Most of the responsible-AI experts polled by MIT Sloan Management Review and BCG agree that the hard problem is no longer making AI good enough. It is keeping people good enough to tell when it is wrong. That changes what a company has to manage: not only the AI, but the skill of the humans around it.
Pilots met this problem first
Airlines have run on autopilot for decades, and they learned what it costs. When pilots rarely fly by hand, the skill decays, and it decays most when it is needed, which is when the automation fails. In 2009, an Air France jet crashed after its crew, who had little hand-flying practice, did not recover from a stall once the automation stopped working. The US aviation regulator now tells pilots to switch the autopilot off on purpose and fly by hand, and to train for automation failures in simulators.
Notice what that advice assumes. The autopilot is good, and nobody wants to give it up. The practice is a cost paid so that a rare bad moment goes well. AI at work needs the same arrangement, because its bad moments are rare too, and that is exactly why nobody practices for them.
The newest people lose the most
An experienced reviewer at least has judgment to lose. A new hire never builds it. The dull first-year work, such as matching invoices, writing first drafts and reconciling accounts by hand, was how people learned what a wrong one looks like. Give that work to AI and the learning goes with it. New York City applied the same logic to children in September 2026, when it banned generative AI in classrooms from pre-kindergarten through eighth grade for a year. Senior people face a milder version: when they let AI draft every decision, they keep their judgment but stop exercising it, and managers who lean on it can lose the knowledge that real decisions used to build.
A review step that keeps the skill alive
Picture a finance team where AI codes and matches 2,000 supplier invoices a month and a clerk approves them. We have already looked at how approval turns into a click on a decision the AI made first. After a year, the clerk has not seen a wrong match in months, and a new colleague has never seen one.
The team could pull 5% of invoices each month, about 100, and have people code them by hand without seeing the AI's suggestion. Then it compares the two answers. At about three minutes an invoice, that is five hours a month, less than one working day for the whole team.
That small habit does two jobs. It is practice, the same way hand-flying is for a pilot. It is also a measurement: if the clerks' answers match the AI's almost every time, including the times the AI is wrong, they were never really reviewing. A team that learns this can move a person to the work where judgment pays, such as new suppliers, unusual amounts and anything with tax or legal consequences, and let the AI run the rest without a human click that adds nothing.
I expect the companies that do well with AI to treat reviewer skill the way they treat equipment: something they own, that wears out, and that needs a maintenance budget. The five hours are cheap next to the first error that nobody catches.