Before buying more AI seats, why are some workers not using the ones you have?
A new NBER working paper finds that AI use at work is widespread but uneven: it reaches many occupations and tasks, yet fewer than half of workers use it in most of them. The useful question is not only what the tool can do. It is why people doing similar jobs make different calls about using it. Before a small team buys more seats, I would sit with three people in the same role: one who uses it, one who tried and stopped, and one who never started. What took longer than expected? Which customer or record did they not want to put in it? What already works well enough without it? Non-use is not automatically resistance. It can be a bad fit, a missing permission, an unclear expectation, or a tool that leaves extra checking to the person who uses it. If the team cannot name which one, another rollout mostly creates more licenses. What would you ask before expanding?
Comments
Compare the same task across the three people rather than asking whether they like the tool. For two weeks, count time to a usable result, minutes spent checking it, and work that bounces back. If the regular user finishes faster but the others lose the saving to review or missing access, that tells you what another license will not.
Before measuring use, ask what opting out costs. If the regular users get the easy cases and everyone else inherits the messy ones, a usage target turns sensible judgment into compliance. Fix the split before buying more seats.
Before the interview, watch each person try the same five-minute job from the ordinary home screen. The useful moment is often the first hesitation: “Which one do I click?”, “Can it see this customer record?”, “Am I supposed to check every sentence?” A seat can be unused because the first try made somebody feel slow or exposed. That will not show up in a license report.
That first hesitation is often not “I don’t know how.” It is “if this gets weird, will I look careless in front of a customer?” A five-minute trial that ends with a draft someone is embarrassed to send can teach the whole team to leave the seat alone. That is not a training problem.
That is why I would not use a completion rate as the rollout dashboard. Give people one low-stakes internal task, say no customer record goes in, and ask what they would be comfortable sending without a manager beside them. If the answer is nothing, the issue is not that they skipped training. The tool has not earned a place in the job yet.