Helping Customers Change How They Work With AI

By Brian Tr · Updated Oct 7, 2026

An AI product only sticks when the people who buy it change how they work. Signing the contract is the easy part. If the support team still answers tickets the old way, or the analysts still rebuild the report by hand "just to be safe," the product is installed but unused, and unused software gets cancelled at renewal.

This page explains that problem through five well-known, published models of change and adoption: Prosci's ADKAR, Kotter's eight steps, Rogers' diffusion of innovations, Davis's Technology Acceptance Model, and Moore's chasm. None of them was written about AI. All of them describe why people resist, delay or adopt a new way of working, which is exactly what an AI vendor is up against. For the metrics that show whether adoption is happening, see AI product adoption metrics. For what happens when it isn't, see why AI products churn.

Why Change Is the Real Bottleneck

Software that automates a task nobody watched is rare. Most AI products touch work that people own, are measured on, and have strong habits around. Switching on the tool doesn't switch the habits.

The evidence that the gap sits with organizations rather than models is consistent. The 2025 MIT NANDA report "The GenAI Divide," as described by Fortune, concluded that enterprise AI struggles mainly because of flawed integration and a "learning gap," not model quality. Fortune reports the research drew on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public AI deployments, and that it found externally purchased tools succeeded about twice as often as internal builds (roughly 67 percent versus 33 percent). The headline that 95 percent of pilots showed little measurable P&L impact is widely quoted. The sample is modest and the bar is harsh, so use it as a direction, not a precise rate. We reached the findings through press coverage, not the report itself.

Employees also adopt faster than companies do. Microsoft's 2024 Work Trend Index surveyed 31,000 knowledge workers across 31 markets in February and March 2024 and found 75 percent already used generative AI at work. Of those users, 78 percent brought their own AI tools rather than using company-provided ones. That's individual adoption running ahead of organizational adoption, which is a different problem from the one a vendor selling into a company has to solve.

Together these point at the same thing. People will try AI on their own. Getting a team to change a shared process around your product is a separate and harder job.

Five Models, Five Lessons for a Vendor

Each model answers a slightly different question. Together they make a usable checklist.

ADKAR: Change One Person at a Time

Prosci's ADKAR model describes five outcomes a person needs to reach for a change to take hold: Awareness (understanding why the change is needed), Desire (choosing to participate and support it), Knowledge (knowing how to change and what success requires), Ability (putting new skills and behaviors into practice) and Reinforcement (sustaining the new way and preventing backsliding).

The useful property of ADKAR is that the elements are ordered and each one can fail independently. A user who knows how to use your AI assistant but doesn't see why they should (low Desire) won't use it. A user who wants to use it but has never been shown how it fits their real tasks (low Knowledge or Ability) will try it twice and stop.

For a vendor, ADKAR is a diagnostic for a stalled account. Ask where the barrier sits. Awareness gaps need a clear business case, which is the work in building an AI ROI business case. Desire gaps usually mean the user fears the tool replaces them, or just that it adds steps. Knowledge and Ability gaps are training and in-workflow guidance. Reinforcement is where most AI rollouts quietly fail: nobody checks usage after launch, and old habits return.

Kotter: Change the Organization, Not Only the User

ADKAR works at the level of the individual. Kotter's eight steps work at the level of the organization: create a sense of urgency, build a guiding coalition, form a strategic vision, enlist a volunteer army, enable action by removing barriers, generate short-term wins, sustain acceleration, and institute change. Kotter set the sequence out in the book Leading Change, and the model has been revised since.

Three of those steps map directly onto vendor behavior.

Guiding coalition. An AI rollout driven only by the person who bought it is fragile. A coalition means a business owner, someone from IT or security, and respected users from the team that will change. The buying committee that approved the purchase is the natural starting point, but it often dissolves after the contract is signed. Reconvene it.

Short-term wins. AI value can take months to show in a P&L. Waiting that long invites doubt. A vendor can help the customer pick one workflow, one team and one measurable result that can be demonstrated in weeks, and publicize it internally.

Removing barriers. The barriers are usually practical: a permissions setup that blocks the data the AI needs, a compliance review nobody scheduled, a manager who still requires the old report format. A vendor who treats these as the customer's problem will watch usage stay flat. A vendor who helps clear them tends to see adoption move.

Rogers: Adoption Spreads Through People, Not Announcements

Everett Rogers' Diffusion of Innovations, first published in 1962, synthesized hundreds of earlier studies. It sorts adopters into innovators, early adopters, early majority, late majority and laggards, and it names five attributes that shape how quickly an innovation spreads: relative advantage, compatibility, complexity, trialability and observability.

Read as a checklist for an AI product, the attributes are practical.

• Relative advantage: is it clearly better than the current way, for this task?
• Compatibility: does it fit existing tools, values and habits, or does it demand a new way of working?
• Complexity: how hard is it to understand and use?
• Trialability: can a person try it on real work without a commitment?
• Observability: can colleagues see someone else getting a result?

Observability is the one vendors underuse. People copy what they see working. A shared dashboard of time saved, a team channel where good prompts are posted, or a visible before-and-after on a single process does more than another training webinar.

Davis: Usefulness Comes Before Ease

Fred Davis's Technology Acceptance Model, introduced in a 1989 paper in MIS Quarterly (volume 13, issue 3, pages 319 to 340), proposes that two beliefs drive whether a person adopts a technology: perceived usefulness, meaning the degree to which they believe it will enhance their job performance, and perceived ease of use, meaning the degree to which they believe it will be free of effort. (Davis, 1989)

Two points matter for AI. First, usefulness is a belief about a person's own job, not a feature list. A user who doesn't see how the tool helps with the tasks they're measured on won't adopt it however clean the interface is. Second, AI introduces a third concern the original model doesn't centre: whether the output can be trusted. Later research extended TAM in many directions, but the vendor takeaway is simple. If users don't trust the answers, perceived usefulness falls, because they have to check everything. That's why trust and evaluation is an adoption issue, not only a security one.

Moore: The Chasm Between Enthusiasts and the Mainstream

Geoffrey Moore's Crossing the Chasm, first published in 1991, argues that there's a gap between early adopters (visionaries who buy on potential) and the pragmatic early majority (who buy on proven results and references). The book recommends focusing on one segment, building a complete solution for its specific problem, and using that segment as a base for the next.

AI vendors hit a version of this inside a single customer. A pilot team of enthusiasts adopts happily. The wider department, which is pragmatic and busy, wants proof that it works on their kind of case, a named colleague who vouches for it, and support when it fails. Treat each team as its own small chasm. Don't assume the pilot group's enthusiasm transfers. The lesson from the pilot-to-contract problem applies after the sale too.

What the Models Add Up To: A Vendor Playbook

Strip the five models to their overlaps and a short list of practices remains.

1. Identify champions early, and equip them. Rogers' opinion leaders and Kotter's coalition are the same people in different words. A champion needs materials, access to your team, and something to show their peers.
2. Redesign the workflow, not just the task. Adding an AI step to an unchanged process often adds work. Agree with the customer which steps disappear, which move, and who owns the new version. Compatibility (Rogers) and ease of use (Davis) both depend on it.
3. Train inside the work. Knowledge and Ability in ADKAR come from practice on real tasks, not slide decks. Short, role-specific examples beat general tours. For the general SaaS version, see onboarding and time to value.
4. Build trust deliberately. Show where the output comes from, where it's likely to be wrong, and how a person can correct it. Show it early, before the first bad answer, not after.
5. Measure behaviour, not logins. Reinforcement means checking whether the new way is actually being used on the real process. A login count is weak evidence. The share of the target workflow completed through the product is better.
6. Stay after launch. Reinforcement is the step most often skipped. Proactive check-ins are how a vendor catches backsliding before the renewal conversation. For how this works specifically for AI products, see customer success for AI products.

Key Facts

• Where the gap sits: the 2025 MIT NANDA study, as reported by Fortune, attributes stalled enterprise AI mainly to integration and a learning gap, not model quality. The widely quoted 95 percent figure rests on 150 interviews, a 350-employee survey and 300 public deployments, so treat it as directional.
• Individuals move first: the 2024 Work Trend Index (31,000 knowledge workers, 31 markets) found 75 percent used generative AI at work and 78 percent of those users brought their own tools.
• ADKAR: Prosci's five elements are Awareness, Desire, Knowledge, Ability and Reinforcement.
• Kotter: eight steps, from creating urgency to instituting change.
• TAM: Davis (1989) names perceived usefulness and perceived ease of use as the two drivers of acceptance.
• Practical test: if a user can't say what the tool does for their own job, or doesn't trust its output, adoption will stall regardless of interface quality.