What jobs will AI replace? The roles employers have actually named

21 Sep 2026 · AIimpacted Editorial · 10 min read · Updated 3 Oct 2026

Short answer: Asked what jobs AI will replace, most answers are forecasts. A narrower question can be checked: which roles have employers themselves named when they cut? Across forty-two companies and 81,486 jobs in the twenty months to August 2026, the named functions are customer support (Salesforce, Just Eat Takeaway, Jumia), HR administration (IBM), payroll and back-office systems (Paycom), help desks, document generation and entry-level programming (OpenText), middle management, finance, legal and internal audit (Cloudflare), financial planning (Domain Money), software engineering (ClickUp, Redis, WiseTech Global, lastminute.com), airline administration (Lufthansa), law-firm business services (Baker McKenzie) and bank back-office process work (Citigroup). Almost every named role sits in the coordination layer of a large company - work that routes information between other people. Nobody in this ledger has said AI replaced a nurse, an electrician or a driver.

Almost every answer to this question is a forecast. A consultancy models 2030, a headline rounds it up, and nobody has to be right until then.

There is a narrower version that can actually be checked: which jobs have employers themselves named? Not which jobs an analyst thinks are exposed - which specific functions a company pointed at while cutting them, in a filing, a memo or an on-record interview. That list is short, it is public, and it is a good deal stranger than the predictions.

As of 3 October 2026 our ledger holds forty-two employers and 81,486 jobs, spanning twenty months from January 2025. Here is what those companies said the work was.

#What jobs will AI replace? The functions employers have actually named

The remaining nineteen - Oracle, HP, Block, Snap, Coinbase, Elastic, C.H. Robinson, Groupon, Gambling.com Group, Acrisure, Allianz, CrowdStrike, DeepL, Artlist, Boozt, Mews, Bird, Monte Carlo and Cognizant - named AI as a cause but not a job type. They account for 38,187 of the jobs - about 47% of the total carries no role description at all, which is itself an answer to the question.

#Who is actually losing jobs to AI?

Ask who is actually losing jobs to AI rather than which jobs, and the concentration is the story.

• Two companies are half of it. Oracle and Citigroup account for 41,000 of the 81,486 jobs. The other forty employers share the rest.
• Two industries are seventeen of the forty-two cases. Information technology (nine events, 29,450 jobs) and banking and finance (eight events, 35,050). Add transport and logistics and three industries cover 86% of the jobs.
• It is overwhelmingly office work at very large employers. Twenty of the forty-two cases are US-based and seven more are cuts the employer described as global. The single exception is instructive: Infopro Digital, a French trade-press publisher, is the only employer here that is neither American nor a technology or finance company - and it is also the smallest cut on the list, at 19 posts.
• The people losing these jobs are mostly not junior. Cloudflare's list is middle management, finance, legal and audit. Domain Money's is qualified financial advisors. This is not only an entry-level story, though the entry-level damage shows up elsewhere - see the jobs that were never created.

So the honest framing is narrow: what we can document is what a handful of enormous technology and finance employers have said about their own white-collar staff. It is not a labour market.

#The pattern: AI is taking the coordination layer

Line the named functions up and they have something in common that the usual “creative versus manual” framing misses entirely.

Customer support routes a question to an answer. HR administration routes a pay query to a policy. Internal audit and revenue recognition check that numbers moved correctly between systems. Middle management routes work between teams. Upwork described what it removed as handoffs and management layers.

None of that work touches physical reality, and none of it carries final accountability. It is the connective tissue of a large organisation - the part that exists because information has to get from one person to another. That is precisely the work a language model does cheaply, and it explains why this list is full of coordinators rather than specialists.

It also explains the one company on the list that does not build software. Angi connects homeowners with plumbers and electricians. When it told regulators that AI efficiencies drove 350 cuts, the plumbers kept working. The people who went were the ones matching, scheduling and following up.

#The headcounts and the shares tell different stories

The two employers that dominate the total are cutting a small slice of themselves. Oracle's 21,000 is about 13% of the company; Citigroup's 20,000 is roughly 8% of its global workforce.

The deepest cuts by share are much further down the table, at companies most people have never counted:

• Block went from about 10,000 staff to under 6,000 - roughly 40%.
• Upwork cut about a quarter of its workforce.
• Cloudflare cut about a fifth.
• Domain Money, a firm of 39 people, cut roughly a third - including about half its financial planning team.

That inversion matters for anyone reading the question personally. The companies making the biggest absolute cuts are trimming; the companies restructuring hardest around AI are small and mid-sized, and at that size a single function can disappear outright. Domain Money had fourteen people in advisory roles. Halving a team of that size is not a trend line, it is the end of a department.

#The clearest single case, and the catch that comes with it

IBM is the only entry with all three pieces: a named system, a named job type and a headcount. AskHR automates 94% of routine HR tasks, and about 200 HR workers went.

The catch is that IBM says the freed-up money went into hiring more programmers and salespeople, and that it planned to triple US entry-level hiring in 2026. Both halves are true at once, and any answer to “what jobs will AI replace” that quotes the first half without the second is selling something. Replacement inside a function is not the same as a smaller company.

Cloudflare is the sharpest version of the same shape: it cut 1,100 people, about 20% of its staff, on the same day it reported record quarterly revenue of $639.8 million, up 34% year on year. These are not companies in trouble shedding cost. That is what makes the cuts worth documenting - and what makes them hard to generalise from.

#What the forecasts get wrong

The standard list of jobs AI will replace leads with writers, illustrators, junior developers and translators - the work that looked most like what the first generative models produced. Twenty months of employer disclosures do not support that ordering at all.

Not one of the forty-two companies said it cut writers - the closest is Infopro Digital, and what it named was sub-editing, the desk that checks and fits other people's copy - and only two, lastminute.com and Livspace, mention design teams at all. Programmers appear, but late and at the bottom of the ladder: OpenText listed “entry-level program” work among the roles a machine would do, and Freshworks said over half its code is now AI-generated or AI-assisted, while IBM said it hired more of them with the money saved by automating HR. What employers named instead was support, administration, audit and middle management - unglamorous work that nobody writes listicles about because it is nobody's dream job.

There is a plain reason for the gap. Forecasts rank jobs by how well a model can imitate the output. Employers cut by where the cost sits and where the workflow can be re-plumbed without anyone outside noticing. Those two rankings are not the same, and so far only one of them has produced actual redundancies.

#What jobs can AI not replace?

The internet is full of confident lists. Ours would have exactly one entry: we do not know.

What our ledger can say is what is absent from it. In twenty months, no employer has told a regulator, a reporter or its own staff that AI replaced a nurse, an electrician, a driver, a cleaner, a teacher or a cook. Every single named function is desk work inside a corporation.

That is absence of evidence, not evidence of safety. Nobody has published a disclosure about a warehouse either, and warehouses are where automation has run longest. Treat any list of “jobs AI cannot replace” that cites no employer statement as what it is: a guess with a headline on it.

#What this list cannot tell you

Three limits, stated plainly, because they are the difference between a floor and a total.

• It counts admissions, not events. A company that cuts a support team and calls it a restructuring never appears here. Most companies do exactly that.
• One entry is a plan, not a completed cut. HP's SEC filings say AI adoption and the resulting efficiencies “enable a reduction in workforce” of 4,000 to 6,000 by the end of fiscal 2028; our ledger records the midpoint, and HP has not said how many have gone so far.
• The largest effect is not a job loss at all. It is the job that is never posted. That channel is invisible to every ledger including this one, and it is the subject of our piece on the three honest answers.

#How to read your own job against this

Not as a risk score - the sample is forty-two companies. As a set of questions worth asking about your own week:

• How much of it is moving information between other people, rather than making or deciding something?
• Does anyone outside your company notice if your specific output stops?
• Could your employer describe what you do in a sentence that begins “routine”? That is the word in IBM's 94%.

If the honest answers point at coordination, the cases above are the ones to read in full. If they point at physical work, final accountability, or a licence with your name on it, nothing in twenty months of employer disclosures speaks to your job at all - which is worth knowing too, and worth remembering next time a forecast says otherwise.

Every event behind the 81,486 is named, dated and linked to its source, and the whole ledger is available as data. If you think we have counted something wrongly, the receipts are there to check.

Figures as of 3 Oct 2026.