Who Will Check the AI's Work in Ten Years?

Workers aged 22 to 25 in the jobs most exposed to AI are now 19% below where they would be if they had kept pace with less exposed peers. Experienced workers in those same jobs are holding steady, and the declines cluster where AI automates work rather than assists it. That's from Stanford's payroll analysis of millions of US workers (Brynjolfsson, Chandar, Chen).

The researchers call these early, descriptive findings, not proof that AI caused the drop. Part of the gap predates generative AI and it narrows with education, but it holds after accounting for interest rates and remote work.

The accounting firms point the same way. Each of the Big Four cut its UK graduate intake in figures reported in 2025, KPMG by 29%, with the cuts attributed partly to AI taking over repetitive junior work (HRreview).

The people being hired less are the people who would become your experts in ten years: the ones who know your systems, your customers, and the ways your business breaks. That's the expertise that can tell when a specific AI answer about your business is wrong.

Why Experts Still Grade the Work

Two properties of today's AI make expert judgment hard to replace.

The first is the jagged frontier. In a study of 758 BCG consultants by Ethan Mollick and colleagues, those using AI finished more work, faster, at higher quality on tasks inside the AI's capability. On one task picked because the AI gets it wrong, they were 19 percentage points less likely to reach the correct answer (Dell'Acqua et al.). Nothing on the screen told anyone which side of the line they were on.

The second is sycophancy. Anthropic researchers found five leading assistants consistently tilted their answers toward what the user already believed (Sharma et al.). So the AI pushes back least at the moment you most need it to.

Automated tests and outcome metrics catch a lot of these errors, and you should build them. But tests catch failures someone already defined, and metrics show problems in aggregate, after the fact. Catching a specific wrong answer before it ships usually still takes someone who knows the answer without asking the AI.

That person also has to trust their own judgment: a survey of 319 knowledge workers found that confidence in the AI went with less critical thinking, while confidence in their own ability went with more (Lee et al., CHI 2025).

Distance from the work matters too. In a study at a global trading firm, marketers using AI wrote articles nearly as well as the analysts who normally write them. Software developers and data scientists, with the same AI, did not (Vendraminelli et al.). That study measured producing the work, not checking it. My read, from years of architecture and code reviews, is that checking follows the same rule: the further you sit from the work, the less you can tell a sharp draft from a plausible one.

What Happened in Customer Support

A widely cited study covers 5,179 support agents using generative AI. Issues resolved per hour rose 14% on average and 34% for novice and low-skilled agents (Brynjolfsson, Li, Raymond).

Salesforce went further, cutting support staff as AI agents took on conversations. Headcount went from about 9,000 to about 5,000. CEO Marc Benioff said satisfaction scores held, and that the remaining staff now work alongside the AI agents, coordinated by an omnichannel supervisor (ITPro).

Gartner predicts that by 2027, half of companies that cut service staff because of AI will rehire for similar work under different job titles (Gartner, Feb 2026). That's a forecast, and only a fifth of the 321 leaders it surveyed had cut for AI at all. Gartner also said most recent cuts were influenced by broader economic conditions, not automation alone.

My bet is that the cuts that come back are the ones made without redesigning the work left behind.

The Pipeline Problem Is Forty Years Old

In 1983, cognitive psychologist Lisanne Bainbridge wrote "Ironies of Automation," a five-page paper on factory control rooms that could have been written about AI (Automatica).

Automate most of a job and the human is left with the hardest parts, while the routine work that kept their skills sharp disappears. The better the automation, the more the rare intervention matters, and the less ready the human is to make it.

Now apply that to hiring. The junior work AI absorbs is the work that turns a new hire into someone who can spot a wrong answer. Cut junior hiring deep enough and fewer people ever learn the job by doing it.

The strongest objection is that AI speeds up learning. In the same support study, agents with two months of tenure and AI performed as well as agents with more than six months without it. During outages, when the AI gave no suggestions, agents who had used it still handled chats faster than before AI, so some of the skill stuck. Agents who tended to ignore the AI saw no gain (Brynjolfsson, Li, Raymond).

So AI can shorten the apprenticeship. But the authors note that top performers contributed many of the examples used to train it, and they suggest part of the effect came from spreading those agents' practices to everyone else. My read: when your products, systems, and customers change, the next version of that tool needs a next set of experts to learn from.

The other objection is that you can hire experts from outside. For general skills, you can. Knowledge of your systems, customers, and failure modes is different: it comes from doing the work in your environment.

Star Wall Street analysts who switched firms saw their performance drop for at least five years on average, and for about two years even between equally capable firms. Only those who moved with their team or to a stronger firm held steady (Groysberg, Lee, Nanda). Cardiac surgeons' patient outcomes at a hospital improved with the operations they did there, not with operations elsewhere (Huckman, Pisano). An outside hire still has to make that climb in your environment.

How people work with AI decides whether they build expertise. A study of 244 BCG consultants sorted them by how they used AI (Randazzo et al.). The 14% who split the work deliberately, keeping the judgment calls for themselves, deepened their domain expertise. The 60% who blended AI into every step got better at using AI and held their domain expertise, but didn't deepen it. The 27% who handed the whole task over and accepted the result learned nothing.

Software shows the stakes most clearly. Headcount for developers aged 22 to 25 fell about 20% from late 2022 to September 2025, in payroll data from one of the Stanford researchers (Chandar, Oct 2025). The validation step there already has a name: code review. Fewer junior developers today means fewer people in ten years who can read your codebase critically.

How to Redesign a Role Without Losing Your Validators

In April I wrote that your AI agent shouldn't have your employee's job description. The same logic applies to the people. Before you cut or reshape a role, work through five steps.

1. Break the role into tasks. Use the work people actually do, not the job description. MIT economist David Autor makes the case that tools are levers for expertise, not substitutes for it (Autor, NBER).
2. Have your current experts sort the tasks, starting with the ones where an error costs the most: inside the AI's frontier, outside it, or unknown. Revisit those high-cost tasks every quarter, because the frontier moves. The Organizational Capability Map does this at the workflow level; this step does it inside a single role.
3. Name a validator for every task the AI touches. Pick by distance from the work. In my experience, someone in an adjacent role can stretch to cover it with AI help, and someone in a distant role can't.
4. Measure the outcome the task exists to produce, before and after, including speed.
5. Build junior roles around supervised validation on your own systems. Juniors check the AI's output and keep the judgment calls for themselves, the way the consultants who deepened their expertise did, and an expert checks their checks. Log what the validators catch, so the role has numbers behind it at budget time, and bring HR in early, since splitting roles into tasks changes job grades.

Two Questions the Business Case Leaves Out

Every AI business case I've reviewed counts the hours saved. None of them asked who would check the work.

Before you approve a role redesign, ask two questions. Who in this group can tell when the AI is wrong about our business? Where does the next person who can do that come from?

If you can answer the first and not the second, the savings aren't savings. They're a loan against expertise you've stopped producing, and the bill arrives when your current experts leave.