Every change program I have seen copies a shape that Kurt Lewin sketched decades ago: unfreeze, change, refreeze. You thaw the old habits, move people to the new way of working, then set the new way solid so it holds. Kotter's version ends the same way, with the new approach anchored in the culture. The consultants who sell AI change management mostly sell this shape too. They bring a readiness assessment, a champion network, a training calendar and a go-live date, and after that comes the reinforcement phase, the slide that shows a green tick and a smiling stick figure.
I think the refreeze step is now where most AI programs fail. The new way of working stops being new somewhere around the third model update. A company that sets its AI workflows solid in October has set them around tools that will be replaced before the next budget cycle.
What the October numbers say
On October 19, 2026, Accenture published the latest round of its Pulse of Change research. It covers 3,000 C-suite leaders and 3,000 employees at companies with more than $300 million in revenue, surveyed in July and August. Most of the coverage focused on the headline numbers on investment and bubble anxiety. The figures I care about sit lower down.
Daily AI use among leaders rose to 41%, from 27% in July. Among employees it rose to 43%, from 30%. That is a big jump in a short window. Over the same period, only 15% of leaders say they have redesigned end-to-end processes, and 20% are actively deploying agents across functions. A quarter, 25%, report widespread and sustained value from AI.
Then comes the line that should worry anyone running a change program: 48% of employees say rapidly changing AI tools make it hard to keep up.
Put that beside the usage numbers. Nearly half the workforce now uses AI daily, and nearly half of employees say the ground keeps moving under them. Those two facts describe the same people. They adopted, and the thing they adopted changed shape on them.
Deloitte reached the same place from a different angle. Its July 9, 2026 essay, AI adoption to adaptation, by Sue Cantrell and colleagues, reports that "84% of organizations have not redesigned jobs or workflows around AI." The authors argue that login counts and active-user numbers tell you whether people opened a tool. They say nothing about whether anyone changed how they work. The fix they propose is a "continuous adaptation loop."
The word "continuous" is doing the real work in that phrase, and most executives read past it.
Why AI change management breaks the old model
Classic change management assumes a destination. You know the target state before you start: the new ERP, the new org design, the new sales process. The gap between now and then is fixed, and the job is to carry people across it. Resistance is friction on a known path. Training is a map of a known place.
AI removes the known place. The tools your people use in October behave differently from the tools they will use in January. This happens partly because vendors ship constantly and partly because what a model can do reliably keeps widening. A workflow that needed a human checkpoint at step four last spring can skip it now. A task that was too long for an agent to finish in July may now run end to end. If your change program wrote down the target state in a requirements document, the document is already a historical record.
This is why the old success measures mislead. An adoption dashboard that shows 90% weekly active users reads as a finish line. In practice it may only show that people found the tool and then settled into using it the way it worked on the day they were trained. Deloitte's adopter group is exactly that: they open the tool, run it the way they were trained to, and their workflow never changes.
Refreezing is the risk
The refreeze step exists so that people do not slide back into old habits. With AI, sliding back into the old pre-AI habit is a small risk. The large one is staying stuck at the first AI habit you learned.
Say a finance team learned in March to paste a variance report into a chat window and ask for commentary. That was a sensible habit in March. Suppose that by October an agent with access to the general ledger can draft the variance report itself, flag the three line items that moved most, and route each one to its owner for a note. The team that refroze around March's habit is still pasting. Its adoption metrics are perfect. Its process is two generations behind, and nobody's dashboard says so, because the dashboard measures use and use is high.
I call this a frozen adoption. It looks like success, and it is the most common thing I see inside companies that ran a good AI rollout eighteen months ago.
The champion network got routed around
The second thing the October data shows is quieter, and it changes how change actually spreads through a company.
Accenture found that 64% of employees try AI tools before asking a colleague for help. Think about what that does to the standard playbook. Change programs run on social proof. You recruit champions in each department, the early adopters show their neighbours, and the new habit spreads through the informal network of who-asks-whom. The whole design depends on people turning to a colleague when they get stuck.
When a majority turns to the model first, the model becomes the first-line change agent. It answers the how-do-I question. It also shapes the answer, because whatever the model suggests becomes the default workflow for the person who asked. Your champion network is still there, but it now sees the second question and misses the first.
That has a practical consequence. If you want to steer how people work, the place to steer is the tool's configuration: its system instructions, its connected data, the skills and templates it offers by default. Every company that deploys an enterprise assistant is already writing a change program into those settings, whether or not anyone in HR knows it.
Uber built the default into the tool
Uber's rollout this month is a useful example. In October 2026, Uber announced it was taking OpenAI's ChatGPT Work to its global workforce of more than 30,000 employees after a 1,000-person pilot, according to RMN Digital. The rollout targets operations, finance, marketing and product teams, with engineers staying on Codex for code. The adoption support listed in the coverage is the interesting part. Uber is relying on an internal AI skills library, vetted platform integrations and structured training modules.
The skills library is the part I would watch. A skills library is a living set of defaults. Each time a team works out a better way to compile a meeting brief or reconcile a dataset, that method can go into the library, and every employee who asks the tool for help gets the improved version. In that setup, change management happens through the contents of the library, and the library can be edited on any day of the year.
Uber's CTO, Praveen Neppalli Naga, framed the company in the announcement as "a company of builders." The phrase is corporate, but the design behind it is sound. A company of builders can treat its workflows as editable. A company of adopters treats them as installed.
Managers are the bottleneck
If the tool is the first-line change agent, the manager is the second, and the data says the second line is thin.
Gallup's research on AI and workplace productivity, last updated September 30, 2026, found that only 36% of employees in AI-integrating organizations "strongly agree" that their manager actively supports their team's use of AI. The gap that support creates is large. Among employees who strongly agree their manager backs AI use, 78% use it frequently. Among those who do not, 44% do. Employees with supportive managers are 9.3 times as likely to strongly agree that AI has transformed how work gets done.
Gallup's frequent-use figures by level show where the problem sits. As of the second quarter of 2026, 51% of leaders use AI frequently, against 36% of managers and 26% of individual contributors. The executives who sponsor the change program are using the tools. The layer whose job is to change the team's daily routine is behind them.
I do not read this as a training failure. Managers are being asked to redesign their team's work around a tool that keeps changing, while carrying the same targets they had before. Deloitte's survey found that 69% of workers report a heavier workload and 43% cite lack of time as their biggest barrier to adapting. Managers sit inside those numbers too. A manager with no slack in the week will freeze the team's AI habits at whatever level first cleared the quarterly target, and that is a rational choice given what the job rewards.
The program needs a budget line
So what replaces the project? My view is that AI change management has to move out of the transformation office and into operations, with a standing budget, owners and a cadence. Here is what that looks like when I build it with clients.
A release desk for workers
Software teams read release notes. Business teams almost never get the equivalent for the tools they use all day. When a vendor ships a meaningful capability change, somebody inside the company should translate it into what it means for specific workflows: what this team can stop doing, what that team can now hand to an agent, which manual checkpoint is now redundant and which one has become more important.
Accenture's survey gives a reason to include cost in those notes as well. It found that 80% of employees are familiar with AI's cost impact, but only 32% have received clear guidance. Workers already sense that a sprawling agent run costs money. Nobody has told them how much, or what is worth it. A release desk that covers capability and cost in the same note closes both gaps.
Process owners with a rewrite cycle
Every core workflow that AI touches needs a named owner whose job includes rewriting it on a schedule. For illustration, a quarterly cycle works for most finance and operations processes: the owner reviews what the current tools can do, tests the workflow against it, and either confirms it or publishes a new version. The rewrite is the deliverable. A process that has not been revisited in two cycles should show up on a report the way an unpatched server does.
This matters because the 15% figure from Accenture, the share of leaders who have redesigned end-to-end processes, describes a one-off event. Redesign done once becomes frozen adoption at the process level. Redesign done every quarter keeps pace with the tools.
Measure movement instead of usage
Drop active users as the headline metric. It was useful for about six months, while the question was whether anyone would touch the tools. Accenture's numbers say that question is answered.
What to measure instead is how far workflows have moved. Suppose you track, for each core process, the share of steps that ran without human handling this quarter compared with last quarter, and the share of AI output that a human changed before it shipped. The first number tells you whether the work is moving. The second tells you whether review is real or ceremonial. Accenture found 50% of employees say AI has raised the expected speed of work and the need to review AI outputs. That review load is a cost you are already paying, and you can only manage it if you count it.
Deloitte offers a sharper behavioural measure. In its survey, 78% of employees say they are continually experimenting, but only 56% say they are recognized or rewarded for being adaptive. That 22-point gap is a direct read on whether your incentives back the change you say you want. If people experiment and the system ignores them, they will stop.
Pay managers for redesign
Gallup's data makes the manager the strongest lever available, so put money on it. Managers should be evaluated on the movement of their team's workflows as well as output. Give them time to do it. Imagine a manager who gets back four hours a week, taken off reporting duties that an agent now drafts, with the explicit expectation that those hours go into redesigning how the team works. That trade costs very little and turns the 64% who reach for the model first into a team whose first reach is a workflow its own manager designed.
Write the defaults on purpose
Because the model answers first, treat the configuration of your enterprise assistant as a change management asset with an owner. Decide what the default instructions say about your company's methods. Decide which skills and templates appear first. Decide which data sources are connected, because an assistant with access to the real system of record will suggest real workflows, and one without that access will suggest workarounds. Uber's skills library is one version of this. The exact form matters less than who owns it and how often they change it.
What a frozen company looks like in a year
It helps to picture the two paths with something concrete.
Say two companies of the same size, in the same industry, both reached 80% daily AI use by the end of 2026. The first ran a textbook program: assessment, champions, training, go-live, reinforcement. It declared success in December and moved the transformation team to the next priority. The second put a standing team on the release desk, gave every core process an owner with a quarterly rewrite, and changed its manager scorecards.
By late 2027 their usage dashboards look the same. Under the dashboards, the first company is running workflows designed around the tools of early 2026, and every capability gain since then has reached it only as a slightly better answer in the same chat window. The second company has been through four rewrite cycles. Its people are not smarter, and its vendor contracts are the same. It just never declared the change finished.
The gap between them will not show in any adoption metric, because both adopted. It will show in cost per transaction, cycle time and headcount growth, about two years later, when it is expensive to close.
The resistance you are really facing
Most change literature treats resistance as fear: fear of job loss, fear of incompetence, fear of the unfamiliar. Some of that is real. Accenture's data complicates the picture, though. In the same survey, 82% of employees say AI increased their productivity, 75% say it increased their job satisfaction, and 66% now feel secure in their roles, up from 48%. On those numbers, the workforce is not afraid of AI.
The resistance I see more often is exhaustion. It is the 48% who cannot keep up. People are willing to change once. They are less willing to relearn the same task every quarter with no recognition, no time set aside and no clear reason why last quarter's method was retired. A permanent change function exists mostly to make that relearning cheap: to tell people what changed and why, to hand them the new default rather than make them find it, and to reward the ones who find it first.
Treat the exhaustion as an operations problem. Fatigue goes down when change arrives on a known rhythm, from a known source, in small pieces. It goes up when change arrives as a surprise, from a vendor email, in a lump. You control the rhythm. The vendors control everything else.
Why this has to be designed
Buying another tool will not fix any of this. Uber's rollout works because of the scaffolding around the tool: the pilot, the skills library and the training modules, which is to say the operating design that decides how a better method spreads through 30,000 people. That design has to fit your org chart, your incentive plan, your data access, your cost controls and the pace at which your vendors ship. None of it comes in the box.
The companies that win the next two years will be the ones that stopped treating AI change management as a project with a finish date and started running it as a standing function with owners, a cadence and a budget. The ones that lose will have excellent adoption numbers and processes set in place around tools that no longer exist. Somebody has to design the rewrite cycle, the release desk, the manager scorecard and the defaults inside the assistant, and has to do it before the tools you trained everyone on are retired. Agor AI Advisory builds those functions with leadership teams, starting from your workflows instead of a vendor's demo, and we stay through the first rewrite cycles so the rhythm holds once we leave. Waiting has a cost, because every quarter you spend frozen is a quarter of capability your competitors are building into their workflows and you are not.