The question most businesses ask about AI and staff is whether they will end up with fewer people. New cross-country research answers it, and the answer is no: firms that adopt AI employ slightly more people than comparable firms that do not.
What moved was not the headcount. It was the shape of it — and shape is a slower, more expensive problem to fix.
01 — What happened
On Monday 21 September 2026, Bharat Chandar of Stanford and Bouke Klein Teeselink of King’s College London released How Does AI Change Labor Demand? Evidence from 41 Countries through the Stanford Digital Economy Lab. It is built on 1.25 billion job postings and 154 million employment records covering 41 countries from January 2021 to March 2026, with AI adoption identified from job advertisements that genuinely involve generative AI work.
The results, as the authors summarise them: adopting firms grow to have about 3.3% more employees, “suggesting modest productivity improvements”. Within that growth, senior employment rose 6.7% over five years while junior employment fell by roughly 2.5%, and the junior share of the workforce dropped by about two percentage points. That decline showed up in around three-quarters of the 41 countries, including the US, the UK, Brazil and Saudi Arabia. The paper’s conclusion: “AI is labor saving for junior workers and labor expanding for seniors in exposed occupations.” Junior losses run somewhat deeper in richer and more digitised economies.
Two other pieces of evidence point the same way. In August, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen reported a 19% employment shortfall for workers aged 22 to 25 in highly AI-exposed occupations relative to less-exposed peers, while finding no widespread, economy-wide displacement. And a ZipRecruiter survey of more than 1,000 hiring professionals in June found 38% of employers had shifted basic data processing away from entry-level workers, and 31% had raised experience requirements for entry-level jobs.
Two caveats. Adoption here is inferred from what firms advertise, which skews the sample towards larger, digitally connected employers. And the Stanford authors are explicit that their August work is descriptive, not proof of cause.
02 — Why it matters
The junior rung is where codified work lives. Retyping an enquiry into the CRM, chasing a missing document, producing the first draft, triaging what came in overnight — this is the work you can write down, which is exactly the work these tools do first. Nobody set out to remove entry-level roles. They removed the codifiable tasks, and those tasks were concentrated in one place.
So the decision arrives at a business owner’s desk looking like a cost question. Someone leaves, or volumes rise, and the choice is framed as: do we hire, or can software handle this now? Increasingly it can handle a decent share. That is a real saving, and taking it is often correct.
But it is not a cost decision. It is a composition decision, and the research shows the composition these firms are converging on: more people who already know things, fewer people learning them. Note that headcount went up. This is not austerity. It is a re-weighting towards experience — which moves the competition, and the price pressure, to the market for experienced people. The 6.7% figure is not just a hiring pattern. It is a bidding war.
03 — The economics
A junior hire is two purchases in one invoice: capacity now, and capability in three years. Automation buys the first. It does not buy the second, and nothing about the saving tells you that.
Take an illustrative example — a fifteen-person professional services or trade business that has historically taken on one junior a year. Suppose 40% of that role is genuinely codifiable: moving data between systems, first-pass triage, document chasing, standard drafts. Automate that 40% and the arithmetic looks clean. You have removed real hours at a fraction of a salary. Skip the hire altogether and the saving is a whole salary.
Now look at year three, when you need someone at mid-level. You have two ways to get one. Promote, which costs two to three years of a junior’s salary plus senior review time, and produces someone who knows your customers, your systems and your exceptions. Or buy, which costs a recruitment fee, three to six months of ramp, and whatever premium a market charges when every firm that adopted AI is bidding for the same experienced people.
The problem is not the comparison. It is the asymmetry in how the two costs present. The saving from automating the codifiable 40% is immediate, measurable and lands in this year’s accounts. The cost of having no bench is delayed, unmeasurable, and arrives as “we can’t find anyone good” — a sentence nobody attributes to a decision made three years earlier. Businesses do not over-automate the junior rung because the maths is wrong. They do it because only one side of the maths is visible.
There is a formal version of this. A model by Enrique Ide, revised in June 2026, finds that improvements in entry-level automation “increase output upon adoption but can reduce growth and welfare” when they divert juniors away from the most experienced practitioners — because tacit knowledge passes between people, not through documentation.
04 — What businesses should do
Split the role before you automate it. List the tasks in the junior role you are about to remove or not fill, and mark each one codifiable or judgement. If more than half are codifiable, you do not have an automation opportunity — you have a role that has been subsidising a broken process. Automate that half and rewrite the brief for the rest.
Use volume as the trigger. If your business handles, say, 500 inbound enquiries a month and someone spends most of their week retyping them between a form, a mailbox and a spreadsheet, automate that now. That work trains nobody. It is why juniors leave.
Replace incidental learning with deliberate learning. If a junior no longer processes 200 records a month, they should be reviewing the automation’s output on those 200, with a senior checking the review weekly. Judging work teaches a business faster than doing it, provided someone corrects the judgements. Only 22% of employers in the ZipRecruiter survey give all staff mandatory AI training. The tools arrived; the teaching mostly did not.
Price the bench explicitly. Before you decide not to hire, write down what a mid-level person will cost externally in three years, fee and ramp included, and compare it to two years of a junior plus review time. You may still decide not to hire — but make it a decision rather than a default.
Do not read the study as a forecast for your own firm. It measures employers who advertise generative AI work; if you have twelve staff and no AI in your job ads, you are not in that sample. Read it as a description of the labour market you will be hiring into.
05 — Where technology fits
For most businesses, the thing that absorbs junior work is not an AI product. It is an integration.
The codifiable half of a junior role is usually “take this information from system A and put it in system B, correctly, every time”. That is a joining problem, and the cheapest, most reliable automation you can buy. You probably do not need an AI system to stop someone retyping enquiries — you need the form, the mailbox and the CRM connected. Where a junior role looks irreplaceable, the reason is often that the person is the integration: the only component that knows how five disconnected tools relate to each other, which is the pattern we set out in the real cost of disconnected business systems.
AI earns its place a layer up, where the input is unstructured and the output needs a judgement — reading a messy enquiry, summarising a call, drafting a reply that depends on context. We set out where that leverage sits in where AI actually creates economic leverage. But that is also where the learning is, which makes the line to hold a narrow one. Automate the transcription and the chasing. Keep the judgement with a person who is still acquiring it, and have someone senior check them.
Sometimes the answer is to buy nothing. If the junior role exists because two systems do not talk, fix the systems and the role changes on its own. If it exists because nobody wrote the process down, no software helps until someone does.
06 — RevenueStack’s take
Our view: the number everyone will quote is the junior decline. The number that matters is 6.7% senior alongside 3.3% total.
Those firms did not shrink. They re-weighted towards people whose knowledge was expensive to acquire. AI has not reduced the value of experience — it has concentrated it and bid up its price, which is the opposite of what most businesses assumed automation would do to their wage bill.
The risk that follows is not redundancy. It is that a generation of businesses takes a defensible cost saving in 2026, quietly stops manufacturing its own experienced staff, and in 2029 goes shopping for mid-level people in a market where every competitor is doing the same thing at the same time.
Automation is a substitute for tasks. It has never been a substitute for a bench. Treat the junior seat as a training investment with some automatable overhead attached, rather than a cost line with a software alternative, and the decision usually gets easier than it first looks.
Sources
- How Does AI Change Labor Demand? Evidence from 41 Countries — Stanford Digital Economy Lab, 2026-09-21
- AI's impacts on jobs around the world — Bharat Chandar, 2026-09-21
- AI Adoption Is Driving Hiring, but Mostly for Senior Roles — Bloomberg, via Insurance Journal, 2026-09-21
- No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19% — Stanford Digital Economy Lab, 2026-08-12
- More Jobs, Higher Bar: The 2026 AI Employer Report — ZipRecruiter Economic Research, 2026-07-29
- Automation, AI, and the Intergenerational Transmission of Knowledge — arXiv (Enrique Ide), 2026-06-08
