Back to blogs

// What's on our mind

What ADKAR and Kotter miss when the change is AI

We look at where the ADKAR and Kotter change models fall short for AI adoption and how a continuous practice loop on live work closes the gap.

What ADKAR and Kotter miss when the change is AI

ADKAR and Kotter still help you plan change, yet both assume a finish line that AI adoption never reaches.

Most AI programs at banks and insurers stall well before they pay back. Close to 95% of generative AI pilots have produced no measurable impact on profit and loss, and the gap rarely traces back to the model itself. It traces back to habits, approvals and workflows that never moved. Classic change management models were built to move exactly those things. So why do they keep falling short with AI?

The answer is timing. ADKAR and Kotter treat change as a project with a defined destination, and AI shifts the destination every quarter. You'll get far more from both models once you run AI adoption as a practice loop on live work, with people learning when to trust the output and when to push back.

Why classic change models assume a fixed end state

ADKAR and Kotter were designed for transitions with a clear before and after, such as a new core banking system, a merger or a reorganization. Each model walks people from the current state to a known future state, then works to lock that state in so the organization doesn't slide back.

That logic works when the target holds still. A bank replacing its loan origination platform knows what the new screens, approvals and reports will look like on go-live day. Training can be scripted, reinforcement can be measured against a stable process, and leaders can declare the change complete once usage settles. AI tools rarely offer that kind of fixed target, and that's where the trouble starts.

What changes when the tools keep moving every quarter

AI adoption has no stable end state because the tools, their costs and their sensible use cases shift every few months. A workflow that made sense in January can look dated by June, so any plan that ends by reinforcing one fixed process starts aging on the day it launches.

The economics alone show the pace. The cost of running a model at a fixed benchmark level fell more than 280-fold in about two years, which reopens use cases that teams had written off as too expensive. A wealth management team that rejected AI meeting summaries on cost in one budget cycle may find them cheap enough to run on every client call in the next.

Each shift resets part of the adoption curve. People who mastered last quarter's prompt patterns now face new features, new limits and new guardrails from risk teams. Change work that stops at go-live leaves them to absorb those shifts alone. That's how early momentum quietly fades.

‍

AI model cost dropThe cost of running a model at a fixed benchmark level fell more than 280-fold in about two years, which reopens use cases teams had written off.More than280-folddrop in model costBeforeAbout two years laterThe cost of running a model at a fixed benchmark level fell more than 280-fold in about two years.

‍

Where ADKAR breaks down for AI adoption

ADKAR breaks down at Knowledge and Ability. Both stages assume a defined skill you can teach once and then confirm, yet AI skills keep moving as the tools move. People can complete training, pass the knowledge check and still lack the judgment to use AI well on live work.

Consider a claims team at an insurer. Awareness and Desire come easily once leaders explain the case for AI-assisted claim summaries, and Knowledge arrives through a two-hour session. Ability is where it falls apart, because summarizing a messy claim file with AI means knowing which details the tool tends to drop. Nobody learns that in a classroom.

‍

"People can complete training, pass the knowledge check and still lack the judgment to use AI well on live work."

‍

Reinforcement then rewards the wrong signal. Completion rates and survey scores look healthy while adjusters quietly rebuild summaries by hand. ADKAR still gives you a useful vocabulary for finding the stuck stage, as long as Ability becomes a loop of practice and feedback on real files that repeats each time the tool shifts.

That repetition has a cost worth naming early. Teams need protected time to practise, and someone has to own the feedback. Budget for both up front, or the last two stages of the model will stall right on schedule.

Where Kotter's 8 steps lose momentum with AI

Kotter's model loses momentum at its final steps, which aim to consolidate gains and anchor the change in culture. Anchoring assumes the change is finished. With AI, the thing being anchored keeps moving, so the culture worth building is one that expects regular adjustment over permanence.

The early steps still earn their place. A sense of urgency, a guiding coalition and short-term wins all help a capital markets desk get started with AI research assistants. The trouble comes later, when leaders institutionalize one approved workflow and freeze it in policy. Six months on, the tool has new capabilities, the policy forbids using them, and analysts route around the rules with personal accounts that compliance can't see.

How ADKAR compares with Kotter for technology change

The main difference between ADKAR and Kotter is the level they work at. ADKAR tracks readiness one person at a time, while Kotter organizes leadership action across the whole organization. For AI, ADKAR helps you find who is stuck and Kotter helps you build coalitions, yet neither covers continuous relearning.

Most AI programs need both lenses, plus the piece both leave out. The comparison below shows where each model helps and where you'll need to add practice and judgment yourself.

‍

Planning questionADKARKotterWhat AI adoption adds
Who the model focuses onIndividual employees moving through 5 stagesLeaders mobilizing the organization through 8 stepsTeams practising on live work while managers coach the habits
What it assumes about the end stateEach person reaches a defined future stateA new way of working gets anchored in cultureThe target shifts as tools and use cases move each quarter
How skills get builtKnowledge and Ability stages, usually through trainingRemoving barriers and empowering broad actionRepeated practice on real tasks, refreshed when the tool moves
How success gets measuredEach person's progress through the stagesShort-term wins and sustained resultsThe share of work running through the new path over time
Where it falls short for AIIt treats ability as a one-time milestoneIt locks in a workflow that soon goes out of dateIt builds review habits so people know when to trust or challenge output
How long the work runsUntil reinforcement holdsUntil the change is part of the cultureAs long as the tools keep shifting, so it never fully stops

‍

Use the table as a planning check. If your plan covers every row in the ADKAR and Kotter columns and nothing in the last one, you've planned a launch and left out the adoption.

Running AI change as a continuous practice loop

A continuous practice loop replaces the one-time transition with a repeating cycle. Teams try AI on a live task, review the output against their own standards, share what worked and adjust the workflow. The cycle repeats on a set cadence, and each new tool release simply feeds the next round.

A payments operations team might run it in two-week sprints. Each sprint picks one recurring task, such as drafting chargeback responses, and pairs analysts to try the AI-assisted version on actual cases. At the sprint review, they keep what held up, record what failed and agree on one update to the standard procedure.

The loop holds up best when it has five fixed parts.

  • A named owner who schedules each cycle and protects time for it
  • One recurring task per cycle, chosen from live work
  • A shared review standard that defines good AI-assisted output
  • A short log of what worked, what failed and why
  • A trigger that restarts the loop whenever the tool or policy shifts

Small cycles keep risk visible. Compliance sees each adjustment before it spreads, and teams don't have to wait for a full program relaunch to use new capabilities.

‍

Continuous practice loopTeams try AI on a live task, review the output against their own standards, share what worked and adjust the workflow, then repeat on a set cadence.Try AI on a live taskReview against your standardsShare what workedAdjust the workflowTwo-week cycleThe cycle repeats on a set cadence, and each new tool release feeds the next round.

‍

Why judgment about AI output belongs in the change plan

Judgment belongs in the change plan because the hardest AI skill is knowing when to trust the output and when to challenge it. Classic models rarely name that skill, so teams either accept AI output too readily or reject it by reflex. Both patterns quietly erase the value of the investment.

Electric Mind builds that judgment through hands-on practice on each team's own work, where people compare AI drafts against their expertise and learn the failure patterns specific to their domain. A credit analyst learns that the tool misreads covenant terms in older agreements, while a client service lead learns it handles routine account queries well. That knowledge can't come from a generic course, and it's what keeps adoption safe inside existing governance and risk frameworks.

Treating change as ongoing work is the real discipline

ADKAR and Kotter remain worth using. They give you shared language, a sequence for leaders and a way to spot where people are stuck. What they can't do on their own is keep pace with tools that shift every quarter, so the discipline lies in pairing them with a loop that never formally closes.

‍

"Judgment belongs in the change plan because the hardest AI skill is knowing when to trust the output and when to challenge it."

‍

That takes patience more than inspiration. The organizations getting value from AI aren't always the ones with the boldest launch. They're the ones that kept practising after launch, reviewed output honestly and reset standards each time the tools moved. We see the same pattern in our work at Electric Mind, where we extend the familiar models with practice on live work and steady coaching on when to trust AI output and when to question it.

Pick one team, one recurring task and one review standard this quarter. Run the loop, keep score with the share of work that actually moves to the new path, and let the models guide the conversation as the tools keep shifting.

Got a complex challenge? Let's solve it together, and for real.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Top