Why a new framework.
ADKAR (Prosci) and Kotter's 8 steps are the two best-known change frameworks. Both are good. Neither was built for a technology that updates quarterly, spreads through unsanctioned AI, and produces value that is hard to attribute. Kotter's original eight-step model was published in Harvard Business Review's 1995 article "Leading Change", almost three decades before generative AI made change continuous. We put ADKAR and Kotter side by side if you are choosing between them.
An AI change framework keeps the human-transition strength of the classics and adds three things they leave out: continuous workflow redesign, AI-specific governance, and a measurement model that survives an AI model upgrade. BCG estimates that about 70% of the value from an AI transformation comes from people and process work such as workflow redesign, culture and governance. The model itself accounts for about 10%.
The seven layers.
Layer 01
Diagnose.
- What it delivers
- Read across people, process, and platform. Name the three to five places AI is producing measurable value and the three to five places it is producing measurable cost. Write the read in prose alongside any score.
- Failure mode
- Maturity-model scoring with no narrative. Everyone agrees the score is roughly right. Nothing changes.
Layer 02
Decide.
- What it delivers
- Make the explicit bets. Which workflows, teams, and AI products you are investing in, and the no list. Connect each bet to a business outcome and at least one customer outcome.
- Failure mode
- Every team is funded and there is no no list, so investments do not compound.
Layer 03
Redesign.
- What it delivers
- Rebuild the workflows touched by each bet with AI in the loop. The workflow owner is on the keyboard during the redesign.
- Failure mode
- AI bolted on top of an unchanged process. Time saved disappears into more rounds of the same work.
Layer 04
Enable.
- What it delivers
- Move people through the redesigned workflow in pairs. Fluency is built role by role, inside each person's own work. Leaders model use themselves.
- Failure mode
- Training programs with no link to the work the trainees do on Monday morning.
Layer 05
Govern.
- What it delivers
- Sanctioned AI list. One named owner of the AI stack. Policy on data, prompts, and approvals, written in language a director can apply without calling legal.
- Failure mode
- Either no policy (shadow AI everywhere) or policy so heavy it pushes use back into the shadow stack.
Layer 06
Measure.
- What it delivers
- Per bet: one adoption metric, one capability metric, one outcome metric. Public to the team. Reviewed quarterly.
- Failure mode
- Only active-user counts are tracked, with no tie to a business outcome.
Layer 07
Re-read.
- What it delivers
- Schedule the next diagnosis before finishing the current one. The AI model landscape will have moved, so the framework runs as a repeating loop.
- Failure mode
- Strategy shipped, never re-read. Plan is a year out of date in nine months.
How this maps to ADKAR and Kotter.
Our Enable layer applies the same logic ADKAR (Awareness, Desire, Knowledge, Ability, Reinforcement) is built on: adoption happens one person at a time. Kotter's discipline around sponsorship and momentum shows up in how we run Decide, Enable, and Measure. This framework adds AI-specific layers around the ground the classics cover, and it does not repackage them.
Related reading.
- - AI change management frameworks, compared: ADKAR, Kotter, Lewin, and 7-S side by side.
- - AI change management software, compared: the adoption platforms, what each is for, and the order to buy them in.
- - AI change management software, compared: what to buy and when.
- - AI readiness assessment: the diagnosis layer in detail.
- - Generative AI adoption: the redesign and enable layers in detail.
- - Our book, The Elephant in the Algorithm: the longer argument behind the framework.
Questions people ask.
What is an AI change management framework?
A repeatable structure for moving an organization from one way of working to a new, AI-included way of working, while keeping the trust of the people doing the work. It covers the human transition (awareness, desire, capability), the workflow redesign, the governance, and the measurement. Classical change frameworks like ADKAR and Kotter cover the first part well; AI-native frameworks add the rest.
Why not just use ADKAR or Kotter?
Use them, but not alone. ADKAR is excellent for the human-side transition. Kotter is excellent for the leadership arc. Neither was designed for the AI-specific failure modes: shadow AI sprawl, prompt-hygiene drift, AI model updates that change the workflow under your feet, and the always-on cadence of AI change. An AI-native framework adds those.
What is the simplest version that still works?
Three things repeatedly: diagnose where AI is producing value and where it is producing cost; redesign the specific workflows touched; bring the people through those redesigns with named owners and measured outcomes. Re-run the cycle quarterly because the AI model landscape will have moved.
Where does software fit?
Software sits below the framework. Adoption platforms, learning platforms, and AI spend trackers all help once you know what you are driving. Buying them before the framework is in place wastes the license. See the <a href='/guides/ai-change-management-software-comparison'>software comparison guide</a> for which to buy when.
ADKAR is a registered trademark of Prosci Inc. Kotter's 8-Step Process, Lewin's change model, and the McKinsey 7-S framework remain the work and intellectual property of their respective creators. Average Robot is not affiliated with, certified by, or endorsed by any of these organizations. References here are for comparison and commentary.