Companies build AI adoption programs around the executive sponsor who funds them and the enthusiasts who show what the tools can do. The managers who decide how the team actually works barely appear in the plan. A manager runs the weekly review, sets the bar, assigns the work, and answers for the number, and adoption becomes an operating practice inside those routines.
An analyst finds a way to cut a reporting day in half, and a year later it’s still her personal trick. When she changes teams the trick leaves with her. Nothing about her discovery changes the team’s definition of the work or what the reviewer checks. It becomes the team’s way when a manager decides it should be, and rebuilds the routine around it.
Executive mandates stall in the same place. The mandate lands as a townhall and a license count, and then the team goes back to its actual week, the Monday meeting and a deadline that hasn’t moved. If nothing in that week changes, the work doesn’t change either. Nobody pushed back. The mandate just never made it into the routines where the work is managed.
Redefine the work, in writing where it counts. Which tasks the team now starts with a model, which outputs a person still writes from scratch, what a reviewer checks when the draft came from a tool, and who answers when model assisted work goes wrong. That’s just management, applied to a new way of producing the work.
A manager’s most useful people here are the heavy users, and a month of watching output makes them easy to find. Take what they’ve learned and turn the useful parts into team practice, like the prompt that handles the common case and the check that catches the usual failure. Forcing everyone to work identically kills the experimentation that found the trick in the first place, so set the standard on the output and the review, and leave the method loose.
While that’s happening, the quality bar needs a reset, because ‘it looks fine’ stops working when volume jumps. The team needs to know what done means for model assisted work: some outputs read line by line, the rest sampled, and the split written down. Set the bar highest where the output feeds someone else’s decision, because a wrong number there costs more than a clumsy sentence.
Two people in the same role will start producing at different rates early, while the standards are still settling. That’s a plain management problem, and treating the slower person as a performance problem in month two teaches the team to hide how the work actually gets done. Ask instead where the released hours go. When nobody decides what the freed hours are for, the team just works a little slower and the gain is gone. Put them on something deliberate, the backlog nobody staffed or the customer work that kept getting deferred, and the difference is visible in the number.
None of it holds until the old process actually stops. Teams keep the old way running next to the new one because it feels safer, but that doubles the work and proves nothing, because everyone still treats the old output as the real one. The manager is the one who has to say the old report stops, because the manager owns the process. A stop decision needs a date and a fallback. After the date, producing the old artifact counts as rework.
Sometimes the model just adds work. The draft arrives faster and the checking takes longer than the drafting used to, especially where errors are expensive and the reviewer can’t yet trust the failure patterns. Managers notice this first because they see the review queue. Say it plainly and pull the tool back from that task.
Managers can’t do this alone. From the center they need baselines they didn’t have to invent, and cover for stop decisions so each one isn’t argued all over again above them. They also need someone above the team to settle what a system may do without a person, because a team level answer to that won’t survive an audit. And the training they get should be about managing model assisted work, since prompting is the easy part.
Nobody should count training attendance as adoption, either. A team can be fully certified and produce its work exactly as before. Attendance tells you the program ran. To see adoption, look at what the review checks now and which processes have stopped.
What leadership wants and what the heavy users have figured out become the team’s way of working through the routines a manager owns. Companies that want adoption to hold should spend less time on the tools and the townhalls and more on the managers who decide what the work is each week.