ADKAR in Practice for AI Programs

COMPEL Specialization — AITE-WCT: AI Workforce Transformation Expert Article 19 of 35

ADKAR’s five-stage model is, for AI workforce transformation, the most practically useful individual-level change lens the field has produced. Its five stages — Awareness, Desire, Knowledge, Ability, Reinforcement — each name a distinct cognitive or behavioural state and each name the specific intervention class that moves a learner through the stage. The model’s popularity is not accidental; in a decade of corporate practice it has survived repeated methodological scrutiny, and in AI-specific applications its diagnostic power is even higher than in traditional change settings because AI adoption splits the population sharply across the five states.

The expert’s use of ADKAR is not as a checklist. The common failure, addressed head-on in this article, is to treat ADKAR as a sequence of activities (“we did awareness, now we do desire”). ADKAR is a diagnostic. Each stage is a question about where the learner actually is; each intervention is targeted at a specific population segment stuck at a specific stage. A well-run ADKAR campaign spends most of its resources on the one or two stages that are actually blocking the population, not on marching everyone through all five.

The five stages — what each actually measures

• Awareness. The learner knows that change is happening, what it is, and why. Measured by: can the learner, unprompted, name the change and articulate the business reason? Signal of blockage: survey items on “I don’t know why we’re doing this” score high.
• Desire. The learner wants to participate in the change, or at least accepts it as legitimate. Measured by: does the learner engage voluntarily, or resist? Signal of blockage: opt-in rates are low; voluntary behaviours (e.g., attending optional sessions) flatline.
• Knowledge. The learner has the information required to change. Measured by: can the learner, in an assessment, correctly name the components of the new way of working? Signal of blockage: assessment scores are low even after training.
• Ability. The learner can apply the knowledge in live practice. Measured by: does observed behaviour match the target behaviour? Signal of blockage: knowledge scores are fine, but adoption metrics stall.
• Reinforcement. The behaviour change is sustained over time, not a spike. Measured by: does adoption at week 12 match adoption at week 2? Signal of blockage: adoption decays rapidly after initial training.

A learner can have Awareness without Desire (they understand but resist), Knowledge without Ability (they know the theory but cannot apply it), Ability without Reinforcement (they can do it but drift back). Each combination is a different diagnostic problem with a different intervention.

AI-specific stage blockers

The AI context amplifies certain blockers.

Awareness is rarely the blocker, but the framing is easily wrong. Most employees in 2026 are aware that AI is coming to their organisation; many are over-aware, in the sense of having consumed public commentary whose tone (existential, utopian, dystopian) does not match what the organisation is actually doing. The Awareness intervention is less “announce that change is happening” and more “calibrate what change is happening, specifically, for this role, in this organisation, in what time horizon.” Generic “AI is transforming the future of work” communications usually score well on delivery metrics and poorly on calibrated awareness.

Desire is the usual blocker for AI adoption. The blocker has three typical sources: job-security concern (will my role retire), identity concern (will I still be valuable), and meaning concern (what is the point of my work if an AI drafts the first version). The Desire intervention requires line-manager conversations, not a communication from corporate. Prosci research across AI deployments has repeatedly flagged that corporate communications are usually ineffective at Desire; a trusted local voice is the only intervention that materially moves the stage.

Knowledge is widely under-served. Organisations that have invested in literacy programmes (§§12–17) address the Knowledge stage at some depth. Organisations that have not tend to assume Knowledge is produced by communication (it is not) or by self-study (it rarely is, at scale).

Ability is the second-largest blocker. A learner can pass a literacy assessment and still fail to apply the knowledge in their daily work. The Ability intervention is applied practice with feedback: pair programming with a colleague, manager-coached use in live work, sandbox environments where the learner can fail safely. Abstract training without application produces Knowledge, not Ability.

Reinforcement is systematically under-resourced. The pattern is predictable: the training campaign runs, adoption spikes, the sponsor declares victory, and 90 days later adoption has decayed to baseline. The Reinforcement intervention is a small-but-continuous cadence of coaching, refresher content, success-story communication, and performance-system alignment. It is cheap to design and expensive to execute, because it has no concluding moment. The expert’s job is to defend the Reinforcement resource from the organisation’s natural instinct to redirect it to the next initiative.

Diagnosing the blocking stage

ADKAR in practice is a diagnostic-first discipline. A disciplined practice runs a short diagnostic instrument across the target population, scores each learner against the five stages, and aggregates to a population profile. The instrument is typically 10–15 items covering all five stages; Prosci publishes a version, and competent practitioners build internal versions calibrated to their organisation. The diagnostic must be run at least once before the intervention design, and ideally twice — mid-campaign and end-campaign — to inform adjustment.

The aggregation output is a population stage profile: a bar chart showing, for the target population, the fraction of learners currently stuck at each stage. The profile names the dominant blocker. A population with 70% stuck at Desire and 15% stuck at Knowledge is a very different intervention problem from a population with 70% stuck at Ability and 20% stuck at Reinforcement. The first needs line-manager dialogue; the second needs applied practice.

A common failure here: practitioners run the diagnostic, and then design the intervention against their pre-existing bias (usually toward Knowledge, because training is the visible default intervention). The expert’s discipline is to let the diagnostic drive the plan, even when the plan is uncomfortable (e.g., “we need to do hundreds of line-manager conversations rather than a comms campaign”).

Segmentation — not everyone is at the same stage

The aggregate profile hides a detail that matters. Within the population, different segments are at different stages. Early adopters may be at Reinforcement while laggards are at Awareness; managers may be at Desire while their teams are at Knowledge; one business unit may have 80% in Ability while another has 30% in Desire.

The intervention plan therefore segments. The common segmentations:

• By adoption stage. Early adopters, early majority, late majority, laggards. Each gets different content and cadence.
• By role family. Senior leaders, line managers, individual contributors. Each needs different framing.
• By business unit or function. Different units have different baselines and blockers.
• By geography. Cultural norms around voluntary participation, feedback, and authority vary.

A segmented plan is more work to design and execute, but its ROI is materially higher than a uniform one. The uniform plan spends its budget on learners who do not need that intervention and fails learners who need a different one.

Movement measurement

Measuring movement through ADKAR is not the same as measuring training completion. Movement is measured by re-running the diagnostic and observing population profile shift. A successful campaign moves the dominant stage rightward over time: at campaign start, 70% at Desire; at week 8, 60% at Knowledge; at week 16, 55% at Ability; at week 24, 50% at Reinforcement.

The measurement cadence is monthly for short campaigns (3–6 months) or quarterly for longer ones. Each measurement cycle triggers a plan adjustment: which interventions need to continue, which need to stop (because that stage is no longer the blocker), which need to start.

Vanity measurement — training hours delivered, completion rates, session satisfaction — is a trailing indicator at best and an entirely misleading indicator at worst. A campaign with 95% training completion and no stage movement has produced nothing but attendance records.

Reinforcement — the underinvested stage

Reinforcement deserves its own treatment. Three intervention classes are relevant, each with a different cost profile.

• Manager coaching cadence. The highest-impact Reinforcement intervention. A 15-minute weekly conversation between manager and direct report, for the first 12 weeks post-training, with a light structure (“what have you tried, what worked, what got in the way, what will you do next week”). Cost: manager time. Impact: substantial. Feasibility: depends on manager capability, which the manager-enablement curriculum (§28) builds.
• Performance-system alignment. Goal-setting and review processes that explicitly include AI-augmented work. Covered fully in §29. Cost: HR redesign time. Impact: long-term, structural. Feasibility: high if performance-system redesign is already scoped.
• Refresh content and communications. Low-cost, moderate-impact. Short refresh modules at weeks 4, 12, 24. Success-story communications that name specific employees and outcomes (with their permission). Cost: content design and comms time. Impact: reinforces without overwhelming. Feasibility: high.

The expert typically uses all three, weighted to manager coaching as the lead intervention because it is both the most effective and the most organisationally revealing — if managers cannot or will not do the coaching cadence, the programme has a manager-enablement problem masquerading as an employee-adoption problem.

Two real-world anchors

Prosci benchmarking — the recurring “Desire is the blocker” finding

Prosci’s Best Practices in Change Management benchmark reports, published periodically since the late 1990s, have consistently identified Desire as the most frequently under-resourced and most frequently blocking stage. The finding is replicated across industries and change-type categories. In AI-specific deployments — both the 2023 benchmark and the 2024 update — Desire remained the dominant blocker, with job-security concern named as the most common underlying driver. Source: https://www.prosci.com/research and Prosci benchmark publications.

The lesson for the expert is not that ADKAR is about Desire; it is that Desire is where most programmes under-invest relative to the diagnostic evidence, and that a disciplined diagnostic-first practice will usually reallocate budget from Awareness (already over-resourced) toward Desire (under-resourced).

MIT Sloan research on manager coaching cadence

MIT Sloan Management Review’s AI-Human Collaboration series (2023–2025) has repeatedly documented the role of manager-level coaching as a determinant of AI tool adoption at the individual-contributor level. The research — published across a sequence of articles and practitioner briefs — identifies manager coaching cadence as the strongest correlate of sustained adoption six to twelve months post-training, with correlations that significantly outperform those of the training programme design itself. Source: https://sloanreview.mit.edu/topic/artificial-intelligence/.

The lesson for the expert is that Reinforcement is not optional. A programme that invests heavily in Awareness and Knowledge but under-invests in Reinforcement will produce an adoption spike followed by a decay to baseline, and the evidence to explain the decay pattern is not mysterious — it is the absence of the manager cadence.

Learning outcomes — confirm

A learner completing this article should be able to:

• Name the five ADKAR stages and the diagnostic signal of blockage at each.
• Design a diagnostic instrument that produces a population stage profile rather than a completion rate.
• Segment a population by adoption stage, role family, business unit, and geography, and apply differentiated interventions.
• Defend Reinforcement investment against the organisation’s natural redirection to the next initiative.
• Argue why Desire is the typical AI-programme blocker and why corporate communications are usually the wrong intervention for it.
• Measure movement across ADKAR stages by periodic re-diagnosis rather than by training-output vanity metrics.

Cross-references

• EATF-Level-1/M1.6-Art05-Change-Management-for-AI-Transformation.md — Core Stream change-management anchor.
• Article 18 of this credential — ADKAR in context of choice framework.
• Article 20 of this credential — Kotter (enterprise scope).
• Article 22 of this credential — change saturation (constraint on pacing).
• Article 28 of this credential — manager enablement (required for Reinforcement).
• Article 29 of this credential — performance-system alignment (structural Reinforcement).

Diagrams

• StageGateFlow — ADKAR five stages with diagnostic signal and primary intervention per stage.
• Matrix — stage × intervention × population segment, showing differential intervention assignment.

Quality rubric — self-assessment