working with ai
whoever checks your ai hasn't been hired yet
if ai writes the first draft, someone still has to know what a good one looks like, and that person has to learn it somewhere. cutting junior hiring now empties the bench that will be checking the machines in 3 years. give the newest people the ai work on purpose.
think back to your first year in a job. somebody handed you the rough first draft, the meeting recap, the competitor roundup nobody else wanted to write. you got it wrong, a senior person marked it up, and you did it again. that loop is how most of us learned what good looks like.
that loop is exactly the work now going to ai. so a reasonable executive asks the obvious question: if a tool writes the first draft, why hire the person who used to?
a lot of them are acting on it. in an april 2026 survey of 415 ceos by the oliver wyman forum and the new york stock exchange, 43 percent planned to pull back on junior hiring in the next year, up from 17 percent a year earlier1. and it may already be showing up in payroll data. a stanford team found that 22 to 25 year-olds in the jobs most exposed to ai are being hired less, an early signal the authors are careful not to call proof2.
what that costs in 3 years
run the plan forward. say that by 2029, ai drafts more of the work than it does today. every one of those drafts still needs a person who can tell a good one from a plausible one, and who knows what to do with the case the tool gets wrong.
that person is today's junior hire, 3 years in. if you didn't hire them, they aren't there. you could plan to hire those people later, from companies that kept training juniors. if 43 percent of ceos are pulling back, there will be fewer of them, and we'd expect them to cost more.
some companies have already gone the other way. ibm said in february 2026 that it would triple its us entry-level hiring in 2026. its hr chief said ai could now do much of what those jobs used to involve, and ibm's answer was to rewrite the jobs and keep hiring3. the oliver wyman forum's own ceo, john romeo, said in the release for that survey that ai is raising the value of entry-level talent1.
the native speakers
our view: younger people are the closest thing these tools have to native speakers. they got there first. in early 2025, pew research center found that 58 percent of americans under 30 had used chatgpt, against 34 percent of adults overall4.
the people starting work also see the risk. kpmg asked 906 of its own summer interns in july 2026 what worried them about ai and their development. the top answer was that leaning on it too much would limit their critical thinking5.
native speakers still need an editor. fluency with the tool and judgment about the work are different things, and the second is what a senior person teaches. the interns' worry is useful, because it names the exact failure the training should guard against: a new hire who accepts whatever the tool produces.
that failure shows up in the data. researchers followed 26,811 middle and high school students in a county in central china for 30 months. once students started using ai on homework, their homework scores rose about 18 percent, and within 6 months their scores on monthly tests taken without ai fell about 20 percent6. the losses were concentrated in students who used ai to produce answers. students who kept doing the work themselves lost little. how people use the tool is what training can change, so train them on that, and give them the work.
know the work before you hand it off
after a panel at a conference, we asked a group of experienced data people who work with pharma data every day a simple question: should you know the process, and the information behind it, before you hand it off to a machine?
we think a lot of people are about to face that question, as more of their work moves to ai. the early data suggests the answer matters, because the skill to check the machine fades when the machine does the work. in poland, doctors who started using an ai tool that flags precancerous growths during colonoscopies spotted fewer of them on their own: in the months after the tool arrived, their detection rate without it fell from 28 percent to 22 percent7. and in a 2025 survey of 319 knowledge workers by microsoft research and carnegie mellon, the more confident people were in the ai, the less critical thinking they reported doing8. if experienced people lose some of the skill once a tool does the work, people who never built it have more to miss.
how to give them the work
have a new hire write the prompt, which is just the instructions the tool follows, for a real piece of recurring work. the weekly competitor roundup. the recap after a conference. the first pass of an internal training deck. then a senior person goes through the result with them, line by line.
some of the time, have them write the first draft themselves and put it next to the tool's version. writing their own teaches them what a good draft takes. the side-by-side teaches them where the tool goes wrong. that second skill is the one we'd expect to matter most once ai writes most first drafts.
in a regulated company, start with work that stays internal and doesn't go through medical, legal and regulatory review: research summaries, meeting notes, internal decks. use only the tools your company has approved, and keep customer and patient data out of it. the mistakes stay cheap while they learn.
if you are setting next year's headcount, write down the 3 recurring jobs a new hire would own, with ai, from their first month. if you can't name 3 yet, work them out with whoever would train that hire, before the headcount decision rather than after.
there's an old saying that you don't want to see how the sausage is made. once a machine makes most of it, someone on your team has to see it, and know it well enough to tell when the machine's batch is off. so hire that person now, and some of the time, have them make it by hand.
sources
- 1oliver wyman forum and the new york stock exchange, ceo survey press release, april 7, 2026. 415 ceos; 43 percent plan to deprioritize hiring for junior roles within the next year, up from 17 percent the year before. includes john romeo's statement that ai is elevating the value of entry-level talent.
- 2erik brynjolfsson, bharat chandar and ruyu chen, "canaries in the coal mine? six facts about the recent employment effects of artificial intelligence," stanford digital economy lab, august 2025, using adp payroll data. reports a 13 percent relative decline in employment for workers aged 22 to 25 in the most ai-exposed occupations, driven mainly by reduced hiring. later revisions report a larger gap; the authors describe the pattern as an early signal, not proof of cause.
- 3bloomberg, february 2026, reporting ibm chief human resources officer nickle lamoreaux's statement that ibm would triple entry-level hiring in the us in 2026 and had rewritten those job descriptions toward customer-facing work. ibm did not give headcount figures. lamoreaux said the entry-level jobs of 2 to 3 years earlier could mostly be done by ai.
- 4pew research center, "34% of u.s. adults have used chatgpt, about double the share in 2023," june 25, 2025. survey fielded february 24 to march 2, 2025; 58 percent of adults under 30 had ever used chatgpt.
- 5kpmg us, "summer intern pulse survey," published august 4, 2026. 906 us summer interns, fielded july 2026. the top concern about ai's effect on their development was over-reliance limiting critical thinking (43 percent). disclosure: the respondents are kpmg's own interns.
- 6david strömberg, victor lei and yanhui wu, "the generative ai learning penalty: evidence from chinese secondary education," cepr discussion paper dp21577, june 2, 2026. 30 months of administrative records for 26,811 students in grades 7 to 12 in one county in central china. ai adoption raised homework scores by about 18 percent; monthly unaided exam scores were about 20 percent lower within 6 months; losses concentrated among the roughly 80 percent of users whose behaviour suggests outsourcing homework. a working paper, not yet peer reviewed, and observational rather than randomized.
- 7krzysztof budzyń et al., "endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study," the lancet gastroenterology & hepatology, 2025. at 4 polish centres, the adenoma detection rate in standard colonoscopies done without ai fell from 28.4 percent (226 of 795) before ai tools were introduced to 22.4 percent (145 of 648) after. observational; the authors describe a deskilling risk, not proof of cause.
- 8hao-ping lee et al., "the impact of generative ai on critical thinking: self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers," chi '25 (microsoft research and carnegie mellon university), april 2025. survey of 319 knowledge workers: higher confidence in generative ai was associated with less critical thinking; higher self-confidence with more.