Several large employers are slowing down on AI. Meta has cut back its plan to reduce staff and rely on AI agents, since the agents' technical and security errors were taking up more and more employee time. Microsoft and McKinsey are limiting how much their staff use AI tools. At Walmart, store and warehouse workers are dealing with AI that hands out schedules and tasks nobody could keep up with. This looks like a retreat, but I read it as companies finding out where the work went.
A faster step moves the work to the next person
A senior Microsoft executive described receiving a document that was clearly written by AI. He had his own AI summarize it and came away with no new thought about the idea. The writer saved an hour and the reader saved ten minutes, but nothing passed between them. Two machines had traded text while two people learned nothing.
Microsoft's own review of its rollout reached the same place: a tool licensed to over 200,000 people does not by itself change how the work gets done. Speeding up one step makes a longer line at the next. And AI output looks right even when it is wrong, so a mistake travels further down the chain before anyone catches it, usually to a person who never had the time that was saved.
The electricity story had a second half
The factory redesign was not only about layout. Once each worker had a motor to start and stop, workers made more decisions, so they needed more training and different contracts. AI has the same second half. Microsoft found that pairing early-career engineers with experienced colleagues raised productivity, which is a change in how people work together, not a change in software.
What a redesign looks like at a desk
Take a finance team at a distributor that writes 30 monthly variance reports. By hand, each takes about 45 minutes. An AI writes a first draft in 5 minutes, but a reviewer who cannot see where the numbers came from has to rebuild them, and that takes about 25. The total is 30 minutes instead of 45, a saving of roughly 7 hours a month.
Now change the process around the tool. Each number in the draft links to the ledger line it came from, and only unusual variances are flagged for a person to read. Checking falls to about 8 minutes, the total to 13, and the saving to about 16 hours a month. The same AI gives more than double the return, and the only thing that changed was the work around it.
What follows
Companies that cut people first and redesigned later will find that the work did not disappear. It turned into checking, which is why a forecast that about 30% of jobs cut for AI will be rehired by 2029 is believable. The companies that gain will be the ones that ask where an AI's output waits for a person, and fix that point first.
If you are judging an AI tool in your own team, measure the time from the request to a result someone has checked and trusts, not the time to the first draft.