Everybody has an opinion on AI and jobs. Most of those opinions are either catastrophising or minimising, and almost none of them are specific. Let me be specific.
“Replaced” here means one thing: a company deployed a tool and reduced headcount in a specific role as a direct result. Not “augmented the role.” Not “transformed the workflow.” People out, tool running, cost saved. That happened. Here’s where.
The 7 That Actually Did It
1. Customer Support Tier-1 (Intercom Fin, Zendesk AI, Freshdesk Freddy)
This is the largest and clearest displacement of the last two years. Tier-1 support — order status, account access, basic troubleshooting, the stuff that follows a script — is being replaced at scale. Not supplemented. Replaced.
Klarna announced in early 2024 that their AI assistant was doing the work of 700 human agents. That number is now higher. Dozens of companies followed the same playbook without the press release, because announcing that you replaced 700 people with software isn’t great PR even if it’s good business.
What survived: escalation handlers, complex case resolution, customer success for high-value accounts. The tier-1 support role as it existed in 2022 is effectively gone at any company that’s done the math.
2. Performance Marketing Copywriters (Jasper, Copy.ai, Anyword)
The junior copywriter producing A/B test variants for paid campaigns is largely gone at companies running volume performance marketing. Not “being helped by AI.” Gone.
The economics aren’t subtle: AI produces 50 headline variants in 30 seconds at zero marginal cost. A human produces 5 in an hour at €30–50. At 10 simultaneous campaigns, the math stopped working for humans somewhere around mid-2024 and nobody’s pretended otherwise since.
Senior copywriters with actual brand strategy? Still employed. Junior variant-generation roles? That job doesn’t exist anymore at companies that figured this out, which is most of them by now.
3. Data Entry and Document Processing (UiPath, Rossum, Microsoft Copilot)
Accounts payable clerks, invoice processors, data migration specialists — the combination of robotic process automation and modern AI document understanding made entire AP departments redundant at companies with high processing volumes. Rossum specifically handles invoice extraction at a level that made human review a compliance formality rather than a value-add. This has been building since 2022. The acceleration in 2025 was significant.
4. Junior QA and Entry-Level Code (GitHub Copilot, Cursor, Devin)
Manual test case execution and boilerplate code generation — the things junior developers and QA engineers spent most of their time on — are now automated. This hasn’t eliminated software engineering. It has significantly compressed the junior layer that used to serve as the profession’s entry point.
If you’re starting a software career in 2026, the path looks different than it did in 2022. The volume of entry-level work that justified entry-level hiring has shrunk. That’s worth being honest about.
5. Content Moderation at Scale (Various, mostly proprietary)
Social platforms and marketplaces have quietly cut large moderation teams through AI-primary moderation systems. This happened without press releases because “we replaced hundreds of human moderators with software” isn’t a headline anyone wants to publish. The headcount numbers are there if you look for them.
6. Financial Report Summarisation (Bloomberg AI, proprietary)
The analyst role that distilled 80-page 10-Ks into two-page summaries doesn’t exist at the scale it once did. AI summarisation pipelines handle the standardised work. What survived: senior analysts providing investment thesis and interpretation. The commodity summarisation layer is automated, and that was always where the headcount lived.
7. Recruitment Screening (HireVue, Greenhouse AI, Workday AI)
Initial CV screening and first-round scheduling, which previously required junior recruiter time, is now automated at most companies using modern ATS platforms. The entry-level recruiter whose job was managing this pipeline has largely been absorbed into higher-value activities or not replaced when they left.
The 5 That Didn’t (Despite the Headlines)
1. Radiologists
The “AI will replace radiologists” story has been running for eight years. It hasn’t happened. AI makes radiologists faster and more accurate — it is not replacing them. Liability, edge cases, and regulatory requirements mean a physician stays in the loop. This will change eventually. It hasn’t changed yet.
2. Lawyers
AI accelerated contract review and legal research dramatically. It did not replace lawyers. What it did was expose the value gap between average lawyers and excellent ones, which is a different problem. If you’re an average lawyer billing for hours that AI can now do in minutes, that’s your concern. If you’re a good lawyer, AI made you more competitive, not redundant.
3. Senior Graphic Designers
Midjourney ate the stock photography industry and displaced designers producing templated assets. Senior designers with genuine creative direction are arguably more in demand because the production quality ceiling has risen and clients expect more from real creative work. The junior pipeline is damaged. The senior profession is intact.
4. Teachers
AI tutoring tools are genuinely impressive and genuinely insufficient. Teaching is more social and contextual than the “AI solves problems” narrative accounts for. The timeline here is slower than predicted, for reasons that aren’t about the technology.
5. Software Architects
Despite every “AI developer” headline, senior software architecture — system design decisions, infrastructure trade-offs, long-term technical strategy — is not being replaced. These decisions are irreducibly complex, the cost of getting them wrong is catastrophic, and there’s no liability framework for offloading them to a model. Possibly the most durable role in the technology industry right now.
The Pattern
If you’re looking at the list above and trying to figure out where you stand, here’s the framework that actually holds: AI replaced roles defined by volume, repetition, and bounded complexity. It hasn’t replaced roles requiring judgment under uncertainty, human relationships, or real accountability for high-stakes decisions.
The question isn’t “can AI do this task?” — AI can produce something that looks like almost anything. The question is: what does getting this wrong cost, and who’s accountable when it does?
Where the answer is “a lot” and “a named human,” AI is a productivity tool. Where the answer is “not much” and “it doesn’t really matter who” — that’s where the headcount has been shrinking, and it’s not going to stop.