Which Jobs Is AI Actually Replacing? The Data

A field guide to which jobs AI is actually cutting, built from a register of 232 source-cited layoffs and what each employer said about them.

Corrected October 10, 2026: a re-check of the register against its sources changed the grade of some cuts in this snapshot, and the figures below are recomputed. CrowdStrike's May 2025 cut and Chegg's May and October 2025 cuts moved from declared AI-related to AI as one reason among several, because each employer's own announcement also gives other reasons, such as restructuring, costs or falling search traffic. Two UPS announcements (about 20,000 jobs in 2025 and up to 30,000 in 2026) were removed from the register, because no credible report tied them to AI; UPS gave its Amazon volume decline as the reason. Klarna's February 2024 statement that its AI assistant did the work of 700 agents is now recorded as a statement about its workforce, not a cut, and Tata Consultancy Services' July 2025 cut was regraded on October 9 to name AI among other reasons. As a result the declared cuts fall from 29 to 26, software engineering among them from 8 to 6, and corporate functions now carry slightly fewer of the declared cuts' roles than customer service. The full list is under Earlier versions of this guide.

Updated October 8, 2026: rewritten on the register as checked for new events through October 3, 2026, which holds 235 job cuts. For this update, the job families recorded for the 119 cuts where the employer named AI and said which roles went were re-checked against their sources, and 68 were corrected. The first version of this guide, published September 16 on about 110 events, found software engineering near the bottom of the cuts employers declared AI-related; it is now at the top. Every change is listed at the end.

Search "what jobs will AI replace" and you get a list of predictions. Nobody asking that question wants a prediction. They want to know whether their own job is on the list, and predictions cannot tell them that, because a prediction is a guess about the future dressed up as a fact about the present. There is a better way to answer the question, and it does not require a forecast at all. It requires looking at the layoffs that have already happened and asking a narrower, harder question: when a company has actually cut jobs and named artificial intelligence as the reason, which jobs did it cut?

That is a question with a real answer, because the events are real, the numbers are disclosed, and the reasons are on the record. We built the AI Layoffs Index to hold exactly that record: a continuously maintained, source-cited register of workforce reductions linked to AI, where every event links to a source you can open and read. As checked for new events through October 3, 2026, and as corrected on October 10, the register held 241 events from 2023 to the end of September 2026. Nine are on-the-record employer statements about AI and the size of their workforce, made without announcing a cut; the other 232 are job cuts, each linked to a filing, a company statement or reputable reporting. This article is what those cuts say about the one question everyone is actually asking.

The short version is that the jobs being cut are not quite the ones the headlines promised, the honest number is smaller than the scary one, and the most common answer to "which jobs?" is one the headlines never give: the employer did not say. Where employers do say, software engineering comes up most often, layers of management come second, and physical work does not appear at all. And part of the damage comes not from AI doing a job but from AI taking away the customers who paid for it. This guide walks through what the receipts actually show, why certain kinds of work are exposed and others are not, which jobs look safe so far, and how to read your own job's exposure without relying on anyone's crystal ball.

Contents

1. The question everyone asks wrong
2. First principles: what actually makes a job exposed
3. What the receipts show: the job families AI cuts land on
4. Software engineering: named most often
5. Customer service: the clearest substitution, and the reversals
6. Back office and corporate functions: the quiet share
7. Middle management: the second most named
8. Writing and editorial: when AI takes the readers, not the work
9. What the exposure research predicted, and where the receipts agree
10. The jobs that look safe (and why)
11. Said is not caused: why the honest number is smaller
12. How to read your own job's exposure
13. What to watch next
14. The bottom line

1. The question everyone asks wrong

The most-searched version of this topic is some variant of "what jobs will AI replace" or "will AI take my job." Those are natural questions and they are also, as usually framed, unanswerable. "Will" is the problem. It invites a forecast, and forecasts about technology and employment have a long and humbling track record of being confidently wrong in both directions. The 1950s expected automation to abolish work. The 2010s expected self-driving trucks to end driving as a profession by 2020. Neither happened on schedule, and the gap between the confident prediction and the messy reality was where all the interesting information lived.

There is a more useful frame, and it is the one this entire site is organized around. Instead of asking what AI will do to jobs, ask what employers have already done and said AI was the reason. This swaps an unfalsifiable guess for an auditable fact. A company that files an SEC Form 8-K disclosing a restructuring and attributes it to automation is not speculating about the future, it is documenting the present, under legal obligation to be accurate. When enough of those documented events accumulate, a pattern emerges that no amount of punditry can match, because it is made of things that actually occurred.

This reframing matters for a practical reason too. A prediction cannot be checked, so it cannot be corrected, so it drifts toward whatever is most dramatic. Evidence can be checked, which disciplines it. The difference shows up immediately in the numbers.

• A prediction says AI will eliminate a category of work entirely and soon.
• The evidence says AI-attributed cuts are real, concentrated, growing, and far smaller than the total layoff wave they are blamed for.
• The gap between them is the AI-washing problem, and it is measurable rather than rhetorical.

Working from evidence rather than prediction changes both the size and the shape of the answer. The size shrinks, because most layoffs blamed on AI in aggregate turn out, event by event, to have other stated causes. The shape sharpens, because the cuts where AI genuinely is the named cause cluster in specific kinds of work rather than spreading evenly across the economy. The rest of this guide is about that size and that shape. We covered the size question in depth in our companion piece on what the 2026 layoff data actually shows, and the live count of roles is on our page answering how many jobs AI has replaced; here the focus is the shape, meaning which jobs.

2. First principles: what actually makes a job exposed

Before looking at which jobs are being cut, it helps to reason from first principles about which jobs should be exposed, so that the data has something to confirm or refute. The common shorthand is "white collar versus blue collar," and it is close to useless. Plenty of white-collar work is safe and some of it is highly exposed, and the dividing line runs straight through the middle of the office, not around it. The better unit of analysis is not the job title but the task.

Today's AI is, at its core, a system that turns patterns in language and data into plausible outputs at near-zero marginal cost. That single capability defines the exposure boundary. A task is exposed to the degree that it consists of transforming information according to learnable patterns, is repeated often enough for those patterns to matter, happens through a screen rather than in physical space, and produces an output whose quality can be checked quickly. Stack those properties up and you get a clean structural test that has nothing to do with collar color and everything to do with the shape of the work.

The properties that raise a task's exposure are consistent and they compound:

• Language or data heavy: the work is reading, writing, summarizing, classifying, or answering.
• Routine and repeated: the same kind of problem recurs, so a pattern generalizes across instances.
• Digitally mediated: it happens entirely through software, with no physical or in-person step.
• Individually checkable: an output can be verified fast, so a human can supervise many of them.

A task with all four properties is a task a language model can now do a meaningful share of, and a job made mostly of such tasks is a job under pressure. A task with none of them, plumbing a house, settling a frightened patient, negotiating a contract where trust is the product, is a task the current generation of AI barely touches. Most real jobs are a bundle of tasks somewhere in between, which is why the honest answer to "will AI replace this job" is almost always "it will replace some of its tasks, and whether that removes the job depends on how many and which ones."

There is also a second, less obvious pathway to job loss that the task framework alone misses, and it matters because it hits a different set of people. The first pathway is task substitution: AI does part of your job, so the employer needs fewer of you. The second is demand destruction: AI does not touch your job at all, but it destroys the market for your employer's product, so the company shrinks or fails and the jobs go with it. The cleanest example in the register is Chegg, the homework-help company, which cut 248 roles in May 2025 and 388 in October 2025, saying "the new realities of AI and reduced traffic from Google to content publishers have led to a significant decline in Chegg's traffic and revenue" - SEC. Chegg's cuts followed a collapse in demand rather than a decision to automate a function, although the same release says its homework-help product is "moving to an AI-powered model that answers all new questions automatically." This distinction is why a low-exposure job can still be at risk (if it sits inside a business AI is disrupting) and why "is my task automatable" is a necessary but not sufficient question. Section 8 shows that this second pathway is no longer a single example.

Hold both pathways in mind, because the receipts, when we turn to them, line up with the framework closely. That alignment is the strongest evidence that the framework is right, and it is also the reason the cuts are landing where they are.

3. What the receipts show: the job families AI cuts land on

Now the evidence. For every cut, the register records which roles went, in the employer's words where it gave them and otherwise from the reporting it cites, and groups them into broad job families: software engineering and IT, back-office operations, corporate functions such as finance, HR and legal, customer service, sales and marketing, management layers, and so on. A restructuring usually spans several families, so a cut counts once in each family it touched, and the counts overlap. For this update every one of those labels was checked against its sources under one rule: a family is recorded only for roles the employer, or named reporting, says were cut, never for the areas where a company says it is applying AI or still hiring.

The register also grades every cut by who connected it to AI, and the grade matters as much as the job families. In 26 of the 232 cuts, the employer itself declared the cut AI-related. In 150, the employer named AI as one reason among several, next to costs, a reorganization or a shift in strategy. In the remaining 56, only the press or other observers linked the cut to AI, while the employer said something else or nothing at all. That leaves 176 cuts where the employer itself named AI, and those are the ones the chart below counts.

The first finding is the one the headlines never report: in 83 of those 176 cuts, the employer did not say which roles went. Companies routinely announce a percentage, a severance budget or a new strategy built around AI and leave the jobs unnamed. Dow, for example, is cutting about 4,500 roles under a program meant to add at least $2 billion in earnings "in part by utilizing AI and automation to deliver step change in growth and productivity," and its filing does not say which roles - SEC. Any confident ranking of "the jobs AI is cutting" has to start from the fact that for almost half of the cuts, nobody outside the company knows.

Among the 93 cuts that do name roles, software engineering and IT appears in 44, almost half. Management layers come next at 27, then sales and marketing (25), corporate functions (24) and customer service (20). Physical work does not register at all: frontline and manual work appears in none of the 176.

These counts are a snapshot of the register as it stood in early October 2026. Each family now has its own page listing every cut in it, kept current as the register grows: software engineering and IT, management layers, sales and marketing, corporate functions, customer service, back-office operations, product and design and writing, editorial and translation.

But "named AI" covers two very different situations, and the difference between them is the central finding of this guide. In most of these cuts, 150 of the 176, the employer named AI as one reason among several, sometimes in the sense of AI doing the work, and sometimes in the sense of moving money into AI. In only 26 did the employer declare the cut AI-related outright. Filter to those 26, and the picture narrows further.

Software engineering leads here too, in 6 of the 26 declared cuts, but 12 of the 26 do not say which roles went, and the declared cuts are small: the register counts 4,569 roles across all of them. The largest single one is Allianz Partners' 1,500, in customer service. That is the shape of the answer in one sentence: employers mention AI most often in cuts that touch engineering, the thing they most often say an AI-linked cut removes after that is a layer of management, and the largest cut they pinned squarely on AI fell on customer service. The next sections take the largest families in turn.

4. Software engineering: named most often

Software engineering is where the picture has changed most since this guide was first written. In September, on a register of about 110 events, engineering appeared in only 2 of the 21 events employers had then declared AI-related (as corrected on October 1), last of every function. On today's register it is the family employers name most often, in 44 of the 93 cuts that say which roles went, and in 6 of the 26 they declared AI-related. Part of that change is coverage: checks of 2025 and 2026 have since added more than a hundred events, many of them at software companies. Part of it is a change in what employers say: all six declared engineering cuts were announced in August 2025 or later.

Most of the engineering count comes from cuts where AI was one reason among several, and the employers' own words show what kind of reason. Meta cut about 8,000 roles in May 2026, roughly a tenth of its staff, and booked the severance to research and development and to marketing and sales - SEC. Its finance chief had told investors that a leaner operating model would help offset the substantial investments Meta is making, and at an April 30 town hall Mark Zuckerberg told staff that getting employees to use AI tools to work more efficiently "is not the thing that's driving layoffs" - TheStreet via Yahoo Finance. Atlassian cut about 1,600 roles, about a tenth of its staff and more than 900 of them in software R&D according to press reports, to "self-fund further investment in AI and enterprise sales." Its chief executive, Mike Cannon-Brookes, wrote that the company's approach is not "AI replaces people," but that "it would be disingenuous to pretend AI doesn't change the mix of skills we need or the number of roles required in certain areas" - SEC. In both, AI is the reason the money moved, not a system doing the departing engineers' work.

Others say more directly that AI changes how many engineers they need, while still naming other reasons. The Australian logistics-software maker WiseTech Global announced a phased cut of up to half of its product-and-development and customer-service headcount, about 2,000 roles, as part of "the shift to becoming an AI-led organization," alongside a focus on higher-margin revenue and a higher-performance culture - WiseTech Global. Visa is cutting about 2,600 roles, concentrated in technology and product, with its chief executive writing that AI is helping to "shape the way work gets done at Visa" - Reuters via Yahoo Finance. Snap cut about 1,000, saying advances in AI let its teams "reduce repetitive work, increase velocity" - Snap, and booked the largest share of the severance to research and development - SEC.

The six engineering cuts declared AI-related outright are smaller, and most of them come from software and internet companies saying AI now does more of the engineering work. ClickUp's chief executive said the cut of 22% of its staff "wasn't about cutting costs," and that its best engineers are "becoming 100x engineers. They're not writing code. They're directing agents that write code" - Zeb Evans on X. The drug-discovery software maker BenchSci laid off 23% of its workforce, 83 people, and has used AI tools to take over their tasks - NOW Toronto. Gambling.com Group's chief executive told investors that "this transition to AI-first ways of working will allow us to move faster and with fewer people" - Gambling.com Group earnings call via The Motley Fool. Fiverr, Yellow.ai and Tailwind Labs complete the six. Across all six, the register counts 486 roles; press estimates for ClickUp and Yellow.ai would add about 390 more, which the register shows but does not count because the companies gave no number themselves. CrowdStrike's George Kurtz wrote that "AI flattens our hiring curve" as the company cut about 500 roles - SEC, and its quarterly filing booked the largest share of the charges to research and development - SEC, but the same memo gives other reasons, realigning the business and simplifying how it executes, so the register records AI as one reason among several there; Chegg's May 2025 cut is recorded the same way.

One of the six is the purest case of the second pathway from Section 2. Tailwind Labs, the small company behind the Tailwind CSS framework, cut most of its engineering team in January 2026. "75% of the people on our engineering team lost their jobs here yesterday because of the brutal impact AI has had on our business," its founder wrote: traffic to Tailwind's documentation, "the only way people find out about our commercial products," was down about 40% from early 2023 even as the framework grew more popular than ever, and revenue was down close to 80% - GitHub. AI did not do those engineers' work. It took away the demand that paid for it.

Engineering also appears in 21 cuts where only the press tied the layoff to AI. Microsoft is the clearest example. Three of its rounds of cuts, from May 2025 to July 2026, are in the register with about 19,800 reported roles between them, and in July 2026 its chief people officer wrote that "the roles eliminated today are not being replaced by AI" - Microsoft. Those cuts stay in the register as context, because the AI link is part of the public record, but none of them counts as a cut where the employer named AI.

The first-principles read reconciles these facts. Software work splits into tasks with very different exposure. Writing routine, well-specified code, generating tests, translating between formats, and producing boilerplate are language-and-pattern tasks that current AI does well, and those tasks made up a real fraction of junior and mid-level engineering work. Architecting a system, deciding what to build, debugging a subtle production failure, and owning the tradeoffs are judgment tasks that AI assists but does not replace. The employers' own words fit that split: the companies declaring engineering cuts outright describe AI doing more of the work and engineers directing it, not engineering without engineers. The US employs roughly 1,687,890 software developers, per the Bureau of Labor Statistics - BLS, so even a modest shift in that task mix moves a lot of jobs. That is why the honest statement is neither "engineers are safe" nor "engineering is being automated," but that the task mix inside engineering is shifting, compressing the demand for routine implementation while raising the premium on design, judgment, and the ability to direct AI tools well.

5. Customer service: the clearest substitution, and the reversals

If engineering is where employers most often mention AI, customer service is where they most concretely describe AI doing the work. It appears in 20 of the 93 cuts that name roles and in 5 of the 26 declared AI-related, among them the largest declared cut in the register. Customer support is the one office function where the AI replacement is a discrete, deployable product (a chatbot, a voice agent, an automated claims system) rather than a vague "productivity gain," which is why the cases here are so specific.

The largest is Allianz Partners, the travel-insurance and assistance arm of Allianz, whose chief executive said in July 2026 that it would cut 1,500 to 1,800 jobs across Europe because of the growing use of AI; the register counts the lower figure - Insurance Journal. Earlier reports citing company sources said the cuts would fall mainly in call centres - AFP via Malay Mail. Where AI was one reason among several, the cases are similar in kind: WiseTech Global's plan reaches up to half of its customer-service headcount (Section 4), and Danske Bank cut 262 roles, including customer-facing roles in personal banking, citing "digital tools and AI based solutions that help handle simpler tasks more efficiently" - Danske Bank.

Several of the best-known customer-service cases are in the register but not counted, and the reasons are instructive. Salesforce's Marc Benioff said in August 2025 that the company's AI agents were handling roughly half of customer conversations and that he had taken support from about 9,000 heads to about 5,000. Salesforce then explained that it no longer needed to actively backfill support engineer roles and had redeployed hundreds of employees into other teams - Fox Business, so the 4,000 that circulates as a layoff figure is mostly roles that were not refilled: the register shows it as reported but does not count it as layoffs. Klarna said in February 2024 that its AI assistant was doing the equivalent work of 700 full-time agents - Klarna, and the next day that it had made no cuts because of it, since its customer service is run by outside partners - Gizmodo. Its own headcount did fall, from about 4,500 to about 3,500 by late 2024, through a hiring freeze and attrition rather than layoffs - TechCrunch. In May 2025 its chief executive conceded that "what you end up having is lower quality" - CX Dive, and Klarna later said it would use human agents to offer a "VIP" customer service - TechCrunch.

The reversals are evidence too. Commonwealth Bank of Australia said it would cut 45 call-center roles after an AI voice bot took over customer queries, then reversed the decision and admitted the roles were not redundant - The Nightly. Of the two most-cited AI-support stories in the world, Commonwealth Bank walked its cut back and Klarna partly walked its approach back. Even the outsourcing industry that supplies this work to everyone else is less clear-cut than the headlines: Teleperformance, one of the world's largest call-center operators, says its 2026 restructuring plans cover about 3,300 employees as part of its AI-related transformation - Teleperformance, but the only roles its efficiency program names are management layers, cut "through delayering, automation and optimized procurement practices" - Teleperformance.

Support work has all four exposure properties from our framework at once: it is language-heavy (it is literally answering questions in text or speech), intensely repetitive (the same questions recur millions of times), fully digital (it happens through a ticketing system or a phone line), and instantly checkable (did the customer's problem get solved). That combination is why the AI systems aimed at support are the most mature, and why the cases here are the most concrete. It is also why support is where the technology's limits show up first. The exposed population is enormous: the US employs about 2,666,000 customer service representatives, according to the Bureau of Labor Statistics, making it one of the largest occupations in the country - BLS. A technology that can handle even a meaningful minority of routine support interactions is pointed at millions of jobs, which is why a handful of high-profile deployments generate so much anxiety.

The honest read of the customer-service signal is therefore not "support jobs are gone" but "support is where AI is being substituted for labor most concretely, and with the most visible trial and error." If your work is answering repetitive questions through a screen, you are closest to the front line of this, and the practical response is not panic but repositioning toward the parts of support AI handles worst: the escalations, the judgment calls, the angry-customer recovery, the genuinely novel problem. Those are the tasks that fail the "individually checkable" and "routine" tests, and they are where human support is consolidating.

6. Back office and corporate functions: the quiet share

Behind the loudest families sit the roles that keep an organization running without facing a customer: billing, claims processing, payments operations, data entry and reconciliation in the back office, and finance, HR, legal and administration in corporate functions. Corporate functions appear in 24 of the 93 cuts that name roles and back-office operations in 14; among the 26 declared AI-related, in 3 and 3. They rarely make dramatic headlines because no single vendor sells a flashy "operations agent" the way Salesforce sells Agentforce, but corporate functions carry almost as many of the declared cuts' roles as customer service, the family with the most.

Of the 4,569 roles the register counts across the 26 declared cuts, the cuts naming corporate functions account for 1,790, second only to customer service's 1,795, most of them from two cases. Cloudflare cut more than 1,100 roles, with its founders saying the cut was not a cost-cutting exercise but about how the company operates "in the agentic AI era" - Cloudflare. Its chief executive, Matthew Prince, later wrote that "the vast majority of those we laid off last week were measurers," a group that, as he described it, includes middle management, finance, legal, internal auditing and revenue recognition - The Indian Express. The payroll-software firm Paycom cut about 540 jobs - SEC, which it attributed to "efficiencies in advanced automation and AI-driven technologies that will impact a limited number of back-office roles" - Dallas Express via Yahoo Finance, saying the cuts affected "only non-client-facing roles" - KGOU.

The Utah physician group Revere Health names the exact tier of work. It laid off 177 people, most of them in accounts receivable or medical coding, as technology that includes "machine learning and automated claims processing" took over billing, collections and denial prevention through a new partner, IKS Health; patient services, it said, would be "uninterrupted" - The Salt Lake Tribune. That case is a useful microcosm. Back-office processing is information work of the purest kind: reading documents, applying rules, filling forms, and chasing exceptions, all through software, all repetitive, all checkable. This is the structural profile our framework flagged as maximally exposed, and it is why a medical group that automates its billing office does not automate its nurses: bodily, high-stakes, in-person work fails the digital-mediation and routine tests at the most basic level, so the cut lands on the processing desk and stops there.

Goldman Sachs shows why the register does not read "where AI is applied" as "who was cut." Its leaders wrote that "to fully benefit from the promise of AI" the firm would make "a limited reduction in roles across the firm," and named client onboarding, lending and regulatory reporting among the processes it would rebuild with AI - Business Insider via Yahoo Finance. Those are the processes Goldman is changing, not the people it let go, so the register records the roles as not specified. For the professions that forecasts most often put at risk, the record is thinner than the forecasts suggest: of the 176 cuts where the employer named AI, 4 name finance or accounting roles and 1 names legal roles (Cloudflare's), each time alongside other roles.

The reason this category matters more than its headlines suggest is scale and diffusion. Back-office and administrative roles are enormous in aggregate and spread across every industry, not concentrated in tech. When the cost of processing a document or reconciling an account falls toward zero, the effect is not one dramatic layoff but a thousand quiet ones, a few dozen roles at a time, at companies that never trend on social media. This is also the category where the "productivity, not replacement" language is most common: firms describe doing the same work with fewer people rather than eliminating a function outright. For a worker, the practical signal is that pure processing roles are, alongside support, the most exposed tier of office work, and the durable positioning is toward the exception handling, the judgment, the cross-system problem-solving, and the relationship management that rules-based automation cannot close out on its own.

7. Middle management: the second most named

A theme runs underneath many of the cuts, and it is easy to miss because it is rarely the headline: middle management. When a company reorganizes around AI, one of the first things it does is flatten, removing layers of coordination whose main job was to route information between people. That coordination work, gathering status, summarizing it upward, translating priorities downward, is itself highly exposed, because it is language work that AI now does cheaply. Management layers appear in 27 of the 93 cuts that name roles, second only to software engineering, and in 3 of the 26 declared AI-related. When employers say what an AI-linked cut removes, "layers" is one of the most common answers.

The clearest declared case is Fiverr, whose chief executive told staff the company would become "an AI-first company that's leaner, faster" with "a smaller team" and "far fewer management layers," as it parted ways with about 250 people across departments - SEC. Cloudflare's cut, above, was mostly middle management and other "measurers," and Gambling.com Group described its AI-first restructuring as producing "a flatter organization, fewer management layers" - The Motley Fool. Where AI is one reason among several, the language is the same. Wix cut about a fifth of its staff in what its chief executive called a company-wide change, giving two reasons: first the shekel's rise against the dollar, and second "the fast evolution of AI capabilities," which he said requires "a faster, leaner, and flatter organization" with fewer levels of leadership - Avishai Abrahami on X. The fintech Chime cut about 150 people, about a tenth of its staff, with its chief executive writing that "smaller teams with fewer layers are moving faster than ever" - Reuters via Yahoo Finance.

Amazon shows why the register reads what each employer said rather than the headline. Its October 2025 memo announcing about 14,000 corporate cuts called AI "the most transformative technology we've seen since the Internet" and spoke of removing layers - Amazon. On the company's earnings call two days later, chief executive Andy Jassy told analysts the cut was "not really financially driven, and it's not even really AI-driven, not right now. It's culture" - TheWrap. Its January 2026 memo announcing about 16,000 more cuts, again about reducing layers, did not mention AI at all - Amazon. The register records both rounds and counts the roles of neither. Flattening is real, but whether a given flattening is about AI is a question each employer answers for itself, and Amazon's answer changed in two days.

The structural logic is worth stating plainly because it is counterintuitive to anyone who assumed seniority was protection. A manager's exposure depends on how much of the role is information-routing versus genuine leadership. The parts that are collecting updates, compiling reports, and relaying decisions are precisely the language-transformation tasks AI absorbs. The parts that are setting direction, making high-stakes calls, developing people, and taking accountability are not. So the layer most at risk is not management in general but the coordination-heavy middle, the role that exists mostly to move information between the people who do the work and the people who set the strategy. For anyone in that position, the durable move is toward the accountable, people-developing, direction-setting parts of leadership and away from being a human message bus, because the message bus is the part that just got automated.

8. Writing and editorial: when AI takes the readers, not the work

One of the smallest job families in this guide carries some of its most distinctive evidence. Writing, editorial and translation work appears in 7 of the 93 cuts that name roles, and in 1 declared AI-related, which is also the closest thing in the register to a model doing a writer's job: the French business-media group Infopro Digital planned to cut 19 sub-editor posts across its titles, according to the journalists' federation, and at its Gisi unit, which estimated that AI can do 70% of a sub-editor's tasks, a court suspended the eight planned cuts in September 2026, citing a serious risk to staff health - BFM TV.

Three of the others tell a different story, the second pathway from Section 2. Reach, the British publisher of the Mirror and the Express, announced about 220 editorial redundancies in September 2026, blaming industry decline and shifts in how people consume news, "with AI overviews and changes to Google algorithms resulting in a major loss of traffic," as The Journal reported it - The Journal. The Washington Post cut more than 300 newsroom jobs, CNN and The New York Times reported - Poynter, after its executive editor told staff that its organic search traffic had fallen by "nearly half" in three years - Nieman Lab, a decline his memo tied to the growth of AI, according to Poynter's account of it - Poynter. Business Insider cut about a fifth of its staff, citing falling search traffic as it went "all-in on AI" - Variety. None of the three said a model now writes its articles. All three described losing the readers who used to arrive through search. The Post had also licensed its journalism to OpenAI in April 2025, to be shown in OpenAI's answers on terms not disclosed, one of the publisher deals in AIData.Broker's record of AI data licensing deals, which is maintained by the same company as this site.

The same mechanism shows up outside the newsroom. Chegg cut 22% of its staff in May 2025, saying Google's AI Overviews "continues to keep web traffic captive in the Google search experience" and that OpenAI had begun giving college students free access to ChatGPT Plus - SEC, then 45% in October 2025, citing "the new realities of AI" - SEC. Tailwind Labs, from Section 4, is the same story in software: its founder tied the loss of documentation traffic, and of the revenue that came through it, to AI. In each case the employer's account is less about a model taking over the work than about AI taking away the readers, students or developers who used to come to it.

This is the pathway the task framework alone would miss, and it changes the advice. A writer whose own tasks are only partly exposed can still lose a job if the business above them depends on search traffic that AI answers now intercept. It is also a pathway that exposure studies, which score tasks rather than business models, cannot see at all. If your employer's revenue depends on people finding it through search, the question to ask is not only "can AI do my tasks" but "can AI answer the question that used to bring the customer to us." For writers and editors weighing their options, AI training is one market where the editorial skill itself is paid, writing reference answers and editing the text that teaches models to write well; the writing roles on AITraining.jobs, from the same team behind this site, show which platforms hire and what they pay.

9. What the exposure research predicted, and where the receipts agree

A fair objection to everything so far is that a register of 232 cuts, however well sourced, is still a sample, and samples can mislead. The strongest test of the pattern is therefore whether it matches independent research built with a completely different method. It does, closely, and that convergence is worth dwelling on because two different instruments pointing at the same answer is far more convincing than either alone.

The foundational study here is the 2023 paper "GPTs are GPTs" by researchers at OpenAI, OpenResearch and the University of Pennsylvania, which estimated task-level exposure to large language models across the US workforce. Its headline finding was that around 80% of US workers could have at least 10% of their tasks affected by LLMs, and roughly 19% could see at least 50% of their tasks affected - arXiv. Crucially, the paper found exposure concentrated in exactly the profile our framework describes: information-processing roles with routine, language-heavy tasks, and it found the lowest exposure in physical and manual occupations. That is the same boundary the receipts trace, derived years earlier from a completely different method.

Real-world AI usage data tells the same story from a third angle. The Anthropic Economic Index, which analyzes what people actually use Claude for, found that computer and mathematical tasks dominate usage at around 36%, exactly the digital, language-heavy work at the top of the exposure ranking - Anthropic. Microsoft Research's study of 200,000 anonymized conversations with Bing Copilot draws the same line: of the 22 major occupation groups, computer and mathematical occupations and sales share the highest AI applicability score (0.29) and healthcare support has the lowest (0.05) - arXiv. Our guide to AI-proof jobs charts those scores group by group.

The macro layoff data agrees on direction while adding a caution about magnitude. Challenger, Gray & Christmas, which records the reasons US employers give when they announce cuts, counted 120,136 announced job cuts attributed to AI in 2026 through September, about 21% of all announced cuts - Challenger, Gray & Christmas, already more than double the 54,836 of all of 2025 - Challenger, Gray & Christmas. The monthly pace has cooled, though: in September AI was the fifth most-cited reason, with 3,961 cuts. Those are announcements and stated reasons, not verified causes, which is exactly the gap Section 11 is about.

The convergence matters because it moves the conclusion from "here is a pattern in our data" to "here is a structural feature of the technology that multiple methods detect." When a theoretical exposure model built in 2023, two live usage datasets, a macro layoff tracker, and a hand-checked register of cuts all point at the same kinds of work, the burden of proof shifts to anyone claiming the pattern is an artifact. The task-shape framework from Section 2 is not a story we told to organize the data; it is the thing these instruments are independently measuring. That is also why the diagram below, which is just the framework drawn out, doubles as a summary of the evidence.

10. The jobs that look safe (and why)

The mirror image of exposure is protection, and the same framework that predicts which jobs are being cut predicts which are not. The occupations that show up least in both the register and the exposure research share a profile: their core tasks are physical, in-person, high-dexterity, or high-trust, meaning they fail one or more of the four exposure properties in a way no amount of language modeling fixes. People search for "AI proof jobs" and "jobs safe from AI" thousands of times a month, and the honest answer is that "safe" is the wrong word but "far less exposed" is a real and defensible category.

The least-exposed work clusters in a few recognizable places, and the reason is structural in every case rather than sentimental:

• Skilled trades: electricians, plumbers, HVAC technicians, whose work is physical, variable, and on-site.
• Hands-on healthcare: nurses, therapists, dentists, home health aides, where the task is bodily and high-stakes.
• In-person services: stylists, chefs, childcare workers, whose value is the physical presence itself.
• High-trust dealmaking: senior sales, negotiation, and relationship roles where the human is the product.

The 2023 exposure research found precisely this, and it is worth being concrete about how precise. The "GPTs are GPTs" paper included an appendix listing occupations with no tasks at all exposed to language models, and all of them were physical: agricultural equipment operators, short-order cooks, dishwashers, meat cutters, motorcycle mechanics, stonemasons, tire repairers, and wellhead pumpers among them - arXiv. Pew Research, rating occupations by how much of their work involves AI-exposed activities in the government's O*NET database, likewise found barbers, childcare workers, and nannies among the least exposed - Pew Research Center.

The register agrees. Of the 176 cuts where the employer named AI, none touches frontline or manual work, the skilled trades or hands-on care; even at Revere Health, a physician group, the cut fell on the billing office. The largest frontline cuts of the period, such as UPS's, came with other reasons: UPS tied its 2025 and 2026 reductions to the decline in its Amazon volume and to automating its network, never to AI. This is not because those workers are lucky but because their tasks fail the digital-mediation test at the most basic level. AI operates on information through screens; work that happens in physical space with physical objects is largely out of its reach with current technology, and robotics runs on a different and slower timetable.

Two cautions keep this from becoming false comfort. First, "less exposed" is not "immune," because even physical jobs contain some exposed tasks (the scheduling, the paperwork, the customer communication around the physical work), and those tasks can thin the administrative support around a trade even when the trade itself is safe. Second, exposure is not destiny in either direction: a low-exposure job can still be affected by second-order economic shifts, as Section 8 shows, and a high-exposure job can persist for years because deploying AI at scale is harder than demonstrating it once. The safe-jobs list is best read not as a place to hide but as evidence for the underlying rule: the further your work sits from routine information processing on a screen, the more insulated it is, and that rule is worth applying to your own situation directly rather than looking your title up on someone's list. Our guide to AI-proof jobs takes this side of the question much further, ranking all 22 occupation groups by how sheltered they are on four kinds of evidence.

11. Said is not caused: why the honest number is smaller

Everything above rests on taking company statements seriously, which means it also has to reckon with the fact that company statements are not neutral. There is a strong incentive, right now, to attribute layoffs to AI whether or not AI caused them, and understanding that incentive is essential to reading the data honestly. This is the single most important caveat in the entire guide, and it is why the register grades attribution rather than simply counting mentions.

Blaming AI for a layoff is, for a public company, an unusually convenient story. It reframes a painful cost cut as a forward-looking technology bet, which investors reward. It signals to the market that the company is an AI adopter rather than an AI victim. And it deflects from less flattering explanations like over-hiring, weak demand, or strategic missteps. The result is a systematic upward bias in AI attribution, which is exactly why the aggregate figures (a Challenger monthly total, or May 2026's widely reported finding that AI was behind 40% of that month's announced cuts) overstate the reality when read as measurements. They are measurements of what companies say, not of what AI did. The incentive can even run both ways inside one company: Intuit presented its 2026 cut as simplifying the company to focus on AI, and its chief executive's email to staff spoke of removing layers of management and streamlining parts of its engineering and product organizations, but he told CNBC's Jim Cramer that "None of it had to do with AI" - CNBC.

The register is built to correct for this, and the correction is large. Of the 232 cuts, the employer declared only 26 AI-related, named AI as one reason among several in 150, and in 56 said something else or nothing at all while the press or other observers linked the cut to AI. In plain terms, in about eight of every nine cuts that get called AI layoffs, the "AI layoff" framing is doing more work than the employer's own words support.

This is why our index counts conservatively and why it separates the disclosed number from the attributed one. When a company reports a headcount figure but does not credibly tie it to AI, or the figure includes contractors, roles that were never refilled, or jobs moved elsewhere, we keep the figure visible but do not count it, a discipline we document in full on the methodology page. The practical upshot for a reader trying to gauge their own risk is to apply the same skepticism: when a company blames AI for a cut, ask whether it named a specific system doing specific work (as Salesforce did with its support agents, or Revere Health with automated claims processing) or whether "AI" is standing in for a decision the company would rather not explain. The specific, product-named cases are the ones that tell you where the technology is genuinely substituting for labor. The vague ones tell you mostly about corporate communications. Reading the difference is the whole skill, and it is the reason a source-cited register beats a headline count.

12. How to read your own job's exposure

All of this converges on a practical exercise you can do for your own role in about ten minutes, which is more useful than any list of "jobs AI will replace" because it works for jobs no list has thought about. The method follows directly from the task-shape framework, and it deliberately ignores your title, because your title is not what is exposed, your tasks are.

Start by writing down how you actually spend a working week, not your job description but your real tasks, and estimate the share of your time each one takes. Then score each task against the four exposure properties: is it language or data heavy, is it routine and repeated, is it fully digital, and is its output quickly checkable. A task that is all four is highly exposed; a task that is none is well insulated; most are partial. The point is not to compute a precise number but to see the shape of your week, because that shape is your real exposure and it is almost always more nuanced than "my job is safe" or "my job is doomed." Then ask the second-pathway question from Section 8 about your employer: does the business depend on customers who could now get what they came for from an AI answer instead?

The interpretation of that shape follows a simple logic, and it is more actionable than a verdict:

• Mostly exposed tasks: expect the role to shrink or change; move toward the judgment and exception parts.
• Mixed: expect the routine share to be automated and the rest to intensify; get fluent at directing AI tools.
• Mostly insulated: expect stability, but watch the administrative edges of the role.

What this exercise reveals, again and again, is that the unit of disruption is the task, not the job, and that most jobs survive in changed form rather than vanishing. The register bears this out: even in the cuts employers declared AI-related, companies were usually cutting a portion of a function and reshaping the rest, not deleting the function entirely. The workers who navigate this best are not the ones in "safe" jobs, they are the ones who identified their exposed tasks early and deliberately shifted their time toward the parts of the work that fail the exposure test. That is a strategy available in almost any role, and it beats waiting to find out whether your title makes someone's prediction list. If you want to see the raw evidence for your own industry, the register is browsable by company and by event, the live count by job family is on our page answering what jobs AI will replace, and the underlying data is free and open to download and analyze yourself.

13. What to watch next

Because this is a live question rather than a settled one, the honest final section is about leading indicators rather than conclusions. The pattern in the data today, most employers not saying which jobs went, software engineering named most often where they do, management layers close behind, customer service carrying the largest declared cut, physical work absent, is a snapshot of a moving system, and a few specific signals will tell you how it evolves before the headlines do. Watching them is more useful than watching the total layoff count, which mixes too many causes to mean much.

The signals that matter most are the ones that would show the frontier moving rather than the volume rising:

• Declared share: whether the fraction of cuts where employers declare AI the reason rises from about one in nine today (26 of 232).
• Named roles: whether employers start saying which jobs they cut; today 83 of the 176 cuts where they named AI do not.
• Engineering scale: whether declared engineering cuts, now mostly at small and mid-sized software firms, spread to the largest employers.
• Demand loss: whether more employers cite AI answers taking their traffic or customers, the pathway in Section 8.
• Reversals and rehires, like Commonwealth Bank's and Klarna's, which mark where AI substitution outran capability.
• Family spread: whether declared AI cuts start appearing in work that is insulated today, such as frontline jobs.

A rise in the declared share would be the single most important development, because it would mean companies are moving from AI as a convenient narrative toward AI as a stated, defensible cause, which is the transition from AI-washing to genuine substitution. A spread of declared cuts into currently insulated families would mean the exposure frontier is advancing, and the task-shape framework predicts where it would go next: toward work that is becoming more digital and more routine even if it was not before. We track all of these continuously, and the live AI Layoffs Index combines the register with the official labor data in one reading. For the questions this article could not fully answer, our FAQ covers the recurring ones directly.

14. The bottom line

The question "which jobs is AI actually replacing" has a real answer, and it is more specific and more reassuring than the headlines, while also being more actionable. The evidence, drawn from 232 source-cited cuts and corroborated by independent exposure and usage research, says this. In almost half of the cuts where employers name AI, they do not say which jobs went, which is itself the most common answer. Where they do, they name software engineering most often, mostly as one reason among several, and the engineering cuts they declare outright come mostly from software and internet companies describing AI doing more of the work. Layers of management are the next most common thing an AI-linked cut removes. The largest cut employers declared AI-related fell on customer service, where the AI replacement is a deployable product and where the reversals happen too. Corporate functions carry nearly as many roles in the declared cuts as customer service, back-office processing is thinned quietly across many companies, and writing and editorial jobs are lost mainly as AI answers take away the audience. Physical, in-person, and high-trust work does not appear. That pattern is not a coincidence of our sample; it is the shape of the technology itself, which operates on information through a screen and nowhere else.

Two conclusions follow, and they point in opposite emotional directions on purpose. The sobering one is that if your work is mostly routine information processing on a screen, you are closest to the front line, and the time to shift toward judgment, exceptions, and directing AI is now rather than later. The reassuring one is that the honest number is far smaller than the scary one, most "AI layoffs" are partly something else wearing an AI label, and the unit of disruption is the task rather than the job, so most roles change rather than vanish. Both are true at once, and holding both is the difference between panic and preparation.

The deeper point is about how to think about this at all. You do not have to accept anyone's prediction, including ours, because the events are public and the sources are linked. The AI Layoffs Index exists so that the question "is AI taking jobs, and which ones" can be answered with a number you can audit rather than a narrative you have to trust, and everything behind it, the register, the company breakdowns, the methodology, and the open data, is there to be checked. In a debate this loaded with incentive to exaggerate in both directions, an auditable answer is the only kind worth having.

This guide reflects the AI Layoffs Index register as checked for new events through October 3, 2026 (241 events after the corrections of October 10, of which 232 are cuts), with job families re-checked on October 8, 2026. The register updates as new AI-linked reductions are verified, and the live counts by job family are on What jobs will AI replace?. Every event cited here links to its source; open them and check.

Earlier versions of this guide

• October 10, 2026. Figures recomputed on the corrected register (the snapshot of October 3 now holds 241 events: 232 cuts and 9 employer statements). CrowdStrike's May 2025 cut (about 500 roles) and Chegg's May 2025 (248) and October 2025 (388) cuts were regraded from declared AI-related to AI as one reason among several: CrowdStrike's memo also names realigning the business and simplifying execution, and Chegg's releases also name reduced Google traffic, macroeconomic pressure and a lower cost structure. UPS's April 2025 plan (about 20,000 operational jobs) and January 2026 plan (up to 30,000) were removed from the register: no credible source we could find presents either cut as linked to AI, and UPS tied both to the decline in its Amazon volume. Klarna's February 2024 announcement became an employer statement rather than a cut, and Tata Consultancy Services' July 2025 cut had moved on October 9 from linked only by others to AI as one reason among several. Declared cuts went from 29 to 26 (one in nine, not one in eight), cuts naming AI among other reasons from 145 to 150, cuts linked only by others from 61 to 56, and cuts where the employer named AI from 174 to 176 (83 of them naming no roles). Declared engineering cuts went from 8 to 6 and their counted roles from 1,234 to 486; the counted roles across all declared cuts from 5,705 to 4,569; corporate functions' share of them from 2,538 (the most of any family) to 1,790, just below customer service's 1,795. Management layers appear in 27 cuts that name roles, not 26, after the TCS regrade. The section on frontline work no longer describes UPS's cuts as AI layoffs.
• October 8, 2026. Rewritten on the register checked through October 3, 2026 (235 cuts and 8 employer statements, up from about 110 events). Before rewriting, the job families of all 119 cuts where the employer named AI and its roles were recorded were re-checked against their sources under one rule (a family only for roles the employer, or named reporting, says were cut); 68 changed, 27 of them to "not specified," because earlier labels had read the areas where a company applies AI, its general description of itself, or internal support functions as roles cut. The counts now match the live page. As a result: software engineering moved from last to first among the cuts employers declared AI-related; customer service is no longer tied for first; management layers became the second most named family; frontline work no longer appears among cuts where the employer named AI; and a section on lost demand (section 8) was added. Cases that changed: Cloudflare's and Paycom's cuts were not customer-service cuts by their own accounts; CrowdStrike's roles come from its quarterly filing, not its memo; Wix's chief executive gave the strong shekel as the first reason and named no development or design roles; Mews and BigID did not say which roles went; Teleperformance named only management layers; Telus's 2024 reduction largely affected management professionals; Allianz Partners gave 1,500 to 1,800 jobs, and the register counts 1,500; Amazon's chief executive said its October 2025 cut was not AI-driven, and Microsoft said its July 2026 cuts were not being replaced by AI; BT's 2023 figure of about 10,000 includes contractors, so the register does not count it, and the BT and UKG examples were dropped. Quotes were re-sourced to the documents that contain them (ClickUp's post, Atlassian's filed post, Jassy's call remark, the Washington Post memo as reported, Chegg's own releases), and Coinbase, now recorded as a management-layers cut, left the engineering section. Challenger's figures were updated through September 2026, Microsoft Research's usage study was added, the "one in five" declared share became one in eight, a chart of illustrative exposure scores (not measured data) was removed, and the claim that searches for AI-proof jobs run to tens of thousands a month was corrected to thousands.
• October 5, 2026. A note at the top recorded that the register had grown from 111 to 243 events and that software engineering's place among declared cuts had changed.
• October 1, 2026. A re-check of the register against its sources regraded PayPal's 2026 restructuring and Duolingo's 2024 contractor cut from explicit to mixed, stopped counting Salesforce's support reduction as 4,000 layoffs (Salesforce said it came from not refilling roles and moving people to other teams), and recorded that Klarna's 700 is an estimate of the work its AI assistant did, with Klarna saying it made no cuts because of it.
• September 30, 2026. An earlier version said customer service appeared in 13 of the 23 events where the employer declared AI the reason. Checking each label against its event's claim and sources gave 8, and the all-events chart had overcounted customer service too.