The heaviest AI users on your team feel like they have cracked a strategy. Most have actually stopped doing the work where the strategy used to form.

A meeting I had last week

I sat with the head of analytics at a regional insurer. He was proud of his team's new pace. Reports that used to take three days now take four hours. He showed me the dashboards. He showed me the prompt library. He showed me the new junior hire who, in his words, "ramped in a week."

Then I asked the junior to walk me through one of his calculations without the tools open.

He could not.

What he had was a vocabulary for the work, not a model of it. He could narrate the output well enough to make it through a meeting. Push one layer deeper and the reasoning was not there. The head of analytics watched this in real time and got very quiet.

The unsettling part: by every dashboard the company watches, this analyst is performing well. Throughput is up. Turnaround is down. Manager satisfaction is high. The hidden cost only shows up the moment the AI is gone.

A new study put numbers on what I saw

Six researchers at UC Irvine ran a three-phase experiment on 132 adults solving logic puzzles. They tested skill without AI. Then they gave half the room AI hints for twenty minutes. Then they tested skill again without AI.

Heavy AI users were measurably slower and less accurate in the post-AI test. 115 seconds per problem versus 78 for light users, with lower correctness. The ability gap between high- and low-skill participants widened with AI use rather than narrowing.

And the result that should worry every operations leader:

Lower-ability participants who had access to the more helpful AI reported a stronger belief they had "found a strategy" than the higher-ability participants did, despite actually performing worse.

The most confident people in the room were the people who had learned the least.

This is the productivity story nobody is writing yet. Your throughput dashboard does not see it. Your engagement survey does not catch it. The only place it shows up is in the moment you take the assistant away. A client meeting. A deposition. A crisis call. The person who looked like an expert cannot reason their way through a basic question.

Pulling the AI back is the wrong instinct. The better move: decide, on purpose, which parts of the workflow your people should keep doing alone, and protect that work from the assistant.

The cheapest output AI gives your team is confidence. Skill costs more, and you can only pay for it in the time your people spend without the assistant.