Microsoft offered buyouts to roughly a quarter of its employees last week, or about 60,000 people. This follows layoffs of 2% of its workforce in July, and the company’s first-ever voluntary retirement offer to 7% of employees in April. Microsoft appears to be creating fiscal room for the $170 billion it’s spending on capital projects this year, much of it on data centers for artificial intelligence.

Paying people to leave is the decent way to shrink a company, and fits the gentler culture that Microsoft CEO Satya Nadella has worked to instill. But Microsoft may have created a retention bonus in reverse — because for companies betting on AI, the people they want most to stay are the ones a tenure-based buyout offer pays most to leave.

This has always been a risk of buyouts, although there are ways to structure them to avoid losing crucial employees. When the US military shrank after the Cold War, for example, the Army cut 245,000 people and paid many to leave, with bigger payments for longer service. It limited the offer to particular combinations of job, rank and years of service, and left out anyone already eligible for retirement pay. The eligible pool skewed toward soldiers with lower test scores and slower promotions. For such soldiers, the offer more than doubled the chance of leaving, from about 13% to more than 29%. For higher-quality soldiers, it rose by 10 percentage points.

Microsoft drew its lines differently. The April program was open only to people whose age plus tenure crossed a threshold, so it went to the company’s older, longer-serving employees. More than 30% of the roughly 8,750 eligible people took it. Last week’s offer, Business Insider reported, pays a week of base salary for every six months of service, or two weeks at more senior levels, up to 39 weeks. Microsoft is letting each business decide whether to participate, which gives it a say over which teams shrink.

Within a participating team, though, performance isn’t a factor in who can accept, and Microsoft will try to retain top talent who express interest. The longer someone has stayed, the bigger the check, up to that cap.

Some executives might see that as a feature, on the theory that their longest-serving people are the most attached to how work got done before AI. For a company betting on AI, however, this might be backward.

In a recent paper, two doctoral students in economics at Harvard tracked résumé data of about 65 million workers at more than 280,000 US firms. In the six quarters after ChatGPT’s release, junior employment at companies building generative AI into their operations fell about 9% relative to other firms, while senior employment kept rising on its earlier trend. The drop came mainly from hiring fewer junior employees, not from letting them go. Some economists argue that remote work explains much of the entry-level slump, and the authors describe their findings as an association, not proof of cause. Still, in their data, the companies adopting AI have so far kept their experienced people.

Meta’s bet — it laid off 8,000 people in April in a bid to become more efficient as it spends more on AI — hasn’t paid off on schedule. CEO Mark Zuckerberg called off a planned second round of layoffs hours before the first began, and in July he conceded that executives had misjudged the timing and that AI agents hadn’t progressed as quickly as he expected. Other companies are still waiting, too. In a recent survey of nearly 9,000 executives in seven countries , companies said AI had raised their annual productivity growth by about 0.2% over the previous three years. They expected about 0.6% a year over the next three.

Whichever way the forecast goes, paying experienced people to leave is a risky bet. If AI delivers, someone has to judge what it produces. SignalFire, a venture capital firm that tracks tech hiring, argues that AI has moved the bottleneck in software from writing code to reviewing it, which is why companies still need their experienced engineers. If AI disappoints, those same experienced workers are the ones who know how to run the business as it stands and are best positioned to train new people and do the work AI was supposed to take over.

And the biggest tech companies might not have those new people if they need them. SignalFire estimates that new-graduate hiring at 12 of the largest, Microsoft and Meta among them, is running about 65% below its 2019 level, and warns that the industry could face a leadership vacuum over the next decade. Firms that hire fewer beginners while also paying veterans to leave are weakening both ends of their talent chain.

Microsoft’s announcement was candid: “The Voluntary Exit Offer is new for us. We’ll learn from it.”No company should regret being humane, especially in this age of AI. But — again, especially in this age of AI — paying your longest-serving people to leave could be a bitter way to learn.