Deal teams are handing AI the work that used to belong to their youngest people. Most senior dealmakers (62%) now call it indefensible to decide on a complex deal without AI, and half already use it regularly when they check a company before buying it. Yet only 22% would let it make the final recommendation on whether to sign. Those two answers fit together only if the person making the final call can still tell when the machine is wrong, and that is the part at risk.
Spotting a bad number takes having built the model
An associate who spends three weeks inside a seller's files learns which numbers move, which claims from management survive a call with customers, and what a wrong assumption looks like three levels down in a model. Nobody teaches this in a classroom. It comes from doing the work. When AI does those three weeks in an afternoon, the output is usually good, and the partner who reads it feels fine. The cost arrives years later, when the people reviewing are ones who never built a model themselves.
The same is true in many other jobs. A finance director who approves AI-drafted forecasts, a recruiter who reviews an AI-ranked shortlist and a school head who signs off an AI-written report are all checking work that a junior used to produce. Checking is a skill that grows out of producing, and AI removes the producing first. Firms treated those junior jobs as a cost, but they were also how the firm trained its next senior people. A firm that hires fewer juniors stops training them unless it pays for that training on purpose.
Aviation answered with practice, not a ban
The regulator did not ask airlines to give up autopilots. It asked them to keep using them and to build hand-flying into normal operations and training, starting with the quiet stretches of a flight when the workload is low. Practice became a scheduled habit instead of something that happened by accident.
A finance team could do the same in a few lines of policy. Before an analyst opens the AI's model of a supplier contract, they write down their own rough answer: the margin they expect, the biggest risk, and what would change their mind. Then they compare it with the AI's version. Where the two differ, someone finds out which one is right. That costs about half a day, a few percent of the three weeks the old way took, and it keeps the analyst's eye in use. The written note also helps later. Firms often revisit a company years after turning it down and cannot remember why they said no.
Hidden defaults are harder to question than a partner
There is a second cost. Every tool carries someone's view of what makes a good comparable, a credible management team or a reason to pass. A junior can learn to argue with a partner's stated preferences, but nobody can argue with a ranking whose reasoning is hidden. A tool that shows what it threw away, and why, teaches the person using it. A tool that returns only a ranked list teaches them to trust it.
Over the next few years most firms will get faster, and nearly all of them will say a person still makes the final decision. The firms that do well will be the ones that kept a few people who can tell when the work in front of them is wrong. Those people will be scarce, and they will be the ones a firm trusts with the yes or the no.