain
Diagnose	What will help or hinder this, here?	CFIR	Weiner readiness (ORIC)
Design	What will we do about each barrier?	ERIC	CFIR–ERIC matching tool
Specify	Exactly who does what, when, how much?	Proctor’s 7 specification elements	—
Implement	Is the work of embedding happening?	NPT	NoMAD at 3 and 6 months
Evaluate	Did it reach, work, get adopted, get delivered, last?	RE-AIM	Proctor’s implementation outcomes
Sustain / stop	Is it institutionalised — or should it be stopped?	RE-AIM Maintenance + NPT	De-implementation frameworks

6.2 The canonical sequence, demonstrated

The cleanest published example of these frameworks working together is the Friendship Bench work in Zimbabwe:

1. RE-AIM used to classify implementation performance across clinics
2. CFIR used to analyse determinants by comparing high-performing to low-performing clinics
3. Strategies designed for the identified barriers, then tested in a cluster randomised trial

v2 check: the published protocol confirms RE-AIM first, then CFIR to compare high and low performers, then strategy design (Verhey et al. 2021). It does not say the CFIR–ERIC matching tool was used. The RE-AIM results show why classification came first: reach was 34% in small clinics, 15% in large and 9% in medium-sized ones (Verhey et al. 2022) [ABSTRACT].

Note the order. RE-AIM came first, to establish which sites were performing differently. Without that, the CFIR analysis would have described context with nothing to explain. This is the corrective to the most common CFIR error identified in Kirk’s review — assessing determinants without an implementation outcome to relate them to.

6.3 A single integrated page

For any implementation, this fits on one side:

The innovation — one sentence, plus the core components that must not change. The setting — Inner Setting named at a specific level; Outer Setting named. Top five determinants — from CFIR, rated, with evidence. Matched strategies — from ERIC, each specified by actor / action / target / temporality / dose / outcome / justification. Normalisation checkpoints — NPT four-construct check at 3 and 6 months. Evaluation plan — RE-AIM five dimensions with denominators pre-specified. Stop condition — what would tell us to stop, and who decides.

If a learning initiative cannot fill that page, it is not ready to launch. If it can, it is better planned than almost anything in the market.

6.4 What to use when you cannot do all four

Realistic triage, in order of value per hour:

1. RE-AIM denominators, pre-specified. Costs almost nothing, prevents the most common self-deception. Do this even if you do nothing else.
2. A five-construct CFIR mini-assessment: Relative Advantage (from the adopter’s position), Compatibility, Relative Priority, Available Resources, and the Mid-level Leader’s Capability/Opportunity/Motivation. Ten interviews, two days.
3. Three named ERIC strategies, fully specified. Forces the plan to be a plan.
4. An NPT four-question check at three months. Fifteen minutes in an existing meeting.

That is a defensible implementation approach in roughly a week of effort, and it is more than almost any learning function currently does.

---

Part 7 — The learning translation layer

This is the part that matters most.

7.1 The central problem: what is the innovation?

Implementation science asks you to name the “thing being implemented.” In healthcare this is usually obvious — a treatment protocol, a screening programme, a device. In learning it is not, and getting it wrong invalidates everything downstream.

You chose all four possible units. They behave very differently, and holding the distinction explicitly is the highest-value move available.

Unit A — A single learning programme as the innovation

Innovation = the programme itself. Deliverers = facilitators, trainers, faculty. Recipients = learners. Implementation success = the programme is delivered as designed, to the intended population, in the intended settings, and continues.

Frameworks behave: cleanly. This is the closest analogue to a clinical intervention. CFIR’s Innovation domain applies directly. RE-AIM’s dimensions are straightforward. Fidelity means “delivered as designed.”

The trap: you can implement a programme beautifully and change no behaviour whatsoever. Implementation outcomes will look excellent. The innovation was simply not the thing that needed to change.

Unit B — A new practice or behaviour as the innovation

Innovation = the behaviour itself (e.g. “managers hold a structured development conversation each month,” “faculty use retrieval practice in lectures,” “engineers run a peer review before merge”). Deliverers = line managers and the people who model it — not trainers. Recipients = the people doing the behaviour, and downstream, their teams. Implementation success = the practice is happening, routinely, and is still happening later.

Frameworks behave: this is where they are most powerful and most demanding. The learning programme becomes an ERIC strategy (cluster 5, train and educate stakeholders) rather than the innovation. That reframing alone changes how a programme is planned, budgeted and evaluated.

This is almost always the right unit, and almost never the one used. It is also the unit that aligns with the transfer evidence: if the behaviour is the innovation, then manager support, peer support and opportunity to perform stop being “transfer factors” and become implementation determinants and strategies, which is what they actually are.

The trap: it is harder to sell, harder to measure, and it makes the learning function accountable for something it does not fully control. That is a real objection and the honest answer is that the accountability already exists, just unstated.

Unit C — An L&D operating model as the innovation

Innovation = the system (capability framework, manager-led development model, new evaluation approach, learning-in-the-flow-of-work model). Deliverers = the L&D function, HR business partners, managers. Recipients = the whole organization; and internally, the L&D team itself. Implementation success = the new way of working is how learning is now done, without the transformation team.

Frameworks behave: Outer Setting and Inner Setting dominate; the Innovation domain matters less because the “innovation” is diffuse. NPT is especially strong here, because operating-model change is precisely the work of making new practices normal. Contextual integration and skill set workability usually carry the failure.

The trap: the Inner Setting for an operating model is often the L&D function itself, and assessing your own function honestly is the hardest assessment to run. Bring in someone external for that one.

Unit D — A learning technology as the innovation

Innovation = the platform, tool, AI tutor, simulation or assessment system. Deliverers = whoever configures, administers and champions it; and managers who require its use. Recipients = learners and administrators. Implementation success = it is used, as intended, by the intended people, and still used in a year.

Frameworks behave: all four apply, and NASSS should be added (see companion report Part 6). NASSS’s seven domains — condition, technology, value proposition, adopter system, organization, wider system, embedding and adaptation over time — were built specifically for this case, and the claim that failure is predicted by complexity across multiple domains simultaneously is the most useful single heuristic available for edtech and AI tooling decisions.

The trap: technology adoption metrics (logins, seats, completions) are Reach and Implementation metrics masquerading as Effectiveness metrics. RE-AIM discipline fixes this instantly.

7.2 The vocabulary translation

Implementation science	Learning equivalent	Note
Innovation	Programme, practice, operating model or technology	Pick one deliberately — Part 7.1
Innovation deliverers	Facilitators, faculty, and line managers	The manager is a deliverer, not a stakeholder. This reframe alone changes plans
Innovation recipients	Learners; and indirectly their teams
Inner setting	Business unit, faculty, school, site, team	Name the level
Outer setting	Sector, group, regulator, awarding body, market
Implementation strategies	What you do to make the learning land	The event is one strategy among many
Fidelity	Delivered as designed, with active ingredients present	Distinguish core from adaptable
Reach	Proportion of eligible population, with representativeness	Not attendance
Adoption	Units and deliverers who took it up	The missing metric in L&D
Maintenance	Still happening at 6 and 12 months	Higher bar than most L&D evaluation
Normalisation	It has become how we work; nobody calls it an initiative	The terminal state you want
De-implementation	Retiring a programme or practice	Needs the same rigour as launching

7.3 Six reframes worth carrying into client conversations

1. The learning event is an implementation strategy, not the intervention. It sits in ERIC cluster 5. There are eight other clusters.

2. The manager is a deliverer. Not an enabler, supporter or stakeholder — a deliverer, with their own Capability, Opportunity and Motivation to assess, and their own fidelity to measure.

3. “Engagement” is four different things. In NPT terms: do they understand it (coherence), have they organised around it (cognitive participation), can they do it in the flow of work (collective action), and is anyone appraising it (reflexive monitoring)? Four different failures, four different fixes, one useless word covering all of them.

4. Attendance is a Reach numerator with no denominator. The denominator is the number of eligible people, and representativeness is the part that tells you whether you reached the people who needed it.

5. Evaluation without adoption and maintenance is a snapshot of the enthusiastic. Effectiveness measured in the units that adopted, delivered well and stayed engaged is a biased estimate by construction.

6. If nobody can say why it works, it will not transfer. Causal ambiguity is the mechanism (companion report, Szulanski). ERIC’s “promote adaptability” strategy — identify what can be tailored and clarify which elements must be maintained to preserve fidelity — is the operational fix.

7.4 Where these frameworks have already been used in education

Not a blank field, though thin [B/C]:

• Medical and health professions education is the most developed. Work on incorporating implementation science principles into curricular design argues explicitly that IS principles can augment traditional curriculum design, with CFIR used as a determinant guide during needs assessment and RE-AIM used to evaluate longitudinally from initial reach through sustained maintenance.
• RE-AIM applied to curriculum — a conference abstract (Nearing et al. 2025; §14.4) maps the dimensions as: reach (who accesses and completes the curriculum), effectiveness (awareness, knowledge, self-efficacy, intentions to apply), adoption (uptake by the network of sites), implementation (how it integrates locally with other training requirements), and maintenance (sustained use by programmes). That is a clean, directly reusable template.
• Reusable learning objects across universities have been evaluated with RE-AIM across three Malaysian universities (Lim et al. 2025; §14.3).
• Higher education classroom practice — work on adapting implementation science for higher education research frames the systematic study of implementing evidence-based practices in college classrooms as an answer to the research–practice gap. This is the closest existing analogue to the faculty-adoption problem.
• School-based universal interventions have been mapped in a scoping review of implementation science use (Gallagher et al. 2023; §14.2).

The corporate L&D application is essentially unwritten. A v2 search confirmed it, while finding framework-based reviews in occupational safety and workplace health (Part 14). That is a gap — and, for consulting and doctorate purposes, an opportunity (companion report, Part 16).

---

Part 8 — Worked applications

Three contexts. Each shows the innovation choice, the CFIR determinants most likely to bind, matched ERIC strategies, the NPT checkpoints, and the RE-AIM plan. These are illustrative reasoning, not findings.

8.1 Corporate L&D — a leadership development programme

Standard framing: “We’re rolling out a leadership programme to 400 middle managers.”

Better framing (Unit B). Innovation = managers hold a structured monthly development conversation with each direct report, using a defined format. The programme becomes one strategy for producing it.

Justification for the choice: the transfer evidence says the programme is the weaker lever; the largest field experiment on producing psychological safety, still a working paper, used managers’ one-to-ones as its conduit (Castro, Englmaier & Guadalupe 2022); and defining the behaviour makes fidelity and maintenance measurable.

Likely binding CFIR determinants

Domain	Construct	Typical rating	Why
Innovation	Relative Advantage	−1	For the manager, it is time cost now for diffuse benefit later
Innovation	Complexity	−1	Requires new skill, new cadence, and often a new tool
Inner Setting	Relative Priority	−2	Competing with delivery targets and other initiatives
Inner Setting	Compatibility	−1	Does not fit a manager week already at capacity
Inner Setting	Incentive Systems	−1	Nothing happens to a manager who doesn’t do it
Inner Setting	Learning-Centeredness	varies	Where this is strong, everything gets easier
Individuals	Mid-level Leaders — Opportunity	−2	The single most common binding constraint
Individuals	Mid-level Leaders — Capability	−1	Most managers have never been taught to run one
Process	Assessing Needs (Deliverers)	blank	Nobody asked the managers

Matched ERIC strategies

Barrier	Strategy	Specification
Mid-level leader Opportunity	Revise professional roles	Remove one existing recurring meeting or report from the manager’s week. HR ops, before launch. Targets Opportunity
Relative Priority	Involve executive boards + Obtain formal commitments	Written commitment from each function head naming what they will personally do and review, monthly
Capability	Make training dynamic (methods, not styles) + Shadow other experts	Behaviour modelling: watch a good one, run one, get task-focused feedback. 2 × 90 min, spaced 3 weeks
Compatibility	Promote adaptability	Specify core (frequency, written record, two required questions) vs adaptable (length, channel, timing)
Incentive Systems	Audit and provide feedback	Monthly completion and quality summary to each function head. Informational, not punitive
Sustainment	Organize implementer team meetings	Peer group of 8 managers, 45 min monthly, protected time, facilitated for the first three
Causal ambiguity	Capture and share local knowledge + Identify early adopters	Interview the 5 managers already doing this well; publish what they actually do

NPT checkpoints at 3 and 6 months

• Coherence: can each manager state, unprompted, what they now do differently and why? (individual specification, internalization)
• Cognitive participation: has the peer group organised itself, or does it need the facilitator? (enrolment, activation)
• Collective action: does the conversation survive a busy week? Who does it fall off for, and why? (interactional workability, contextual integration)
• Reflexive monitoring: has anyone changed the format based on what they learned? (reconfiguration) — the strongest single signal of normalisation

RE-AIM plan

Dimension	Metric	Pre-specified denominator
Reach	Direct reports receiving ≥1 conversation/month	All direct reports of eligible managers, not just those of participating managers
	Representativeness	By tenure, performance band, location, contract type, shift vs office
Effectiveness	Team psychological safety (with within-group agreement); voluntary turnover; internal mobility	vs matched non-wave units
Adoption	Managers holding ≥1 conversation ÷ managers trained; units participating ÷ eligible units
Implementation	Fidelity: written record present, two required questions covered. Cost: manager hours × 12
Maintenance	Managers still at ≥80% cadence at 6 and 12 months ÷ managers who ever reached cadence

Design note: roll out in four waves by unit, order randomised. That converts the rollout into a stepped wedge at no additional cost and makes the effectiveness estimate causal.

8.2 Higher education — faculty adopting a new teaching practice

Scenario: a university wants teaching staff to adopt retrieval practice and spaced review in modules.

Innovation choice (Unit B). Innovation = the teaching practice, not the CPD workshop. Deliverers = faculty. Recipients = students. Note the two-level structure: faculty are recipients of the implementation effort and deliverers of the innovation to students. Say which you mean at every point or the analysis collapses.

Distinctive CFIR pattern in HE — the Outer Setting carries far more weight than in corporate settings:

Domain	Construct	Note
Outer Setting	Policies & Laws	Accreditation and quality regimes explicitly named in the construct definition
Outer Setting	Performance-Measurement Pressure	Student satisfaction metrics, league tables — and note these can push against retrieval practice, which students often rate as harder and less enjoyable
Outer Setting	Partnerships & Connections	Disciplinary bodies and communities often matter more to faculty identity than the institution
Inner Setting	Relative Priority	Research vs teaching allocation is the structural fact
Inner Setting	Work Infrastructure	Teaching loads and timetabling
Individuals	Opinion Leaders	Disciplinary seniority, not management hierarchy
Individuals	Mid-level Leaders (heads of department)	Control the only real lever: workload allocation
Innovation	Compatibility	Fits some disciplines’ assessment patterns far better than others
Innovation	Trialability	High — a lecturer can try it in one session. Exploit this

The satisfaction trap, named precisely. Desirable difficulty produces better learning and lower immediate satisfaction. In an institution where Performance-Measurement Pressure runs on student satisfaction, the CFIR analysis predicts that the innovation has negative relative advantage for the individual faculty member even when it has strong positive advantage for the institution and the student. That is not resistance; it is a correctly perceived incentive conflict, and it must be addressed structurally (Incentive Systems, Performance-Measurement Pressure) rather than by persuasion.

Matched ERIC strategies

• Inform local opinion leaders — disciplinary, not managerial
• Conduct cyclical small tests of change — one module, one term, reversible (exploiting high trialability)
• Capture and share local knowledge — what colleagues in this discipline actually did
• Develop academic partnerships — with the education research unit, giving the practice scholarly legitimacy (this addresses NPT legitimation directly)
• Audit and provide feedback — module-level attainment and retention data back to teaching teams
• Alter incentive/allowance structures — recognition in promotion criteria for teaching innovation; the structural fix for the satisfaction trap
• Change record systems — so that the practice and its effects are visible in module data

NPT emphasis in HE: legitimation is usually the binding sub-construct. Faculty need to believe it is right for someone in their role, in their discipline, to teach this way. That is an identity question, not a knowledge question, and it is why disciplinary opinion leaders outperform central teaching-and-learning units.

RE-AIM plan: Reach at student level (proportion of enrolled students in modules using the practice, with representativeness by widening-participation category — this is where the equity dimension bites). Adoption at module and department level. Implementation as fidelity to the core components. Maintenance across the following academic year, when the enthusiasm and the project funding have gone.

8.3 AI capability adoption

The case where these frameworks are most useful and least used. Add NASSS alongside.

Innovation choice. The most common framing — “roll out Copilot/ChatGPT Enterprise” (Unit D) — is the weakest, because it makes licences the unit and logins the metric. Stronger: define role-specific practices (Unit B). “Client managers draft first-pass proposal structures with AI and review against a checklist.” “Analysts use AI for exploratory data summarisation with mandatory source verification.”

CFIR determinants with the strongest AI-specific signal

Construct	AI-specific reading
Innovation Relative Advantage	Role-specific and frequently negative for the individual once review burden is included. The most common fatal error
Innovation Trialability	Often destroyed by policy lockdown before experimentation is permitted
Innovation Complexity	High: requires new skill, new judgement, new verification behaviour, often new workflow
Innovation Evidence-Base	Genuinely strong inside the capability frontier and negative outside it — see below
Inner Setting: IT Infrastructure	Access, data boundaries, integration into the tools people already use
Inner Setting: Learning-Centeredness	Psychological safety determines whether people admit uncertainty and share failures — the mechanism by which the frontier is learned
Inner Setting: Incentive Systems	Do people get credit for AI-assisted output, or suspicion?
Individuals: Recipients — Capability	Domain judgement, not prompt technique, is the binding capability
Process: Assessing Needs (Recipients)	Shadow-use data is a free needs assessment nobody reads

The observability problem, stated as a determinant. AI use is invisible — it happens inside one person’s window. In Rogers’ terms observability is near zero; in complex-contagion terms there is no social reinforcement; in NPT terms communal appraisal cannot occur because nobody sees anyone else’s practice. Engineering observability is therefore the highest-leverage cheap intervention available, and it is an implementation strategy, not a communications activity.

The jagged frontier as a capability target. The experimental evidence shows large gains inside the AI capability frontier and degraded accuracy outside it, with users unable to reliably detect which side of the frontier a task sits on. That means the capability to build is calibrated judgement about where the tool fails in this specific domain — acquired by practice with immediate correctness feedback, i.e. behaviour modelling and deliberate practice, not a webinar on prompting.

Matched ERIC strategies

• Conduct cyclical small tests of change — sanctioned low-stakes sandboxes with clear boundaries, before lockdown
• Capture and share local knowledge — prompt and workflow libraries built from what people actually do, in-role, per team
• Shadow other experts — paired working sessions; the highest-value single activity, and the one that fixes observability
• Identify early adopters — shadow-use data tells you exactly who they are
• Audit and provide feedback — at team level on practice, not individual on usage. Individual usage surveillance destroys the psychological safety the whole thing depends on
• Promote adaptability — specify what must be verified (core) versus how (adaptable)
• Revise professional roles — where AI shifts what a role is for, say so; leaving it implicit is the source of most quiet resistance
• Change record systems — make AI-assisted work labellable so that it becomes visible and discussable

NASSS overlay. Score complexity across the seven domains: condition/problem, technology, value proposition, adopter system, organization, wider system, embedding and adaptation over time. NASSS itself concludes, cautiously, that complexity across multiple domains poses the greatest challenge (Greenhalgh et al. 2017). A working heuristic, mine rather than the framework’s: if three or more domains are complex rather than complicated, expect non-adoption or abandonment and scope accordingly. My judgement is that most enterprise AI programmes would score complex on at least four.

RE-AIM plan

Dimension	AI-specific metric
Reach	Proportion of eligible role-holders using the defined practice (not the tool), with representativeness by seniority, function, tenure, location
Effectiveness	Task-level quality and time against a comparison; plus explicit negative effects — error rate on out-of-frontier tasks, verification burden, deskilling signals
Adoption	Teams that adopted the practice ÷ eligible teams; managers requiring it ÷ managers approached
Implementation	Fidelity to the verification step; adaptations logged; true cost including review time
Maintenance	Practice still in use at 6 and 12 months; and whether shadow use converged with sanctioned use or diverged further

The shadow-use question, framed properly. High personal-tool use alongside low official adoption is decoupling — formal structure separated from actual activity (companion report, Part 7). Treating it as a compliance problem suppresses the signal. Treating it as the most honest needs assessment available turns it into the foundation of the implementation.

---

Part 9 — Instruments

Working instruments. Adapt the wording; keep the structure.

9.1 CFIR determinant interview guide (learning-adapted)

Semi-structured. 45–60 minutes. Ask the role-relevant sections. Every construct definition is written as “the degree to which…”, so questions convert directly.

Opening

• Describe what is being introduced, in your own words. (Coherence check, before you start)
• Who decided this was needed? (Innovation Source; Tension for Change)
• What happens today, and what will be different? (Differentiation)

Innovation domain

• Compared with how you do this now, is the new way better — for you personally, in your job? (Relative Advantage — insist on the personal frame)
• What makes it complicated? How many other things have to change for it to work? (Complexity)
• Could someone try this on one team/module/case without committing? Could they stop? (Trialability — probe reversibility)
• Can it be adjusted to fit how your area works? What must not be changed? (Adaptability; core/adaptable specification)
• Where did this approach come from, and do you trust that source? (Innovation Source)
• Is there evidence it works? Has anyone shown you any? (Evidence-Base)
• What will this cost you — in time, not money? (Cost)

Outer Setting

• What outside your organization is pushing this? (External Pressure — probe societal, market, performance-measurement separately)
• Are there standards, accreditation or regulatory requirements involved? (Policies & Laws)
• Is anything happening externally that could derail it? (Critical Incidents)

Inner Setting

• Is the current situation bad enough that people want it to change? (Tension for Change)
• Does it fit how work actually flows here? Where does it not fit? (Compatibility)
• Compared with everything else landing this year, where does this sit? (Relative Priority — ask for a ranked list)
• What happens to someone who does this well? What happens to someone who ignores it? (Incentive Systems — both directions)
• What do you have and not have to do this properly? (Available Resources)
• Where would someone go for help? (Access to Knowledge & Information)
• When something goes wrong here, what happens? Can people say they’re struggling? (Learning-Centeredness — the psychological safety probe)
• Do teams here talk to each other about how they do things? (Relational Connections; Communications)

Individuals — ask per role

• What is your part in making this work? (Individual specification / Role clarity)
• Do you know how to do it? What would you need to learn? (Capability)
• Do you have the time and the authority to do it? (Opportunity)
• Do you want to? What would make you want to more? (Motivation)
• Who around here do people actually listen to on this? (Opinion Leaders)

Implementation Process

• Who is running this, and is there a real team? (Teaming; Implementation Leads)
• Did anyone ask you what you needed before designing it? (Assessing Needs — Deliverers)
• Is there a plan you have seen, with milestones and success measures? (Planning)
• Is there any way to feed back what is not working, and does anything happen? (Reflecting & Evaluating; Adapting)

Closing

• If this fails in a year, what will have caused it? (The single most productive question in the guide)
• What is the one thing that would make the biggest difference?

9.2 CFIR rating sheet

Construct	Rating (−2…+2 / blank)	Informants raising it (n/N)	Key evidence (verbatim)	Discriminating between sites?
Innovation Relative Advantage
Innovation Complexity
Innovation Trialability
…

Rules: two coders independently; disagreements resolved by discussion, not averaging; blank ≠ 0; every non-zero rating needs a quote; the final column is only completable in a multi-site design and is where the diagnostic value concentrates.

9.3 ERIC strategy selection and specification worksheet

Step 1 — barrier to candidate strategies

CFIR barrier (rated ≤ −1)	Candidate ERIC strategies	Endorsement level (if using matching tool)	Selected?	Why / why not

Step 2 — full specification, one row per selected strategy

Element	Entry
Strategy name (ERIC)
Actor
Action
Action target (whom, and which determinant)
Temporality
Dose (amount, frequency, duration)
Implementation outcome affected
Justification (evidence or reasoning)

Step 3 — coverage check

ERIC cluster	Strategies used	Count
1. Use evaluative and iterative strategies
2. Provide interactive assistance
3. Adapt and tailor to context
4. Develop stakeholder interrelationships
5. Train and educate stakeholders
6. Support clinicians (deliverers)
7. Engage consumers (learners and managers)
8. Utilize financial strategies
9. Change infrastructure

If clusters 1, 4, 6 and 9 are empty and cluster 5 is full, the plan is a training plan.

9.4 NPT normalisation check — learning-adapted

Modelled on NoMAD’s structure. Three general items on a 1–10 scale, then four subscales. Administer at 3, 6 and 12 months. Report subscale means and the spread, not just the average.

General normalisation (1–10)

1. How much is this practice currently a normal part of your work?
2. How likely is it that this practice will become a normal part of your work?
3. How familiar does the practice feel now?

Coherence (1 = strongly disagree, 5 = strongly agree)

1. I can clearly describe how this practice differs from what we did before. (differentiation)
2. My team has a shared understanding of the purpose of this practice. (communal specification)
3. I understand specifically what I am expected to do differently. (individual specification)
4. I can see the value of this practice for my work. (internalization)

Cognitive participation

1. There are people here who actively drive this practice forward. (initiation)
2. People in my team have organised themselves to make this work. (enrolment)
3. It feels appropriate for someone in my role to be doing this. (legitimation)
4. We have agreed how we will keep this going. (activation)

Collective action

1. This practice fits into how my working day actually runs. (interactional workability)
2. I trust that my colleagues are doing this properly. (relational integration)
3. I have the skills and the workload capacity to do this well. (skill set workability)
4. Our systems, policies and resources support this practice. (contextual integration)
5. When it is busy, this is one of the first things to be dropped. (reverse-scored; interactional workability under load)

Reflexive monitoring

1. Information is collected about how this practice is going. (systematization)
2. My team discusses whether this practice is working. (communal appraisal)
3. I have formed my own view about whether this is worth doing. (individual appraisal)
4. When we find something isn’t working, we change it. (reconfiguration)

Interpretation heuristics

• Low Coherence with high Cognitive Participation → willing people doing the wrong thing. Fix specification before adding energy.
• High Coherence and Cognitive Participation with low Collective Action → the classic training-transfer failure: they understand it and want to, and the work does not permit it. Fix opportunity and workload, not motivation.
• Everything acceptable except Reflexive Monitoring → it will decay silently. This is the most common profile at six months and the one that predicts failed maintenance.
• Item 20 (reconfiguration) is the strongest single indicator of genuine normalisation. A practice that has never been modified by its users has probably never been used.

9.5 RE-AIM scorecard

Complete the denominator column before launch. That is the discipline; everything else follows.

Dimension	Metric	Numerator	Denominator	Timing	Owner	Result
Reach	Participation rate	Participants	Eligible individuals	Wave close
Reach	Representativeness	Participant profile	Eligible non-participant profile on 4–6 pre-specified characteristics	Wave close
Effectiveness	Primary outcome		vs comparison group	3 months
Effectiveness	Negative effects	Pre-specified list, actively collected	—	3 months
Adoption	Setting adoption	Adopting units	Eligible units	Wave close
Adoption	Deliverer adoption	Deliverers who delivered	Deliverers approached	Wave close
Adoption	Representativeness	Adopting unit profile	Non-adopting unit profile	Wave close
Implementation	Fidelity	Core elements delivered	Core elements specified	Each delivery
Implementation	Adaptations	Count, type, reason, core/peripheral	—	Continuous
Implementation	True cost	Delivery + deliverer time + recipient time	—	Wave close
Maintenance (setting)	Continuation	Units still delivering	Units that adopted	6 and 12 months
Maintenance (individual)	Sustained outcome	Outcome ≥6 months post last contact	Participants	6 and 12 months

Composite impact estimate: Reach × Adoption × Fidelity × Maintenance. Compute it. It is usually a small number, and it is the honest one.

9.6 Fidelity and adaptation log

Date	Site / unit	Element	Core or adaptable?	Adaptation made	Reason	Who decided	Effect observed

Rule: adaptations to core elements require explicit sign-off and are logged as fidelity failures unless approved. Adaptations to the adaptable periphery are logged as learning and reviewed for possible incorporation into the design.

9.7 One-page implementation summary

Innovation: (one sentence) Core components (must not change): (3–5 items) Adaptable periphery: Inner Setting (level named): Outer Setting: Top 5 determinants (construct, rating, evidence): Selected strategies (name, actor, action, dose, target determinant): Normalisation checkpoints: 3 / 6 / 12 months, NPT four-construct check RE-AIM plan: denominators pre-specified for all five dimensions Stop condition: (what would tell us to stop, and who decides) Owner after go-live:

---

Part 10 — Protocols

10.1 A six-week implementation diagnostic

Deployable as a standalone consulting engagement.

Week 1 — Scope

• Define the innovation in one sentence; agree the unit (Part 7.1) with the sponsor. Expect this to take longer than you plan; it is the whole engagement in miniature.
• Name the Inner Setting level and the Outer Setting.
• Specify core vs adaptable components. If nobody can, flag causal ambiguity as a finding in its own right.
• Select CFIR constructs and record the justification.
• Identify roles and build the interview sample.

Weeks 2–3 — Collect

• 12–20 interviews (or 5–8 per site across 3–4 sites for a comparative design), covering high-level leaders, mid-level leaders, opinion leaders, deliverers and recipients. These numbers are my working rule, not CFIR guidance (§10.2).
• Document review: existing evaluation data, prior initiatives affecting this group, the change portfolio.
• If an implementation is already running, administer the NPT check (9.4).
• If pre-implementation, administer a readiness measure (commitment and efficacy separately).

Week 4 — Analyse

• Code to constructs; rate valence and strength; two coders.
• Identify discriminating constructs if comparative.
• Map determinants to candidate ERIC strategies.

Week 5 — Design

• Select and fully specify 5–8 strategies across at least four ERIC clusters.
• Build the RE-AIM plan with denominators.
• Define the stop condition.

Week 6 — Report and hand over

• One-page summary (9.7) plus a determinants annex with evidence.
• Present the uncomfortable findings first; the value of the engagement is concentrated there.
• Agree owner, cadence and the 3-month NPT checkpoint before you leave.

10.2 Sampling guidance

Design	Interviews	Notes
Single site, exploratory	10–15	Cover all five roles; saturation usually by 12
Single site, decision-grade	15–20	Add recipients beyond the enthusiastic; deliberately sample non-participants
Multi-site comparative	5–8 per site × 3–6 sites	The strongest design. Choose sites with known different implementation outcomes
Rapid / pragmatic	6–8 + one workshop	Restrict to 5–8 constructs; document the restriction

v2 note: these numbers are my working rule of thumb. The CFIR User Guide gives no number. It ties sampling to how the Inner Setting is defined and to role, profession and tenure, and its 2013 demonstration used 24 interviews across five facilities (Reardon et al. 2025; Damschroder & Lowery 2013) [VERIFIED].

Sampling rule that matters most: deliberately include people who did not participate, did not adopt, or dropped out. Reach and Adoption are defined by them, and they hold the information the enthusiastic cannot give you.

10.3 Evaluation cadence

Timing	Activity
Pre-launch	Determinant assessment; readiness; denominators fixed; core/adaptable specified
Wave close	Reach, Adoption, fidelity
Month 3	NPT check; adaptation log review; early Effectiveness signal
Month 6	NPT check; Effectiveness with comparison; Maintenance (setting) first read
Month 12	Full RE-AIM; Maintenance both levels; decision: sustain, adapt, scale, or de-implement
Month 18	Optional: has it normalised — i.e. has it stopped being called an initiative

10.4 What to do with each NPT failure profile

Profile	Diagnosis	First move
Low Coherence	People don’t know what the practice is	Re-specify. Write the behaviour down at individual level. Stop communicating and start defining
Low Cognitive Participation	Nobody owns it	Named roles, protected time, legitimation from a credible in-group source
Low Collective Action	It doesn’t fit the work	Structural: workload, tooling, policy, sequencing. Do not add training
Low Reflexive Monitoring	No feedback loop	Put it on an existing agenda with data attached; create the reconfiguration route
Low everything	Wrong innovation, wrong time, or saturated portfolio	Consider stopping. That is a legitimate finding

---

Part 11 — Limitations, honestly

11.1 What none of these frameworks can do

They cannot tell you if your intervention is any good. CFIR’s Outcomes Addendum is explicit: CFIR constructs primarily influence implementation outcomes, not innovation effectiveness. Perfect implementation of a weak learning design produces a well-executed waste. Your content evidence base is a separate obligation.

They give no effect sizes and no weighting. A determinant rated −2 and another rated −2 are not necessarily equally consequential, and nothing in the framework tells you which to fix first. Judgement fills the gap.

They do not specify mechanisms — except NPT, which is why it behaves differently and why the four are complementary rather than redundant.

They are heavy. 48 CFIR constructs, 73 ERIC strategies, 16 NPT sub-constructs, 5 RE-AIM dimensions with multiple components each. Full application is a research project. Pragmatic adaptation is legitimate and must be documented.

Framework proliferation is itself a problem. There are dozens of implementation frameworks with heavy conceptual overlap. Selecting one is a judgement call with limited guidance, and the field’s own reviews acknowledge this. Part 13 organises the most useful of them by the job they do.

11.2 What the evidence says about how well they are actually used

This is the sobering part, and it should moderate any claim you make about adopting them.

• CFIR: Kirk et al. found only 11 of 26 studies (42 per cent) assessed implementation outcomes, with widespread unjustified construct selection and shallow integration into analysis [A].
• ERIC: strategies are chronically under-specified in published use; the matching tool rests on expert consensus, with an average of 47 strategies endorsed per barrier by at least one expert, and imperfect correspondence to real-world effectiveness data [B].
• RE-AIM: maintenance, representativeness and cost are systematically under-reported; qualitative methods rarely used to explain results [B].
• NPT: critiqued, by a minority of users, for its terminology, its emphasis on agency and its normative focus; limited treatment of power and equity [B].

The pattern is consistent: the frameworks are better than their typical use. That is an opportunity — competent use is genuinely differentiating — and a warning: adopting the vocabulary is not the same as adopting the discipline, and the vocabulary is what usually gets adopted.

11.3 When not to use them

• When the answer is obvious and small. A three-person team adopting a shared template does not need CFIR.
• When the real problem is that the intervention is wrong. These frameworks will help you implement the wrong thing thoroughly.
• When you cannot measure any implementation outcome. Determinant assessment without an outcome to relate it to is description, not diagnosis.
• When the honest answer is “stop.” None of these frameworks will tell you to stop; three of them are biased toward embedding. Hold the de-implementation question separately and deliberately, with the tools in §13.7.
• When using them would be ceremonial. If the assessment will not change the plan, the assessment is theatre — which is exactly the decoupling pattern these frameworks are supposed to expose.

11.4 Transferability caveats for learning contexts

• All four were developed in healthcare and public health. Their assumptions about roles, data availability, protocol fidelity and regulatory context do not transfer uniformly.
• CFIR’s transferability to low- and middle-income settings has been questioned in a systematic review; the same caution applies to sectors with very different organizational forms.
• Fidelity is a cleaner concept for a clinical protocol than for a teaching practice or a management conversation, where variation is often the point. Specifying core versus adaptable is how you handle this, and it requires more judgement in learning than in medicine.
• “Innovation recipients” in learning are frequently also deliverers of something else — a manager receives the capability programme and delivers the practice. Two-level structures are normal in learning and unusual in the source literature. Say which level you mean, every time.

---

Part 12 — Common failure modes when applying this to learning

1. Choosing the programme as the innovation when the behaviour is the innovation. The single most consequential error. Everything downstream inherits it.
2. Treating managers as stakeholders rather than deliverers. They have fidelity, capability, opportunity and motivation, and none of it gets assessed.
3. Reporting attendance as Reach. No denominator, no representativeness, no information.
4. Skipping Adoption entirely. The metric that tells you whether the organization took it up, absent from virtually all L&D reporting.
5. Assessing determinants with no implementation outcome to explain. Kirk’s finding, reproduced faithfully by practitioners.
6. Using the CFIR–ERIC matching tool as an oracle. It is expert opinion with wide dispersion, useful for broadening candidates and forcing explicit reasoning.
7. Naming strategies without specifying them. “We’ll do facilitation and champions” is not a plan.
8. Filling only ERIC cluster 5. Training and education is one cluster of nine and 11 strategies of 73, and the transfer evidence says it is not where the strongest levers are.
9. Never specifying core versus adaptable. Then adaptation during rollout silently removes the active ingredient, and the evaluation cannot tell you.
10. Measuring maintenance at three months and calling it sustained. RE-AIM’s original individual-level convention is six months after the last contact; current guidance tailors the period, but three months is short for almost any capability.
11. Running the NPT check only at the end. Its value is as an early-warning system at three months, when reconfiguration is still possible.
12. Adopting the vocabulary without the denominators. The most likely failure mode for anyone who reads this document and moves quickly.

---

Part 13 — The wider toolkit: what to add to the core four, by job

Bottom line. The core four leave eight jobs under-served:

• choosing and specifying strategies
• diagnosing individual behaviour
• phasing a rollout
• adapting without losing the effect
• sustaining
• stopping
• measuring cheaply
• designing and reporting a study

Each job has a well-cited tool, and most are free. But most are used thinly and almost none has been tested against outcomes. Treat them as structured prompts that make you look at the right things, not as validated predictors. Almost all the evidence comes from health care and public services.

How thinly they are used.

Tool	Evidence of use	Source
EPIS	49 projects, all public sector. On average two of its four phases measured; mean depth of use 2.8 on a 5-point scale	Moullin et al. 2019 [ABSTRACT]
PARIHS	Used in 367 of 1,475 articles read, often partially and thinly described	Bergström et al. 2020 [ABSTRACT]
Implementation Mapping	29 studies; prioritisation inconsistently described; outcome evaluation often missing	Warhurst et al. 2025 [ABSTRACT]
Theoretical Domains Framework	60 intervention-design studies among 3,540 papers screened	Dyson & Cowdell 2021 [ABSTRACT]
NIRN Active Implementation Frameworks	No trial or systematic review of outcomes found; the indexed literature is mostly case studies and protocols	Sweep, October 2026

13.1 Selector

Job	Tool	Why this one	Cost to apply
Quick strategy options for known barriers	CFIR–ERIC matching tool (§3.4)	Experts diverge (about 47 strategies endorsed per barrier), so use it as a prompt list	1–2 hours
Strategies for a costly or sensitive rollout	Implementation Mapping	Stakeholder-led; links determinants to mechanisms to protocols	Weeks
One-page plan and report	Implementation Research Logic Model (IRLM)	Forces the determinant → strategy → mechanism → outcome chain	Half a day
Testing why a strategy should work	Causal pathway diagram plus a proximal measure	Exposes failed preconditions before launch	One workshop
Individual behaviour of learners and managers	TDF with COM-B	Training reaches 2 of TDF’s 14 domains	8–15 interviews plus coding (my estimate)
Phased system rollout	EPIS or NIRN	Phases and bridging factors (EPIS); teams, drivers and practice profiles (NIRN)	Low to moderate
Local change without losing the effect	Functions vs forms; FRAME log; MADI for impact	Records who changed what, why, and whether it fits the core functions	Low
Still alive at 12 months?	PSAT or CSAT (short forms)	Validated capacity domains	One survey
Retiring a course or practice	McKay’s criteria, then a de-implementation framework	Stopping differs from starting	Low
Acceptability, fit and feasibility pulse	AIM, IAM and FIM	Free; four items each; under five minutes	Minutes
Pre-launch readiness	ORIC; R = MC² as a heuristic	ORIC has validation data; R = MC² does not	Minutes
Testing effect and uptake together	Hybrid type 1, 2 or 3 study	Type chosen by how strong the effectiveness evidence already is	Design-level
Why it worked, for whom	Realist evaluation (RAMESES II)	Context–mechanism–outcome theory testing	High
Reporting	StaRI plus TIDieR	27 and 12 items	Hours
Adopting an AI tool	NASSS plus CFIR, with the 2026 AI caveats	Adaptive AI breaks the stability assumptions of the core four	Moderate

13.2 Choosing and specifying strategies

The problem. The CFIR–ERIC matching tool aggregates expert votes. In its source survey, 169 of 435 invited experts ranked strategies for 39 barriers. On average 47 different strategies (range 35–55) were endorsed per barrier, with few consistent links (Waltz et al. 2019). That makes the tool a way to widen options, not to choose among them. Powell et al. (2017) review the methods available for the choosing step: concept mapping, group model building, conjoint analysis and intervention mapping [ABSTRACT].

Implementation Mapping (Fernandez et al. 2019) is the most developed of these. It has five iterative tasks:

1. Assess needs and identify the adopters and implementers.
2. State outcomes and performance objectives, identify determinants, and build matrices of change objectives.
3. Choose theoretical methods (mechanisms of change) and strategies.
4. Produce protocols and materials.
5. Evaluate.

A scoping review to August 2024 found 32 publications covering 29 studies, mostly from the US. They described how they prioritised determinants and strategies inconsistently, and often did not evaluate outcomes (Warhurst et al. 2025) [ABSTRACT].

The two approaches divide the work:

• The matching tool takes an afternoon.
• Implementation Mapping takes weeks, with the people whose work must change in the room.

Use the tool to widen the options and Implementation Mapping when the bundle is expensive or politically sensitive. An enterprise AI-skills programme qualifies.

The Implementation Research Logic Model (Smith, Li & Rafferty 2020) is a one-page chain:

• determinants, using CFIR domains
• strategies
• mechanisms of action
• outcomes, in Proctor’s format

There are variants for comparative, multi-context and optimisation designs, and fillable templates in the open paper [VERIFIED]. The evidence is self-report only. Of 66 survey respondents among more than 130 researchers and partners trained, 77.7% found it moderately or very helpful for rigour, and 74.0% found it useful as a roadmap. It has had no outcome test. For learning work it replaces v1’s one-page summary (§9.7): one row per strategy, showing barrier, strategy, mechanism and measure. That is the page a sponsor reads.

Mechanisms. §3.7 reported that none of 46 studies established a mechanism for an implementation strategy. The field’s answer is the causal pathway diagram (Lewis et al. 2018; refined in Klasnja et al. 2024) [ABSTRACT]. It has four steps:

1. Specify the strategy’s core component.
2. Link the strategy to a mechanism.
3. Name the proximal and distal outcomes.
4. State the preconditions and moderators.

Preconditions are where bundles quietly die. One clinical bundle combined an alert, education and a champion, and it failed. Only 54.2% of patients had the laboratory results the alert needed, so a precondition of the main strategy was absent for nearly half the target population (Austad et al. 2025) [ABSTRACT].

The learning equivalents are familiar:

• managers who have no time to coach
• practice tasks that the job never presents
• AI tools to which learners have no licence

For each strategy, write down:

• the mechanism
• the preconditions
• one proximal behaviour measure
• one distal outcome

If a precondition fails, drop or fix the strategy before launch.

13.3 Diagnosing individual behaviour: TDF and COM-B

CFIR’s individual domain is thin, and NPT is about collective work. The Theoretical Domains Framework fills the gap. Its 14 domains were distilled from 112 constructs by an expert card sort (Cane, O’Connor & Michie 2012) [ABSTRACT]:

• Knowledge
• Skills
• Social/professional role and identity
• Beliefs about capabilities
• Optimism
• Beliefs about consequences
• Reinforcement
• Intentions
• Goals
• Memory, attention and decision processes
• Environmental context and resources
• Social influences
• Emotions
• Behavioural regulation

The clustering was only modest: the average silhouette was 0.29.

COM-B is the coarse layer above TDF. Capability, opportunity and motivation are the hub of the Behaviour Change Wheel, which links them to nine intervention functions and seven policy categories (Michie, van Stralen & West 2011). The TDF guide (Atkins et al. 2017) [VERIFIED] synthesises 33 theories. It presents TDF as a lens, not a theory, and maps it onto COM-B.

The learning point is arithmetic. Knowledge and Skills are 2 of TDF’s 14 domains. A training course addresses those two. When a programme fails to change behaviour, the cause usually sits in the other twelve:

• identity
• beliefs about consequences
• reinforcement
• environmental resources
• social influences

That is the transfer literature’s finding (companion review) restated in implementation terms. A workable sequence:

1. Use COM-B as a three-question screen in manager and learner interviews: can they, does the environment let them, do they want to?
2. Use TDF only for the components that fail.

Limits.

• TDF is used mainly qualitatively, and most studies apply it rigidly and deductively, so determinants outside its domains go unseen. One review of 186 qualitative studies found findings confined to TDF’s own domains (McGowan et al. 2020) [ABSTRACT].
• Designing interventions from TDF findings remains underdeveloped (Atkins et al. 2017).

13.4 Phasing and infrastructure: EPIS, NIRN, i-PARIHS and the MRC framework

EPIS (Aarons, Hurlburt & Horwitz 2011) [ABSTRACT] divides implementation into four phases: Exploration, Preparation, Implementation, Sustainment. In each phase it looks at:

• the outer context (the system)
• the inner context (the organisation)
• bridging factors between them
• the innovation itself

It was built for public-service sectors, and its site offers free measures and worksheets (episframework.com).

It adds two things CFIR lacks:

• Determinants that change by phase. CFIR is phase-agnostic.
• Bridging factors. For learning, these are the accreditors, vendors and procurement rules that connect an organisation to its environment.

Its Dynamic Adaptation Process separates core elements from adaptable characteristics (Aarons et al. 2012). Use in practice is shallow (see the table above): inner-context factors appeared in 90% of projects, bridging factors in only 31%.

The NIRN Active Implementation Frameworks (Fixsen et al. 2009; for education, Fixsen et al. 2013) [ABSTRACT] are five linked frameworks:

• usable innovations
• stages: Exploration, Installation, Initial Implementation and Full Implementation, which NIRN says takes two to four years
• implementation drivers: competency, organisation and leadership
• implementation teams
• improvement cycles

The hub is education-oriented and free (implementation.fpg.unc.edu) [VERIFIED]. Its most transferable tool is the practice profile: a description of the practice that is teachable, learnable and assessable, written before any training is designed.

The outcome evidence is thin. No trial or systematic review of AIF outcomes was found. The nearest tests are two Swedish school cluster trials of a multifaceted strategy (implementation teams, improvement cycles, training) against a single one:

• 19 schools: no significant difference at 12 months, on an unvalidated self-report outcome (Toropova et al. 2022) [ABSTRACT].
• 55 schools: small effects on recipient-rated fidelity, d = 0.16 to 0.41 (Rödlund et al. 2025) [ABSTRACT].

Both measured whether staff followed the guideline, not whether students learned. [C]

i-PARIHS (Harvey & Kitson 2016) [VERIFIED] makes facilitation the active ingredient, working on three things:

• the innovation
• the recipients, individual and collective
• the context, inner and outer

Success is defined as agreed goals achieved and embedded. The framework answers earlier critiques of PARIHS: little prospective evidence, blurred elements, too much focus on the facilitator’s role.

Its main test, the FIRE trial (Seers et al. 2018) [VERIFIED], randomised 24 long-term care sites in four European countries to three arms:

• standard dissemination
• 12 months of goal-focused facilitation
• 24 months of holistic facilitation

The outcome was documented compliance with continence recommendations, measured over 24 months in 2,313 records. There were no significant differences between arms, and all arms improved.

This null is weak evidence against facilitation, for three reasons:

• Probably underpowered. The trial was powered on an intraclass correlation of 0.01; the observed baseline values were 0.40–0.55, which the authors called “much higher than expected”. (Inferring low power from that is my reading.)
• Wrong framework version. It tested facilitation under the original PARIHS, not i-PARIHS.
• Context limited the models. The process evaluation found that context constrained what facilitation could do (Rycroft-Malone et al. 2018) [SECONDARY].

My reading: the defensible lesson is narrower than either side claims. Coaching does not rescue a hostile context, so diagnose context first. Then budget facilitation as a named role with a stated dose, not as goodwill. [B/C]

The MRC/NIHR framework for complex interventions (Skivington et al. 2021) [VERIFIED] sits a level above all of these. It has:

• four research perspectives: efficacy, effectiveness, theory-based, systems
• four phases, not necessarily sequential: develop or identify, feasibility, evaluate, implement
• six core elements to revisit at every phase: context, programme theory, stakeholders, key uncertainties, refinement, economic considerations

It names no determinants and no strategies. Its value for learning is its question-first logic: what is the key uncertainty this pilot must resolve? An L&D buyer is rarely asking the efficacy question (“can this work under ideal conditions?”). Write the programme theory and the key uncertainties before piloting anything.

13.5 Adapting without losing the effect

Fidelity and adaptation are not opposites. The concept that reconciles them separates core functions from forms (Perez Jolles, Lengnick-Hall & Mittman 2019) [ABSTRACT]:

• A core function is the basic purpose an element serves. In learning, deliberate practice with feedback is one.
• A form is the particular way it is delivered: format, length, language, medium.

Forms can be adapted freely. Functions must be protected. The idea is usually credited to Hawe, Shiell & Riley (2004) [UNVERIFIED].

Recording adaptations.

• FRAME (Wiltsey Stirman et al. 2019) records eight things about each adaptation:
  • when and how it was made
  • whether it was planned or reactive
  • who decided
  • what was changed
  • at what level of delivery
  • the type of context or content change
  • whether it was consistent with fidelity
  • the reasons
• FRAME-IS (Miller et al. 2021) applies the same logic to implementation strategies [ABSTRACT].
• MADI (Kirk et al. 2020) [VERIFIED] adds the effects: ripple effects, mediators and moderators. It also separates proactive from reactive and systematic from unsystematic changes.

For learning, upgrade v1’s adaptation log (§9.6) to FRAME’s eight fields.

Why fidelity matters.

• A review of more than 500 studies found that the level of implementation affects outcomes, and identified 23 contextual factors that affect implementation (Durlak & DuPre 2008) [ABSTRACT] [A/B].
• The standard conceptual framework for fidelity (Carroll et al. 2007) [VERIFIED] has four elements: adherence, dose or coverage, quality of delivery, and participant responsiveness. Two moderators affect them: intervention complexity and facilitation strategies. The authors note that the framework “requires testing”.

There is very little research on whether adaptations help or harm. The dynamic sustainability view (below) argues that the usual default, assuming adaptation is “program drift”, blocks improvement.

13.6 Sustaining

Definitions are scarce. Of 209 articles in four syntheses, only 24 (11.5%) defined sustainability. The consensus definition has five parts (Moore et al. 2017) [ABSTRACT]:

• after a defined period
• the programme or its strategies continue, and/or
• behaviour change is maintained
• the programme may evolve
• while its benefits continue

How many programmes are sustained? Nobody knows, as a single number. The best review (Wiltsey Stirman et al. 2012) [VERIFIED] covered 125 studies. Only 7% were educational, and 65% gave no definition of sustainability. Its findings:

• Partial sustainment was “relatively high”.
• Few studies with independent observation reported high site-level continuation.
• Fewer than half of observed providers kept high fidelity.
• Of 56 studies that tracked implementation after support ended, 19 found lower levels, 17 higher, 3 unchanged and 17 mixed.

No review pools these into a share of programmes sustained, so any such figure in a pitch deck is invented. [B]

Frameworks.

• The Dynamic Sustainability Framework (Chambers, Glasgow & Stange 2013) [ABSTRACT] expects continued learning, ongoing adaptation for fit and continuing improvement. It rejects “voltage drop” as the default expectation.
• The Integrated Sustainability Framework (Shelton, Cooper & Wiltsey Stirman 2018) [ABSTRACT] lists multilevel determinants.

Shelton and colleagues later noted that these frameworks give a strong foundation of determinants but little guidance on measuring sustainability outcomes (Shelton et al. 2020) [VERIFIED].

Tools.

• PSAT (Luke et al. 2014): 40 items in eight domains, tested with 592 staff from 252 programmes; mean alpha 0.88.
• CSAT, the clinical version (Malone et al. 2021): 35 items in seven domains, tested with 286 clinicians.
• Short forms with three items per domain (Malone et al. 2024): alphas 0.82–0.92.

All three are [ABSTRACT].

For learning. Run a short CSAT-style check at 3 and 12 months. Ask what has been adapted, not only what has stopped. Fund a standing improvement cycle rather than a launch-only project. Strategies for sustainment are themselves barely studied: of 26 articles, 7 defined sustainment and 6 reported strategies (Hailemariam et al. 2019) [ABSTRACT].

13.7 Stopping: de-implementation

v1 warned that three of the four core frameworks are biased toward embedding (§11.3). De-implementation now has its own literature:

• When to stop. Three criteria (McKay et al. 2018) [ABSTRACT]: the practice is not effective or is harmful; it is not the most effective or efficient option; or it is no longer necessary.
• How stopping differs from starting (Norton & Chambers 2020) [ABSTRACT]: in multilevel factors, types of action, strategies, outcomes and unintended consequences.
• What to measure. Proctor’s eight implementation outcomes can be mapped onto de-implementation, with degrees and thresholds (Prusaczyk et al. 2020) [ABSTRACT].
• The frameworks. A scoping review found 27 (Walsh-Bailey et al. 2021) [ABSTRACT]. 21 are multilevel and 18 linear. Only 13 had been applied empirically.
• What works. Strategy effects are in §3.7 (Kien et al. 2024). Tailoring, relationship and workflow strategies gave the most consistent reductions (low to moderate certainty). The overview did not compare them directly with education.

For learning. Apply McKay’s three criteria to the course catalogue once a year. Plan each exit as a project with outcomes: archive, redirect or retrain.

13.8 Measuring cheaply

AIM, IAM and FIM (Weiner et al. 2017) [VERIFIED] measure three things:

• Acceptability (AIM): “meets my approval”, “is appealing to me”, “I like”, “I welcome”.
• Appropriateness (IAM): “seems fitting”, “suitable”, “applicable”, “a good match”.
• Feasibility (FIM): “seems implementable”, “possible”, “doable”, “easy to use”.

Each has four items on a five-point scale. Their properties:

• alphas 0.85–0.91
• test–retest reliability 0.73–0.88
• sensitive to change
• free, and each takes under five minutes
• no norms or cut-offs
• predictive validity untested at publication

For learning, ask learners and managers all 12 items before and after a pilot.

Readiness.

• ORIC (Shea et al. 2014) [VERIFIED]: ten items in two factors, change commitment (alpha 0.92) and change efficacy (alpha 0.88). Its predictive validity was untested at publication.
• R = MC² (Scaccia et al. 2015) [ABSTRACT] defines readiness as motivation × general capacity × innovation-specific capacity. It is a useful heuristic, but no free validated tool was found.

Choosing measures.

• PAPERS (Stanick et al. 2021) rates how pragmatic a measure is: useful, compatible, acceptable, easy.
• Applied to 150 implementation-outcome measures, the ratings showed that psychometric evidence beyond internal consistency and norms was often missing (Mettert et al. 2020) [ABSTRACT].

Prefer measures that are cheap, short and published with their psychometrics.

13.9 Designing and reporting

Hybrid studies (Curran et al. 2012) combine effectiveness and implementation questions:

• Type 1 tests the intervention and observes implementation.
• Type 2 tests both.
• Type 3 tests an implementation strategy and observes the intervention’s effect.

The 2022 reflections (Curran et al. 2022) [VERIFIED] add two points. Call them hybrid studies, not designs, because the type follows the study’s aims and primary outcomes. And choose the type by how strong the effectiveness evidence already is.

For learning, a pilot that compares two rollout methods while tracking performance is a type 3 study. Saying so makes the evaluation honest about what it can conclude.

Other designs.

• Stepped-wedge trials suit staged rollouts but carry a higher risk of bias from misspecified time trends, worst with few, heterogeneous clusters. The CONSORT extension requires a justification for the design (Hemming & Taljaard 2020; Hemming et al. 2018) [ABSTRACT].
• Realist evaluation asks what works, for whom, in what circumstances. Its reporting standard, RAMESES II, has 20 items (Wong et al. 2016) [ABSTRACT].

Reporting.

• StaRI (Pinnock et al. 2017): 27 items in two strands, the implementation strategy and the intervention’s effectiveness.
• TIDieR (Hoffmann et al. 2014): 12 items describing an intervention.

TIDieR’s 12 items are the minimum another business unit would need to run the same programme. Use them to document every learning programme you expect to scale.

13.10 Sector derivatives and RE-AIM extensions

Schools. SISTER adapted ERIC’s 73 strategies into 75 for schools (Cook et al. 2019) [ABSTRACT]: 52 surface changes, 5 deeper changes, 6 deletions and 7 additions. In education, start from SISTER, not ERIC. A school effort guided by EPIS mapped about 20 SISTER strategies across 19 high schools (Gunderson et al. 2021) [ABSTRACT].

Community settings. ISAC lists 40 strategies drawn from interviews with researchers and practitioners (Balis et al. 2024). ISAC Match adds a four-step selection process with equity considerations (Balis et al. 2025) [ABSTRACT].

RE-AIM updates.

• An extension adds dynamic sustainability, iterative assessment, and equity and cost as drivers (Shelton, Chambers & Glasgow 2020).
• Iterative RE-AIM (Glasgow et al. 2020) [ABSTRACT]: teams pick the one or two weakest dimensions, act on them, and meet again after about six weeks. Five VA projects found it useful, mostly for reach and adoption.
• Holtrop et al. (2021) correct 13 common misconceptions about the framework.

For learning, run a six-week RE-AIM check in the middle of a rollout and fix the weakest dimension (§10.3).

13.11 Frameworks for AI adoption

NASSS (Greenhalgh et al. 2017) remains the best general framework for technology adoption (§7.1, §8.3). It covers seven domains:

• the condition
• the technology
• the value proposition
• the adopter system
• the organisation
• the wider context
• adaptation over time

Programmes that were complex in several domains were “rarely, if ever” mainstreamed. NASSS has since been applied to AI decision support in radiology (Naicker et al. 2026) [ABSTRACT].

Does CFIR fit AI? Nilsen, Hald & Neher (10 September 2026) [ABSTRACT] argue that it depends on the kind of AI:

• Locked AI behaves much like a conventional digital intervention.
• Adaptive and generative AI bring new features to the fore: opacity, probabilistic output, dependence on local data, drift, vendor-controlled updates, regulatory uncertainty, tensions with professional identity, and the need for calibrated trust.

Nilsen and colleagues argue that these features challenge CFIR’s assumptions that the innovation stays stable and has clear boundaries. The same logic would apply to the other three frameworks (my inference). It is the theoretical development since v1 that matters most for AI work.

Frameworks are rarely used in AI implementation research.

• Only 34.7% of 202 qualitative studies of clinical AI applied any framework, and 40 of the 50 frameworks used appeared only once (Hogg et al. 2023) [ABSTRACT].
• The AI-specific frameworks are lifecycle and governance checklists, all from health care [ABSTRACT]:
  • SALIENT (van der Vegt et al. 2023)
  • FAIR-AI (Wells et al. 2025)
  • generative AI with technology acceptance and NASSS (Reddy 2024)
  • generative AI in health professions education (Toohey et al. 2025)
• Their genuinely new element is monitoring after deployment: drift checks and control over updates. Nothing in the core four does this.
• No workplace or higher-education framework for AI adoption was found.

For learning.

1. Classify each AI tool as locked, adaptive or generative before applying CFIR.
2. For adaptive and generative tools, add three items to the sustainment plan: a named owner, an update check and a drift check.
3. For AI-skills programmes, pair NASSS (the system) with TDF (the individual).
4. Remember that the “innovation” changes under you. Any fidelity judgement made about a generative tool must be dated.

13.12 What this changes in the core workflow

Stage (§6.1)	v1 tool	Add from this Part
Diagnose	CFIR, ORIC	COM-B screen, then TDF for individual barriers; AIM/IAM/FIM pulse; NASSS and the Nilsen AI typology for AI tools
Select	CFIR–ERIC matching tool	Matching tool as a prompt; Implementation Mapping for costly rollouts; one IRLM row per strategy, with mechanism and preconditions
Deliver	Fidelity and adaptation log	Functions vs forms; FRAME fields; a named, dosed facilitator; EPIS bridging factors
Evaluate	RE-AIM with denominators	Iterative RE-AIM at six weeks; label pilots as hybrid types; report with TIDieR
Sustain or stop	RE-AIM maintenance, NPT	PSAT/CSAT at 3 and 12 months; McKay’s three criteria applied to the catalogue once a year

---

Part 14 — What exists in schools, universities, health professions education and workplaces

Bottom line. The four frameworks have now been applied in every sector this document addresses except one. Corporate training and HRD still have no study using CFIR, ERIC, RE-AIM or NPT. Even where the frameworks have been used, the studies are descriptive: they map barriers and report reach, but they do not test whether using the framework changed anything.

The strongest evidence about what moves practice in these sectors comes from the training literatures instead, and part of it is correlational. Three things carry the strongest numbers:

• support in the work environment (correlational)
• coaching with feedback (causal studies)
• engaging, practice-based methods (quasi-experimental)

This agrees with §3.7 on coaching and on feedback tied to action. For audit and feedback, though, repeated delivery now predicts smaller effects. Effects also shrink when a programme is scaled up.

14.1 Workplaces

The framework literature stops at the edge of corporate learning. Searches of Europe PMC and Crossref combining the frameworks with workplace, HRD, transfer-of-training and corporate terms found no application to corporate training [negative search, two indexes].

What does exist sits next door:

• Occupational safety and health. A scoping review coded 30 knowledge-translation studies to CFIR and ERIC. The activities were mostly “train and educate” and “develop relationships” (Chabot et al. 2026) [ABSTRACT].
• Workplace health promotion. A CFIR-coded review of 53 studies found most determinants in the Inner Setting. Evidence strength and leadership engagement were the facilitators; structure, relative priority and resources the barriers (Bhandari et al. 2026) [ABSTRACT].
• RE-AIM reviews of worksite programmes. These mainly expose reporting gaps: implementation and adoption are poorly reported, and maintenance rarely (Schaller et al. 2022) [ABSTRACT].

v1’s claim stands, narrowed: the corporate L&D application is unwritten. Occupational safety and workplace health are not.

The training literatures carry the effect sizes.

• Safety training changes behaviour, not health. A systematic review of 22 randomised studies found strong evidence that it changes worker behaviour, and insufficient evidence that it changes symptoms, injuries or illness (Robson et al. 2012) [ABSTRACT].
• Engaging methods beat passive ones. Across 95 studies and nearly 21,000 workers, knowledge rose and accidents and injuries fell as methods moved from lecture and video, through feedback, to behavioural modelling and hands-on practice (Burke et al. 2006) [ABSTRACT].
• Support predicts transfer, correlationally. In a meta-analysis, peer, supervisor and organisational support jointly accounted for 32% of the variance in training transfer, with peer support the largest share. All three related to whether the trained behaviour was sustained (Hughes et al. 2020) [ABSTRACT].
• Ability, personality, motivation and environment all relate to transfer, more strongly for open than closed skills. Same-source measurement inflates those relationships (Blume et al. 2010) [ABSTRACT].

Two workplace data points bear on AI and trust. Both are working papers.

• Psychological safety. In a randomised encouragement design, changing the content of managers’ one-to-ones toward employees’ needs and aspirations raised team psychological safety by about 1.8 points on a 100-point index (about 0.16 SD) in six weeks (Castro, Englmaier & Guadalupe, CEPR DP20218) [SECONDARY]. The effect was largest in the teams that started lowest. Sample sizes differ between versions of the paper, and there is no follow-up.
• AI take-up. In Danish survey data linked to registers, 83% of workers whose employer encouraged chatbot use had used one, against 47% of those whose employer did not (Humlum & Vestergaard 2025) [SECONDARY]. Employer-run training narrowed the gender gap in use. Encouragement was not randomised, and earnings effects above 1–2% were ruled out.

What this means for learning work (my judgement). In AI-capability and training rollouts, employer and peer support has the strongest numbers and the least framework coverage. Measure it twice:

• as a determinant: the CFIR Inner Setting and leadership engagement
• as an outcome: use, transfer and sustainment

Do that before adding another framework.

14.2 Schools

Coaching changes teaching more than attainment, and shrinks at scale. A meta-analysis of 60 causal studies found teacher coaching improved instruction by 0.49 SD and achievement by 0.18 SD. Large effectiveness trials showed “only a fraction” of the effect of small efficacy trials (Kraft, Blazar & Hogan 2018) [ABSTRACT].

Performance feedback works, on single-case evidence. Feedback to educators on how faithfully they implement a practice meets the What Works Clearinghouse standards for single-case evidence (Fallon et al. 2015) [ABSTRACT].

Implementation-context measures exist for schools. School versions of the implementation leadership and climate scales held their structure in 196 school consultants. One part failed: the rewards subscale, because schools rarely allow individual financial incentives (Lyon et al. 2018) [ABSTRACT].

The scoping review v1 cited without naming is Gallagher et al. 2023 [ABSTRACT]. It covers 22 papers, mostly using determinant frameworks, with inner-setting factors the strongest.

Train-the-trainer can match expert consultation at lower cost. In a cluster trial in urban schools, it produced similar fidelity and student outcomes for about 30% less (Eiraldi et al. 2024) [ABSTRACT].

For strategies, start from SISTER’s 75 (§13.10).

What this means for learning work (my judgement). Plan for dilution. A pilot’s effect size is the ceiling of what a scaled programme will show, not its forecast.

14.3 Universities

The oldest review is also the bluntest. Across 191 articles on changing undergraduate STEM teaching, only 21% of implementation studies gave strong evidence (Henderson, Beach & Finkelstein 2011) [ABSTRACT].

Two strategies were clearly ineffective:

• developing and disseminating curricular materials
• top-down policy

The effective strategies shared three features:

• they addressed or changed teachers’ beliefs
• they lasted at least a semester
• they treated the institution as a system

A later paper lays out eight change strategies as a repertoire, with no comparative effect data (Borrego & Henderson 2014).

Framework applications are recent and descriptive.

• Reusable learning objects. A RE-AIM evaluation across three Malaysian universities covered 23 objects and 7,622 users. Users were self-selected, and the evaluation was pre–post only (Lim et al. 2025).
• An online course. CFIR guided the dissemination of an online nutrition-research course (Daudelin et al. 2025).
• College classrooms. Soicher and colleagues adapt implementation science to the college classroom (Soicher et al. 2020).

All are [ABSTRACT].

What this means for learning work (my judgement). For faculty adoption:

• design for belief change over a semester or longer
• treat the department as the unit
• never rely on releasing good materials

The 2011 review already said this.

14.4 Health professions education and continuing professional development

Faculty development satisfies participants and changes self-reported practice, with little evidence of organisational change. Two BEME reviews (53 and 111 studies) found exactly that (Steinert et al. 2006; Steinert et al. 2016). Their recurring features:

• experiential learning
• feedback
• a community of practice
• longitudinal design
• institutional support

Continuing medical education improves performance more reliably than patient outcomes. Interactive, multi-method, repeated and longer activities do better (Cervero & Gaines 2015) [ABSTRACT].

CPD is starting to be studied as an implementation strategy. One example: dentists who took a one-hour course wrote 0.60 fewer opioid prescriptions over six months than matched controls (Polk et al. 2026) [ABSTRACT].

Two v1 examples, identified.

• The “RE-AIM applied to curriculum” template comes from a conference abstract (Nearing et al. 2025) [VERIFIED]. It reports no results, so treat it as a template, not evidence.
• v1’s “curricular design” source is Gottlieb et al. 2024, a conceptual paper.

What this means for learning work (my judgement). CPD has the longest record of strategy-style evaluation. Use it for effect-size analogies in L&D, labelled as clinical-workforce evidence.

14.5 Champions and train-the-trainer

Two strategies are ubiquitous in learning work and thinly evidenced.

Champions. Most studies record a champion simply as present or absent. In an integrative review of 199 articles, only four studies were randomised (Miech et al. 2018). A systematic review of 35 studies found an association with implementation in five of seven organisation-level studies, and mixed or scarce evidence otherwise (Santos et al. 2022) [ABSTRACT].

Train-the-trainer.

• Knowledge. A meta-analysis found a gain of SMD 0.58, with very high inconsistency (Poitras et al. 2021).
• Competence and adherence. In one trial against expert consultation, train-the-trainer raised competence but did not significantly change adherence (Raghavan et al. 2024) [ABSTRACT].
• No cascade-loss evidence. No study was found showing that fidelity falls at each step of a cascade. v1’s claim to that effect is withdrawn (Appendix D).

14.6 The sector picture in one table

Sector	Framework use	Strongest evidence found (design)	What is missing
Corporate L&D, HRD	None found	Work-environment support (correlational meta-analysis); engaging safety-training methods (quasi-experimental)	Any framework application; any test of a framework-guided rollout
Occupational safety and workplace health	CFIR-coded reviews; RE-AIM reviews	Safety training changes behaviour (randomised studies)	Health and injury outcomes
Schools	Determinant frameworks; SISTER; school context scales	Coaching, 0.49 SD on instruction (causal studies); performance feedback (single-case)	Effects at scale; student outcomes from implementation strategies
Universities	RE-AIM and CFIR descriptive studies	Belief-changing, semester-long, systemic strategies (review-level)	Controlled comparisons of change strategies
Health professions education, CPD	RE-AIM evaluations; CPD as a strategy	CME improves performance more than outcomes	Organisational change from faculty development

---

Part 15 — The bridge to aviation and AI

Bottom line. The aviation report and this document describe one failure from two sides.

The aviation report shows what happens when an artefact is imported without its system. Checklists spread and the mechanism did not travel. Investigators ran out of recommendations. Training was in date and the accident happened anyway.

Implementation science supplies the vocabulary and the instruments to see that failure coming, and to measure it when it happens. The bridge has four spans, and a mapping of the aviation design kit onto the frameworks.

15.1 Completion is not fidelity

The single most useful datum for this bridge comes from 45 gynaecological operations recorded on video in one operating room. The record and the video told different stories:

• What was completed. Sign-in and time-out were completed every time.
• What the items said. Only 47% of checklist items were actually performed as written, against 89% recorded in the medical record (Møller et al. 2025) [ABSTRACT].

That is Dixon-Woods’s phenotype of compliance without its genotype (aviation report, §19.2), measured.

The implication for any evaluation plan:

• Report completion (was it done?) and fidelity (was it done as designed?) as separate indicators under RE-AIM Implementation.
• Measure fidelity by observation or record, never by sign-off.

The same applies to every learning programme whose evidence is a completion record.

15.2 Team training and checklists through implementation lenses

Team training works.

• The effect. A meta-analysis of healthcare team training found effects at every Kirkpatrick level, from d = .37 to d = .89, except that programmes involving feedback did worse (Hughes et al. 2016) [ABSTRACT].
• What happens after it. TeamSTEPPS implementation studies are qualitative and local, and I found none that tracks sustainment with an implementation framework (Velez et al. 2026) [ABSTRACT].

Checklist rollouts have been studied with the frameworks.

• NPT and NoMAD. One rollout bundled champions, audit and feedback, education and prompts. Of 150 staff surveyed, 59 responded (39%), and 77.6% of those saw value in it (Gillespie et al. 2018) [ABSTRACT].
• CFIR. After a national scale-up in Benin, self-reported use was 86%, against 31% before training. Median fidelity was 5 of 7, with no control group (White et al. 2019) [ABSTRACT].
• Strategy count. A review of 47 papers from low- and middle-income settings found no association between the number of implementation strategies used and implementation outcomes (White et al. 2021) [ABSTRACT].

That last finding agrees with §3.7’s second rule: bundle by diagnosis, not by count.

A correction to the usual citation. The best-known English study of checklist barriers (Russ et al. 2015) is often cited as a framework study. It used thematic analysis, not NPT or CFIR.

15.3 The debrief as an implementation strategy

The aviation report identifies the structured debrief as the most defensible single investment it found. Two meta-analyses put its effect on performance in training at d = .67 and .79 (aviation report, §12.5 and Part VI, component 5). The evidence that any particular debriefing method improves patient or operational outcomes is weak, as the reviews below show.

Yet the implementation side is almost empty:

• The methods. A review of 70 studies found simulation debriefing methods “not based on robust empirical evidence” (Duff et al. 2024).
• Clinical tools. Of 21 clinical debriefing tools, one improved patient outcomes, and few said how to turn the discussion into change (Phillips et al. 2024).
• Sustainment. No study of whether simulation or debriefing programmes are sustained, using CFIR or RE-AIM, was found.
• Classification. A 2026 position paper argues that simulation should be treated as an ERIC strategy in its own right (Sanseau et al. 2026).

All are [ABSTRACT].

For learning work, the nearest ERIC labels for a debrief are Audit and provide feedback and Facilitation. Specify it with Proctor’s seven dimensions (§3.5) like any other strategy.

The feedback tension from §3.7 resolves here. Individual feedback is more effective than team feedback in the clinical trials. Aviation shows what individual feedback needs before it is safe:

• a debrief that is never graded
• diagnosis separated from consequential judgement
• legally or contractually protected data

Where aviation has these conditions, they were built deliberately: protection exists for event data but, as IATA has flagged, not yet for individual competency data (aviation report, §8.7 and §18). Individual feedback on practice is safe to use only where all three conditions hold. Otherwise, use team-level data. In the team-training meta-analysis, programmes involving feedback did worse (§15.2), a reminder that how feedback is delivered matters as much as whether it is given.

15.4 Learning-response systems

The NHS Patient Safety Incident Response Framework (PSIRF, 2022) is the closest healthcare attempt at an aviation-style learning system. It shifts incident response from blame-oriented root-cause analysis to systems-based learning.

Its evidence so far is about process only:

• Staff reaction. Staff welcomed the shift, but skills and investment are uneven (HSSIB 2025) [SECONDARY].
• What the reports recommend. A content analysis of mental-health incident reports found that most recommended actions were still administrative controls (Challinor et al. 2026) [ABSTRACT].

The lesson matches the aviation report’s first precondition. A new system label does not produce stronger actions by itself. Evaluate a learning-response system by the strength of the changes it produces, not by its adoption.

15.5 AI rollouts

Ambient AI scribes are the most-studied AI rollout I found in professional work.

They have trials.

• Uptake and effect. In a pragmatic randomised trial of 238 physicians, scribes were used in about 30% of visits. One of the two products cut time spent on notes by 9.5%; the other had no significant effect (Lukac et al. 2025).
• Framework studies. A scoping review mapped 29 scribe studies to CFIR 2.0 (Yu et al. 2026), and RE-AIM/PRISM has been used to study physicians’ views (Shah et al. 2025).

All are [ABSTRACT].

No deskilling check found. I found no study that ran the withdrawal condition: clinicians’ own performance with the scribe removed. The nearest signal is a retrospective matched cohort of 20,302 primary-care notes, in which AI-scribed notes recorded more psychiatric symptoms but were followed by less intervention (aOR 0.83; Castro et al. 2026) [ABSTRACT].

The bridge is concrete.

• Classify the tool by the Nilsen typology: locked, adaptive or generative (§13.11).
• Treat employer support as a determinant: employer encouragement was strongly associated with take-up in Danish survey data (§14.1; not randomised).
• Report Reach (how many eligible physicians used the tool, and how they differ from those who did not) separately from Implementation (the share of visits in which it was used: about 30% in the trial). Neither is an Effectiveness finding.
• Add assistant-absent performance, stratified by baseline, as an Effectiveness indicator. RE-AIM’s current definition already asks for differences between subgroups (§4.2).

15.6 The aviation design kit, mapped onto the frameworks

The aviation report’s Part VI specifies eight design components. Each has a natural home in the frameworks, which is how an implementation plan can carry them.

Aviation component	CFIR (diagnose)	ERIC (act)	RE-AIM or NPT (evaluate)
1. Data channel before training	Inner Setting: culture, information and knowledge access	Develop and implement tools for quality monitoring; Develop and organise quality monitoring systems	NPT Reflexive Monitoring (systematisation); RE-AIM baseline measures
2. Diagnosis apart from jeopardy	Inner Setting: culture (psychological safety), incentive systems	Audit and provide feedback, kept developmental	NPT Reflexive Monitoring (individual and communal appraisal)
3. Observable behaviour, calibrated raters	Innovation: evidence base; Individuals: capability	Develop and implement tools for quality monitoring; Conduct ongoing training (for raters)	RE-AIM Implementation: fidelity measured by observation (§15.1)
4. Consequence setting and instructional setting	Innovation: design; Inner Setting: available resources	Make training dynamic; Conduct ongoing training	RE-AIM Effectiveness, measured in the consequence setting
5. Debrief	Process: reflecting and evaluating	Audit and provide feedback; Facilitation	NPT Reflexive Monitoring (communal appraisal)
6. Cycle set by decay	Process: reflecting and evaluating	Conduct ongoing training; Purposely re-examine the implementation	RE-AIM Maintenance, with an interval set from decay data
7. Withdrawal and stratification	Innovation: adaptability; Individuals	Develop and implement tools for quality monitoring	RE-AIM Effectiveness by subgroup, including performance without the assistant
8. Faded scaffold	Innovation: adaptability, complexity	Promote adaptability; Tailor strategies	FRAME log of planned adaptations; NASSS adaptation over time

The mapping is mine. Its use is practical: it lets an aviation-derived design travel inside a standard implementation plan, with its measures attached, instead of arriving as an artefact.

---

Part 16 — Reading list

The four primary sources

1. Damschroder, L. J. et al. (2009). Fostering implementation of health services research findings into practice: a consolidated framework for advancing implementation science. Implementation Science, 4:50.
2. Damschroder, L. J., Reardon, C. M., Opra Widerquist, M. A. & Lowery, J. (2022). The updated Consolidated Framework for Implementation Research based on user feedback. Implementation Science, 17:75.
3. 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.
4. 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.
5. May, C. & Finch, T. (2009). Implementing, embedding, and integrating practices: an outline of Normalization Process Theory. Sociology, 43(3), 535–554.

Operational companions

1. CFIR Outcomes Addendum (2022), Implementation Science, 17:7 — the implementation vs innovation outcomes distinction.
2. CFIR User Guide (2025) — the five-step method.
3. Waltz, T. J. et al. (2015). Use of concept mapping to characterize relationships among implementation strategies and assess their feasibility and importance. Implementation Science, 10:109 — the nine clusters.
4. Waltz, T. J. et al. (2019). Choosing implementation strategies to address contextual barriers: diversity in recommendations and future directions. Implementation Science, 14:42 — the CFIR–ERIC matching tool.
5. Proctor, E. K. et al. (2011). Outcomes for implementation research: conceptual distinctions, measurement challenges, and research agenda. Administration and Policy in Mental Health, 38(2), 65–76.
6. Finch, T. L. et al. (2018). Improving the normalization of complex interventions: part 2 — validation of the NoMAD instrument. BMC Medical Research Methodology, 18:135.
7. Feldstein, A. C. & Glasgow, R. E. (2008). A practical, robust implementation and sustainability model (PRISM). Joint Commission Journal on Quality and Patient Safety, 34(4), 228–243.

Critical reading — do not skip

1. Kirk, M. A. et al. (2016). A systematic review of the use of the Consolidated Framework for Implementation Research. Implementation Science, 11:72.
2. Systematic reviews of RE-AIM use over time (Gaglio, Shoup & Glasgow, 2013, AJPH 103(6); and the 2021/2022 updated review in Frontiers in Public Health).
3. Systematic review of NPT use in feasibility studies and process evaluations (2018) — the agency and normative-focus critiques.
4. Comparison of CFIR–ERIC matching tool recommendations to real-world strategy effectiveness data, Implementation Science (2023).

Education and learning applications

1. Gottlieb, M. et al. (2024). Incorporating implementation science principles into curricular design. AEM Education and Training.
2. Adapting implementation science for higher education research: the systematic study of implementing evidence-based practices in college classrooms (2020).
3. Greenhalgh, T. et al. (2017). NASSS framework, Journal of Medical Internet Research, 19(11), e367 — for any technology-mediated learning innovation.

Free tools

• cfirguide.org — constructs, definitions, Excel downloads, tools and templates, translations
• re-aim.org — dimension guidance, standard survey items, planning tool, PRISM instruments
• normalization-process-theory.northumbria.ac.uk — NPT toolkit and NoMAD

---

Appendix A — The 73 ERIC strategies, translated for learning

Listed alphabetically as in Powell et al. (2015), Table 3. ERIC definitions are abbreviated; the source wording uses “clinical innovation,” “clinician” and “patient/consumer,” which are rendered here as innovation, deliverer and recipient. The third column is my translation, not part of ERIC.

Relevance key: ●●● directly usable in learning contexts · ●● usable with adaptation · ● largely healthcare-specific

#	Strategy	ERIC definition (abbreviated)	Learning translation	Rel.
1	Access new funding	Access new or existing money to facilitate implementation	Levy funds, capability budgets, grant funding, cost recovery	●●
2	Alter incentive/allowance structures	Work to incentivize adoption and implementation	Recognition, promotion criteria, objectives. Prefer informational and social over material (motivation literature; implementation trials are inconclusive, §3.7)	●●●
3	Alter patient/consumer fees	Fee structures where recipients pay less for preferred options	Internal chargeback design; making the desired option the cheapest for a unit	●
4	Assess for readiness and identify barriers and facilitators	Assess degree of readiness, barriers that may impede, strengths that can be used	The CFIR assessment plus a readiness measure. The step before design	●●●
5	Audit and provide feedback	Collect and summarize performance data over a period; give it to practitioners and administrators	Team-level practice dashboards to managers and heads. Task-directed, not individual surveillance. The largest evidence base (292 trials; median +2.7 points, mean +6.2); see §3.7 on the individual-vs-team tension	●●●
6	Build a coalition	Recruit and cultivate relationships with partners in the implementation effort	Cross-functional sponsor group with real authority	●●●
7	Capture and share local knowledge	Capture how implementers made something work in their setting, then share it with other sites	Direct attack on causal ambiguity. Interview the units that made it work	●●●
8	Centralize technical assistance	Develop and use a centralized system to deliver technical assistance on implementation issues	A central implementation support desk for units rolling it out	●●
9	Change accreditation or membership requirements	Alter accreditation standards, or membership requirements, to encourage use	Professional body requirements, internal certification, standards bodies	●●
10	Change liability laws	Participate in liability reform making deliverers more willing to deliver	Rarely applicable; occasionally relevant to safety-critical training	●
11	Change physical structure and equipment	Adapt physical structure and/or equipment to accommodate the innovation	Rooms, tooling, simulators, screens, workspace layout	●●
12	Change record systems	Change record systems to allow better assessment of implementation or outcomes	Make the practice visible in a system. If you cannot see it you cannot manage it	●●●
13	Change service sites	Change the location of service sites to increase access	Where learning and practice happen; in-the-flow vs off-site	●●
14	Conduct cyclical small tests of change	Implement changes cyclically using small tests before going system-wide, with systematic measurement	Trialability engineered. One team, one term, reversible	●●●
15	Conduct educational meetings	Hold meetings targeted at different stakeholder groups to teach them about the innovation	Briefings, workshops, faculty meetings. Small average effect as the main component (median +4 points); format comparisons inconclusive (§3.7)	●●●
16	Conduct educational outreach visits	A trained person meets deliverers in their practice setting to educate them, intending to change practice	Learning partner going to the team, in situ. Large, consistent evidence of small effects (median +5.6 points), not updated since 2007 (§3.7)	●●●
17	Conduct local consensus discussions	Include local deliverers and stakeholders in discussing whether the problem is important and the innovation appropriate	Participation on the whether and the how. Supports autonomy and surfaces feasibility information	●●●
18	Conduct local needs assessment	Collect and analyze data related to the need for the innovation	Distinct from a learning needs analysis: this is need for the practice	●●●
19	Conduct ongoing training	Plan for and conduct training in an ongoing way	Spacing, in framework language	●●●
20	Create a learning collaborative	Facilitate groups of deliverers or organizations and foster a collaborative learning environment	Cross-unit cohort with shared measurement. The Michigan mechanism	●●●
21	Create new clinical teams	Change who serves on the team, adding different disciplines and skills	Redesign who does the work — pairs, triads, embedded specialists	●●
22	Create or change credentialing/licensure standards	Certify people in the innovation; change certification, licensure or CE requirements	Internal accreditation, CPD requirements, standards-driven training	●●
23	Develop a formal implementation blueprint	Aim/purpose, scope, timeframe and milestones, performance measures; used and updated over time	An implementation plan distinct from a delivery plan	●●●
24	Develop academic partnerships	Partner with a university or academic unit for shared training and research skills	Legitimacy plus evaluation capability. Addresses NPT legitimation	●●●
25	Develop an implementation glossary	Develop and distribute a list of terms describing the innovation, implementation and stakeholders	Shared language. Cheap; supports NPT coherence	●●●
26	Develop and implement tools for quality monitoring	Introduce the right protocols, algorithms, standards and measures into quality-monitoring systems	Practice quality measures, not satisfaction measures	●●●
27	Develop and organize quality monitoring systems	Develop systems and procedures monitoring processes and/or outcomes for assurance and improvement	The measurement infrastructure that makes maintenance possible	●●●
28	Develop disincentives	Provide financial disincentives for failure to implement or use	Use sparingly. Implementation trials are inconclusive; the case against rests on the motivation-crowding literature (companion review, Part 8)	●●
29	Develop educational materials	Develop and format manuals, toolkits and supporting materials to make learning easier	Job aids, playbooks, checklists — for deliverers as much as learners	●●●
30	Develop resource sharing agreements	Develop partnerships with organizations that have needed resources	Shared facilitators, shared platforms, consortium arrangements	●●
31	Distribute educational materials	Distribute guidelines, manuals and toolkits in person, by mail or electronically	Small average effect alone (median +4 points), similar to meetings and feedback; cheap; uncertain when added to other strategies	●●●
32	Facilitate relay of clinical data to providers	Provide near-real-time data on key process/outcome measures through integrated channels	Practice data back to the team fast enough to act on	●●●
33	Facilitation	Interactive problem solving and support in a context of recognized need and a supportive relationship	The largest pooled effect of any strategy (OR ≈ 2.8), on a thinner, publication-biased evidence base with a null pragmatic trial. In practice a bundle of feedback, goal setting and tailoring. Closest analogue: skilled internal coaching	●●●
34	Fund and contract for the clinical innovation	Payers issue RFPs, use contracting, develop funding formulas to motivate delivery	Procurement and contracting levers; commissioning training provision	●●
35	Identify and prepare champions	Identify and prepare individuals who dedicate themselves to supporting and driving implementation, overcoming indifference or resistance	Champions are usually named and abandoned. “Prepare” is the operative word. No verified effect size; feedback involving a local champion with an existing relationship works better (§3.7)	●●●
36	Identify early adopters	Identify early adopters locally to learn from their experiences	Positive deviance. Nearly free, rarely done	●●●
37	Increase demand	Influence the market to increase competition and maturity	Internal demand generation; making the practice sought after	●●
38	Inform local opinion leaders	Inform people identified by colleagues as opinion leaders, hoping they influence others	Note: identified by colleagues, not appointed by management	●●●
39	Intervene with patients/consumers to enhance uptake and adherence	Develop strategies with recipients to encourage and problem-solve around adherence	Working with learners on their own barriers to sustained practice	●●●
40	Involve executive boards	Involve governing structures in the implementation effort, including reviewing implementation data	Board or exec review of implementation metrics, not just outcomes	●●●
41	Involve patients/consumers and family members	Engage or include recipients and families in the implementation effort	Learners on the implementation team, not just in the programme	●●●
42	Make billing easier	Make it easier to bill for the innovation	Internal cost coding, time recording, chargeback simplification	●
43	Make training dynamic	Vary delivery methods to cater to different learning styles and work contexts; make training interactive	⚠ The learning-styles rationale is unsupported. Keep “vary methods, fit the work context, make it interactive”; discard the stated reason	●●●
44	Mandate change	Leadership declares the priority of the innovation and their determination to have it implemented	Works for compliance-type behaviours; poor for judgement-type ones	●●●
45	Model and simulate change	Model or simulate the change prior to implementation	Walkthroughs, dry runs, simulation. Under-used in L&D rollouts	●●●
46	Obtain and use patients/consumers and family feedback	Increase recipient feedback on the implementation effort	Feedback on implementation, distinct from course feedback	●●●
47	Obtain formal commitments	Written commitments from key partners stating what they will do	Named, written manager and sponsor commitments	●●●
48	Organize clinician implementation team meetings	Support teams implementing the innovation with protected time to reflect, share lessons and support each other	The “protected time” clause is what makes it work	●●●
49	Place innovation on fee for service lists/formularies	Get the innovation onto reimbursable lists	Getting the practice into the approved catalogue or standard operating model	●●
50	Prepare patients/consumers to be active participants	Prepare recipients to be active, ask questions, inquire about evidence	Preparing learners to expect and demand transfer support	●●●
51	Promote adaptability	Identify how the innovation can be tailored locally and clarify which elements must be maintained to preserve fidelity	The core/adaptable specification in a single strategy	●●●
52	Promote network weaving	Build on existing high-quality relationships and networks to promote information sharing, collaborative problem-solving and shared vision	Complex contagion, operationalised	●●●
53	Provide clinical supervision	Ongoing supervision focusing on the innovation; train the supervisors	Manager coaching capability, and coaching the coaches	●●●
54	Provide local technical assistance	Deliver technical assistance on implementation issues using local personnel	Embedded local support rather than central	●●●
55	Provide ongoing consultation	Ongoing consultation with experts in the strategies used to support implementation	Retained implementation advisory, not content expertise	●●●
56	Purposely reexamine the implementation	Monitor progress and adjust practices and strategies to continuously improve	The reconfiguration mechanism in NPT terms	●●●
57	Recruit, designate, and train for leadership	Recruit, designate and train leaders for the change effort	Implementation leads, properly resourced	●●●
58	Remind clinicians	Reminder systems to help recall information and prompt use	Prompts, nudges, calendar structures, checklists in the workflow	●●●
59	Revise professional roles	Shift and revise roles among practitioners; redesign job characteristics	Structural. Almost never on an L&D plan and often the actual answer	●●●
60	Shadow other experts	Provide ways for key individuals to directly observe experienced people using the practice	Vicarious mastery. Second strongest self-efficacy source	●●●
61	Stage implementation scale up	Phase efforts starting with small pilots or demonstration projects, gradually moving system-wide	The stepped wedge, as a strategy	●●●
62	Start a dissemination organization	Identify or start a separate organization responsible for disseminating the innovation	A dedicated academy, centre of excellence or spin-out	●●
63	Tailor strategies	Tailor implementation strategies to barriers and facilitators identified through earlier data collection	The CFIR→ERIC step itself	●●●
64	Use advisory boards and workgroups	Create and engage a formal multi-stakeholder group to provide input and recommendations	Genuine advisory group with a route to decisions	●●●
65	Use an implementation advisor	Seek guidance from experts in implementation	Distinct from subject-matter or content expertise	●●●
66	Use capitated payments	Pay providers a set amount per recipient for delivering	Per-head funding models for training provision	●
67	Use data experts	Involve, hire or consult experts to inform management on the use of data generated by implementation	Analytics capability inside the implementation team	●●●
68	Use data warehousing techniques	Integrate records across facilities and organizations to facilitate implementation across systems	Joining learning, HR and performance data	●●
69	Use mass media	Use media to reach large numbers of people	Internal campaigns. Necessary, never sufficient	●●●
70	Use other payment schemes	Introduce payment approaches (catch-all)	Funding model changes	●
71	Use train-the-trainer strategies	Train designated people or organizations to train others	Ubiquitous in L&D; note it is a strategy. Cascades risk diluting the core functions; direct evidence on loss per step was not found	●●●
72	Visit other sites	Visit sites where a similar effort has been considered successful	Site visits and exchanges. Cheap, powerful, rarely funded	●●●
73	Work with educational institutions	Encourage educational institutions to train people in the innovation	University and provider partnerships shaping the pipeline	●●●

Reading the list. 54 of the 73 are rated directly usable in learning contexts with no more than vocabulary substitution. The healthcare-specific residue is concentrated in the financial cluster. Anyone claiming these frameworks don’t transfer to L&D has not read the list.

---

Appendix B — CFIR 2.0 quick-reference checklist

For rapid assessment. Full definitions in Part 2.

I. Innovation — Source · Evidence-Base · Relative Advantage · Adaptability · Trialability · Complexity · Design · Cost

II. Outer Setting — Critical Incidents · Local Attitudes · Local Conditions · Partnerships & Connections · Policies & Laws · Financing · External Pressure (Societal · Market · Performance-Measurement)

III. Inner Setting Persistent: Structural Characteristics (Physical · IT · Work Infrastructure) · Relational Connections · Communications · Culture (Human Equality-Centeredness · Recipient-Centeredness · Deliverer-Centeredness · Learning-Centeredness) Innovation-specific: Tension for Change · Compatibility · Relative Priority · Incentive Systems · Mission Alignment · Available Resources (Funding · Space · Materials & Equipment) · Access to Knowledge & Information

IV. Individuals Roles: High-level Leaders · Mid-level Leaders · Opinion Leaders · Implementation Facilitators · Implementation Leads · Implementation Team Members · Other Implementation Support · Innovation Deliverers · Innovation Recipients Characteristics (assess per role): Need · Capability · Opportunity · Motivation

V. Implementation Process — Teaming · Assessing Needs (Deliverers · Recipients) · Assessing Context · Planning · Tailoring Strategies · Engaging (Deliverers · Recipients) · Doing · Reflecting & Evaluating (Implementation · Innovation) · Adapting

The pragmatic eight, if you only have time for a rapid assessment: Relative Advantage (adopter’s view) · Complexity · Trialability · Tension for Change · Compatibility · Relative Priority · Available Resources · Mid-level Leader Capability/Opportunity/Motivation.

---

Appendix C — Verification log

v2 note. The v1 log below is kept as written, with corrected rows marked “(v2 corrected)”. The full v2 verification record is a claim-by-claim register with rulings and sources, kept in the research files (not published here). Corrections that changed what the document says are listed in Appendix D.

C.1 Verified against primary or official sources

Item	Source	Status
Full CFIR 2.0 domain and construct list, with official definitions	cfirguide.org constructs page (official CFIR site)	Verified verbatim; definitions quoted as published
CFIR original: 39 constructs, five domains, Implementation Science 4:50 (2009)	Multiple secondary sources	Verified
CFIR update 2022: Implementation Science 17:75; based on literature review plus survey of CFIR authors; centred recipients; added equity determinants; maps back to original	Springer / Implementation Science	Verified
CFIR Outcomes Addendum: anticipated vs actual; implementation vs reach vs innovation effectiveness; implementation vs innovation determinants	Implementation Science / cfirguide.org	Verified
CFIR User Guide 2025, five-step guide	PMC	Verified
ERIC: 73 strategies, three-phase modified Delphi, Powell et al. 2015, Implementation Science 10:21	Verified; full Table 3 with all 73 definitions retrieved from source PDF	Verified verbatim
ERIC nine clusters and their names	Waltz et al. 2015 concept mapping study	Cluster names verified. See C.2
Cluster merge detail (two clusters merged into “Develop stakeholder interrelationships”; strategy #62 moved from Engage consumers to Change infrastructure)	Waltz et al. 2015	Verified
CFIR–ERIC matching tool: 435 invited, 169 responded; top-seven ranking; Level 1 ≥50%, Level 2 20–49.9%; average 47 strategies endorsed per barrier (original CFIR constructs); caution advised	Waltz et al. 2019 full text	Verified (v2 corrected)
VA mixed-methods comparison of matching tool recommendations to real-world effectiveness	Implementation Science 2023	Verified
RE-AIM: Glasgow, Vogt & Boles 1999, AJPH 89(9), 1322–1327; five dimensions; >2,800 citing publications by 2019	re-aim.org; Glasgow et al. 2019	Verified (v2 corrected the count)
RE-AIM maintenance individual-level standard: ≥6 months after most recent intervention contact	re-aim.org	Verified verbatim
RE-AIM under-reporting pattern: maintenance least reported; representativeness, cost, sustained implementation under-reported; qualitative methods rare	Systematic reviews (AJPH 2013; Frontiers 2021/22)	Verified
PRISM: four contextual domains plus RE-AIM outcomes; 200+ applications	re-aim.org and citation analysis / scoping review	Verified
NPT: four constructs; action theory concerned with what people do rather than attitudes or beliefs	normalization-process-theory.northumbria.ac.uk (official)	Verified verbatim
NPT sub-constructs — Coherence (differentiation, communal specification, individual specification, internalization) and Cognitive Participation (initiation, enrolment, legitimation, activation)	Official NPT site and applied studies	Names verified; v2: §5.2 definitions are paraphrases, two of these depart from the official wording
NPT sub-constructs — Collective Action (interactional workability, relational integration, skill set workability, contextual integration)	Official NPT site and applied studies	Names verified; v2: §5.2 definitions are paraphrases, two of these depart from the official wording
NPT sub-constructs — Reflexive Monitoring (4.1 systematization, 4.2 communal appraisal, 4.3 individual appraisal, 4.4 reconfiguration)	Official NPT site, reflexive monitoring page	Verified verbatim, definitions and worked examples quoted from source
NoMAD: 23 items, 3 general (0–10) + 20 across four subscales (CO 4, CP 4, CA 7, RM 5); CFA (parcels) CFI .95, TLI .93, RMSEA .08, SRMR .03; α CO .71, CP .81, CA .78, RM .65; overall .89; 831 returned, 413 complete	Finch et al. 2018 full text	Verified (v2 corrected)
Proctor et al. 2011 eight implementation outcomes	Multiple	Verified
Kirk et al. 2016 CFIR systematic review: 26 studies, 11 (42%) assessing implementation outcomes, four recommendations	Implementation Science full text	Verified (v2 corrected the share)
NPT critiques: terminology, agency emphasis, normative focus, equity/power gap (the “single-traversal” critique in v1 had no source and was removed in v2)	May et al. 2018 systematic review; May, Finch & Rapley 2026	Verified (v2)
Education applications: medical education curricular design, RE-AIM curriculum mapping, higher education adaptation, school-based scoping review, reusable learning objects	Multiple 2020–2024 sources	Verified

C.2 Inferred or partially verified — treat with care

• Waltz et al. 2019: now verified (Implementation Science 14, 42; v2).
• Friendship Bench: now checked against the protocol (v2). RE-AIM → CFIR → strategy design is confirmed; use of the CFIR–ERIC tool is not stated (§6.2).
• Not completed in v2:
  • the EEF guidance report Putting Evidence to Work
  • effect sizes for performance feedback to teachers (Noell et al. 2014)
  • discontinuation rates among faculty who adopt research-based teaching (Dancy et al. 2016)
  • Stains et al. (2018) observation data
  • Robson et al. (2012) effect sizes
  • Blume et al. (2010) ρ values

C.3 Judgements, not findings

The following are my reasoning, clearly marked as such throughout the document: all L&D translations in the third columns; the Unit A–D typology in Part 7.1; the six reframes in §7.3; the worked applications in Part 8 (illustrative reasoning, not case findings); all instruments in Part 9, which adapt published structures but are not themselves validated; the protocols in Part 10; and the failure modes in Part 12.

The NoMAD-style instrument in §9.4 is an adaptation and is not validated. It preserves NoMAD’s structure and sub-construct mapping, but the items are rewritten for learning contexts. Use it as a structured conversation and a trend indicator, not as a psychometric measure. If you need a validated instrument, use NoMAD itself and translate the wording minimally.

---

Appendix D — What changed from v1

v2 keeps v1’s structure and adds to it. This appendix lists every change that alters what the document claims. Changes to wording, links and formatting are not listed. The full claim-by-claim record is kept in the research files (not published here): rulings, sources and verification notes.

D.1 Corrections

#	Where	v1 said	v2 says	Source
1	§0.3, App. B	CFIR 2.0 has “48 constructs and subconstructs”	48 constructs and 19 subconstructs in five domains	Damschroder et al. 2022
2	§3.2	Cluster contents “representative, not authoritative”	Full cluster membership from the source; e.g., Provide ongoing consultation belongs to Train and educate	Waltz et al. 2015
3	§3.1	One strategy was moved between clusters	Three were moved; the nine clusters reflect team merges and moves, not the raw algorithm	Waltz et al. 2015
4	§3.2, §7.3, Part 12	Training is “one ninth of the available levers”	One cluster of nine, holding 11 of 73 strategies (15%), rated high on importance and feasibility	Waltz et al. 2015
5	§3.4	“169 implementation experts were invited”	435 invited, 169 responded (39%). The average of 47 strategies per barrier refers to the 39 original CFIR constructs	Waltz et al. 2019
6	§3.4	A 2023 VA study “found the relationship imperfect”	Sites did not use recommended strategies more often. But 70% of the top-20 recommended strategies were associated with better outcomes, against 48% of the strategies actually used	Yakovchenko et al. 2023
7	§3.4	“A 2026 scoping review” (unnamed)	Named: 53 studies; only 15.7% validated the strategies the tool generated	Meng et al. 2026
8	§3.6	Multifaceted interventions outperform single techniques	Withdrawn. No compelling evidence that multifaceted interventions beat single-component ones	Squires et al. 2014; Grimshaw et al. 2004
9	§3.5	Proctor et al. “require” seven elements as a reporting standard	They recommend seven dimensions; StaRI is the reporting standard	Proctor et al. 2013; Pinnock et al. 2017
10	§0.3, §4.1	RE-AIM cited in about 700 publications	More than 2,800 by 2019	Glasgow et al. 2019
11	§4.2–4.3, Part 12, §9.5	Maintenance at six months or more after the last contact is “a usable, unambiguous standard”	Six months is the older convention; current guidance tailors the period to the project	re-aim.org
12	§4.2	Dimension definitions	Current official wording added: reasons for participation or non-participation; subgroup differences; staff; adaptations	D’Lima et al. 2022
13	§4.5	Setting-level maintenance reported less often than individual-level	Reported more often (60% vs 20%) in the review’s sub-sample	D’Lima et al. 2022
14	§4.6	RE-AIM has no satisfaction dimension “deliberately”; four of five dimensions lack a Kirkpatrick equivalent	“Deliberately” removed (no source). The recount is three dimensions and half of a fourth	Internal consistency
15	§4.4	PRISM domains in their 2008 wording	2024 labels	Glasgow et al. 2024
16	§5.1	NPT “is not a determinant framework and should not be used as one”	An explanatory theory of implementation work, with classificatory and process elements	May, Finch & Rapley 2026
17	§5.2	NPT as four constructs with four sub-constructs each	The 2022 coding manual: 12 primary constructs in context–mechanism–outcome form; the 16 sub-constructs become secondary codes	May et al. 2022
18	§5.2	Sub-construct definitions presented as the theory’s own	Labelled as paraphrases; four depart from the official wording	NPT toolkit
19	§5.3	NoMAD’s general items on a 1–10 visual analogue scale; 831 completed questionnaires	0–10 items on familiarity and on current and future normality; 831 returned, 413 complete	Finch et al. 2018
20	§5.3	NoMAD examined alongside ORIC	Confirmed, with a caveat: test–retest reliability was suboptimal for both	Batterham et al. 2024
21	§5.4	NPT is bad at telling you what to do about a deficit	An NPT-based taxonomy now offers 24 general and 96 micro-strategies	May et al. 2025
22	§5.5	A “single-traversal architecture” critique, attributed to NIHR Open Research	Removed. No source states it	May, Finch & Rapley 2026
23	§5.5	Studies critiqued NPT’s agency emphasis and normative focus	About 10% of 108 studies critiqued NPT, mostly its terminology; two the agency emphasis, one the normative focus	May et al. 2018
24	§7.1, §8.3	NASSS: complexity in three or more domains predicts non-adoption	NASSS concludes cautiously that complexity in multiple domains is the greatest challenge. The threshold is labelled as my heuristic	Greenhalgh et al. 2017
25	§2.5, §11.2	Kirk et al.: “over half did not identify implementation outcomes”	Only 11 of 26 studies (42%) assessed implementation outcomes	Kirk et al. 2016
26	§2.7, §8.1	“The psychological safety RCT”	Labelled as a 2022 working paper; no journal version located	Castro et al. 2022
27	App. A	“Roughly 55 of the 73” strategies directly usable	54	Count
28	§3.1, §3.6	Strategy labels “meaningless”; reviews “consistently” find inadequate specification	Softened to what the sources say	Waltz et al. 2015; Proctor et al. 2013
29	§3.3, App. A	Evaluative notes on strategies: facilitation the “best-evidenced”, outreach “strong”, materials the “weakest”, a cascade “degrades fidelity at each step”	Replaced by the effect sizes in §3.7	§3.7
30	§2.4	A blank rating means the construct is “absent from the data”	A blank means not asked about, or not corresponding; a strong rating needs two or more interviewees with explicit examples	Damschroder & Lowery 2013
31	§6.2	Friendship Bench used RE-AIM → CFIR → CFIR–ERIC matching tool	RE-AIM → CFIR → strategy design confirmed; use of the matching tool not stated	Verhey et al. 2021, 2022
32	§10.1–10.2	Interview numbers presented without source	Labelled as my rule of thumb; CFIR gives no number	Reardon et al. 2025
33	§7.4	Unnamed education sources	Named: a conference abstract (curriculum), Lim et al. 2025, Gallagher et al. 2023	Part 14

D.2 Additions

• §3.7, what the strategies actually achieve. Effect sizes for about a dozen strategies, three rules for setting expectations, the tension between individual and team feedback, and a transfer caveat.
• Part 13, the wider toolkit. Organised by job: choosing and specifying strategies, individual behaviour, phasing, adaptation, sustainment, de-implementation, measures, study design, sector derivatives and AI.
• Part 14, the evidence from schools, universities, health professions education and workplaces.
• Part 15, the bridge to aviation and AI.
• Smaller additions.
  • the NPT 2022 box
  • RE-AIM’s current wording
  • citations for the Nilsen typology and COM-B
  • verification tags
  • links on nearly all load-bearing citations

D.3 What did not change

The core argument stands. Learning and development evaluates the innovation and not the implementation (§0.2), and its most consequential error is choosing the programme as the innovation when the behaviour is the innovation (§7.1). Nothing found in v2 weakens either point.

§3.7 strengthens the practical conclusion. When single strategies are worth a few percentage points and bundles do not reliably add up, diagnosis and measurement matter more than the size of the menu.

---

v1 compiled 10 August 2026; v2 compiled 24 October 2026. Companion to Learning and Behaviour Change: A Critical Review of the Literature (v2) and to Learning to Fly the Machine That Flies (v2). Framework versions current as of v2: CFIR 2.0 (2022) with the 2025 User Guide; ERIC (2015) with the 2019 matching tool; RE-AIM (1999; definitions as updated) with PRISM (2024 labels); NPT (2009) with the 2022 coding manual, the 2025 strategy taxonomy and NoMAD (2018).