When staff stop using AI after rollout: a playbook for managers

A year ago, 29% of the professionals surveyed by the Thomson Reuters Institute expected generative AI to become central to their work within twelve months. In the institute's 2026 AI in Professional Services Report, the share who say it is central today rose only three points, from 13% to 16%. If your firm rolled out AI in the last year, you probably see the same gap on your own team: the licenses are approved and the training has run, but several people tried the tool on a recurring task and then went back to doing it the old way.

The usual response is another training session, and it rarely brings the expected change in behavior. When people stop using AI for a task, the cause is more often where the work is received than how the prompt was written: the draft needed so much checking, or the person who received it rewrote it so often, that the whole task took longer than before. So the goal of this playbook is narrow: get one task back into routine use, or decide on purpose that AI does not help with it.

The five steps below trace one task, test it with three questions, match the fix to one of four causes, change how the work is received and retire the old version. Two short hypothetical cases run through the steps, chosen because they end with different fixes:

• The weekly status note. An associate drafted the internal status note for an engagement with the firm's approved AI assistant for three weeks, then went back to writing it by hand.
• The client document summary. An analyst used the same assistant to summarize long client documents, such as a contract or an annual report, before partner meetings, and stopped after the partner said the summaries were too general to prepare from.

Why another training session is usually the wrong first fix

At most firms that have rolled out AI, access and basic training are no longer what is missing. The people who use the tools use them often, and the gap shows up on particular tasks. The Thomson Reuters Institute surveyed 1,514 professionals in legal, tax and accounting, corporate risk and government roles in 27 countries in October and November 2025. The share whose organizations use generative AI rose from 22% to 40% in a year, and 82% of current users use it at least weekly. BCG's AI at Work 2025, a survey of more than 10,600 leaders, managers and frontline employees in 11 countries and regions, found that regular use among frontline employees "has stalled at 51%."

The standard advice describes the conditions that make use likely. Gallup's April 2026 study, a survey of 23,717 employed US adults in February 2026, points to AI built into existing workflows, manager support with clear expectations and clear AI policies. BCG adds at least five hours of training with in-person coaching, and Microsoft's Copilot feedback guide for Champions relies on champions, the staff who promote a tool to colleagues, to gather feedback. That advice is sound, and a firm without it should start there. But in both of our cases those conditions were in place, so the manager has to look at the task itself.

Step 1: Pick one task and follow one recent instance of it

Choose one recurring task that someone tried with AI and then went back to doing by hand, and follow a single recent instance from the first draft to the person who received it. Write down what that person changed and how long the whole task took, including checking. Each task stops for its own reason, and a usage report cannot show what happened to the draft.

In the status note, the AI draft misstated two milestone dates because it worked from last week's notes rather than the current project tracker. The manager corrected the dates without saying anything, so the associate, unsure whether the next draft would pass, kept writing a version by hand as a backup. In the summary, the facts were right, but the partner found that each summary described the document in general terms and asked the analyst to go back to the usual bullet points.

Step 2: Run the acceptance test with both people

The acceptance test is three questions a manager asks about one task that someone stopped using AI for: was the output usable, did the person receiving it accept it, and did the whole task take less work than before? Ask them of both the person who did the work and the person who received it, because each sees a different part of the problem.

The third question matters because time saved in drafting can disappear in checking. A Workday survey of 3,200 full-time employees who actively use AI at organizations with more than $100 million in annual revenue, fielded in November 2025, found that nearly 40% of the time AI saves is lost to rework, and only 14% consistently get clear, positive net outcomes. Workday sells software and reported this in a press release, so treat the figures as directional. The first question matters as well, because reliability and accuracy concerns are the top reason respondents in the Thomson Reuters Institute report give for why AI should not be applied to their work.

The two cases fail at different questions, and that difference decides the fix:

• Status note. The output was partly usable, because the format and tone were right but the dates were wrong. It was not accepted, because the manager rewrote it without comment. The task took more work, because the associate prepared two versions each week.
• Client document summary. The output was not usable for its purpose, because it did not tell the partner what to raise in the meeting. It was not accepted, and the task took more work, because the analyst wrote the usual bullet points as well.

Step 3: Match the fix to one of four causes

Each "no" points to one of four causes, and only one of them is fixed by training. In Gallup's study, 88% of employees who strongly agree that AI integrates well with the systems and processes they use are frequent users, against 55% of those who do not, which points to workflow and sources as much as skill. Among employees who do not use AI at all, 39% do not believe it can help with their work, which is sometimes the right conclusion for a task. Skill still matters, because BCG found that only one-third of employees say they have been properly trained.

• Coaching gap. The person does not yet know how to give the tool what the task needs. This was the summary case: the analyst asked for a summary without saying who would read it or what the partner needed, such as the client's obligations, the deadlines and the open questions. A short session on this one task, using two past summaries the partner liked as examples, is the whole fix.
• Workflow gap. The person receiving the work still asks for the old format, or a later review expects a structure the AI draft does not follow. Changing the request or the review step fixes this; training does not.
• Source or access gap. The tool cannot see the information the task depends on. This was the status-note case: the assistant could not see the current project tracker. Connecting the tracker, or pasting in the current dates, fixes it, and a prompting class would not. A good draft still depends on a current record, which is why we treat the source as part of the system in build, buy, or fix the process first.
• A task AI does not help with. Part of the work is a judgment the person has to make, such as the assessment of client risk in the status note. If AI adds checking there without saving work, stop using it for that part and write down why, so the team does not keep trying.

Step 4: Change how the person receiving the work accepts it

Employees go back to the old way when the person who receives the work rewrites it silently or still asks for the old format. So the fix often sits with the manager or partner: agree on the standard in advance, mark errors on the AI-assisted draft instead of fixing them, and stop asking for the old version.

This matters because a weak AI draft changes how the reviewer sees the person who sent it. BetterUp Labs and the Stanford Social Media Lab surveyed US full-time desk workers in September 2025 about "workslop," AI-generated work that looks polished but is unhelpful or off the mark. Forty percent had received it in the past month, each instance took an average of 1 hour 51 minutes to deal with, and managers received it more often than individual contributors, at 54% against 38.5%. About half of recipients saw the sender as less creative, capable and reliable, which gives an associate good reason to keep a backup. Manager support also goes with higher use: in Gallup's study, among employees whose organizations make AI available, 78% of those who strongly agree that their manager actively supports AI use are frequent users, against 44% of those who do not, although that is an association, not proof of cause.

For the status note, the manager and associate agree that a passing note contains current dates from the project tracker, the open risks and the decisions the partner needs to make. For the summary, the partner says once what a useful summary contains and reviews the next two against that list.

Step 5: Retire the old version once the new one meets the standard

The work stops being done twice only when someone decides the new version is good enough and removes the old one. Do this only after the AI-assisted version has met the agreed standard for a few cycles, so nobody loses their backup while the new version still fails. Firms have been slow to make this change: in Workday's survey, 89% of organizations had updated fewer than half of their roles to reflect what AI can do.

For the status note, the hand-written backup stops after three notes in a row pass, and the old template comes out of the shared drive. Keeping that standard after the tools or the team change is a separate job. In our case study A shared system for repeat client analysis, we built a consultancy a library of reusable task instructions, each explaining how a task is done and how its output should be checked, along with saved examples that reviewers rerun after a change to see whether the results differ. The same approach works for a status note: write down the standard once and check new versions against it.

Diagnostic questions for your team

Ask these about any task your team has gone back to doing the old way:

• Which recurring tasks did someone try with AI in the last three months and then stop?
• For one of them, what did the person receiving it change in the last AI-assisted version?
• Is anyone still preparing the old version as a backup?
• Can the tool see the documents and systems the task depends on?
• Has the person receiving the work said what a passing version contains?
• Is any part of the task a judgment the person has to make themselves?

The first step to take this week

Pick one task from the first question above and book 20 minutes with the person who did it and the person who received it. Bring one recent instance, ask the three acceptance-test questions, and write down which answer was "no." That answer tells you whether the next step is coaching, a change to the request or review, a missing source, or a decision to stop.

Questions buyers ask

Should we run more AI training for staff who stopped using the tools?

Only when the acceptance test points to a coaching gap. Training helps when the person does not yet know how to give the tool what the task needs, and BCG found that regular use is sharply higher among employees who received at least five hours of training and had access to in-person coaching. It does not fix a missing source, a review step that still expects the old format, or a task that AI does not make shorter. If you do run training, build it around one task and your team's own past examples rather than a general session on prompting.

Are usage dashboards enough to measure adoption?

No, because a dashboard shows who opened the tool, not whether the work it produced was accepted. A team can log in every day and still prepare the old version as a backup. Only 18% of respondents in the Thomson Reuters Institute report say their organizations collect metrics on the return from AI. For each task, track two things instead: whether the AI-assisted version was accepted without a rewrite, and whether anyone still prepares the old version.

Should we make AI use mandatory for the tasks people stopped using it for?

We recommend against it until the task passes the acceptance test. A requirement to use the tool does not change why the output was rejected, so people comply by producing an AI draft and then a backup, which makes the task longer. In Gallup's study, 46% of employees who do not use AI prefer to keep doing the work the way it is done now, and a rule that adds work will not change their minds. Set the expectation on the result instead: agree what a passing version contains, and once the AI-assisted version has met it for a few cycles, remove the old template.

How long should we keep trying before deciding AI does not help with a task?

Judge it over three or four cycles after you have fixed the cause, because the weeks before the fix tell you little. A task with a missing source or a silent rewrite will keep failing until that is fixed, however long you wait. Once coaching and sources are in place, stop if the output still fails the recipient's standard after those cycles, or if checking it takes as long as doing the work. Record the decision and the reason, so the team can revisit the task when the tools change. Models continue to improve at a rapid pace, so just because a task fails now does not mean it will always do so.

Adoption after rollout comes back one task at a time, at the point where the work is received, so start with one task and the person who receives it. If you want help deciding which tasks to fix, change or stop across your firm, see our AI strategy and opportunity design service.