Change Management for an AI Rollout That Sticks

Research from SAP and Oxford Economics published on September 10 reported that 97% of surveyed Canadian businesses adopting agentic AI say they are not fully ready for it. That is vendor-sponsored research and the number is doing rhetorical work, so treat it loosely. The direction matches what a survey of 1,000 Canadian small business owners found last week, where 9% credited AI with meaningful gains. Neither gap is about tooling.

Why this is a different problem

Traditional software change management assumes a forcing function. The old system goes away, the new one is the only route to doing the job, and the management task is getting people through the transition without breaking anything.

AI tools have no cutover. The spreadsheet still opens. The old template still works. Anyone who finds the new thing awkward on Tuesday can go back on Wednesday and no process stops them. Nothing forces the issue, so nothing surfaces the problem either.

There is a second difference. Most software has a finished state: the data is migrated, the workflows are built, the project closes. An AI tool is never finished, because what it is good at changes every few months and the prompts people wrote last year quietly stop being the right ones, which we set out in the case for trimming your prompts.

The eight-week shape

Most rollouts are watched hardest in the first fortnight, when the numbers are flattering and mean nothing, and left alone during weeks three to six when the outcome is actually being determined. Flip that attention and you will catch the reversions while they are still fixable.

Four conditions that predict it sticks

1. One task, named, with a before time. Not a tool rollout. A task rollout. Writing quotes, summarising site visits, drafting follow-up emails. A named task has a finish line and a measurement; a tool has neither, which is the pattern behind the businesses seeing returns in last week's survey.

2. A person who does the work, not the person who bought the tool. Colleagues are persuaded by watching someone in the same role get visibly faster. They are not persuaded by an owner who has never done the task describing how much time it will save. Give that person a few protected hours rather than adding it to a full week, which is the argument in AI champions.

3. A written procedure, not a demo. Demos are forgotten within a week. A one-page procedure that says what to paste in, what to check, and when not to use it is what people return to in week four. The shape of that document is in standard operating procedures an AI can follow.

4. A stated answer on jobs. Somebody is wondering and nobody is asking out loud. Say plainly what this means for headcount, including if the answer is that you do not know yet. Unaddressed, the question becomes quiet non-adoption that looks like technical friction.

Resistance is information

Three different objections get filed under the same heading, and they need different responses.

The output is not good enough. This person may well be right, and they are the most valuable voice in the room. Someone who knows the work well enough to spot a plausible-looking error is exactly who you want reviewing it. Ask for examples and take them seriously.

This replaces my job. Reassurance does not work here and is often dishonest. A straight answer does, even when the honest version is that the role will change and you do not yet know how much.

It is awkward and I am busy. The most common one, and the easiest. Ten minutes at their desk fixes more of this than any amount of training material, because the obstacle is usually one specific step nobody explained.

Treating all three as attitude loses you the first group, which is the group that would have told you the truth about whether the tool works.

Measure the task, not the seats

Licences issued and logins recorded are the metrics vendors give you, and they measure your purchase rather than your change. A team where everyone logged in once shows perfect adoption on that dashboard.

Three numbers, weekly, on one line. How long the target task takes now. How many people completed it that way this week. How many reverted to the old method. The third number is the honest one and nobody collects it.

At week eight, compare against the before time you recorded and make a decision. Keeping a tool nobody uses because cancelling feels like admitting a mistake is the most common outcome, and it costs more than the subscription, because the next proposal is harder to fund. What to do with the hours when it does work is in capturing the hours AI saves.

Frequently Asked Questions