Klarna's AI did the work of 700 people, Harvey compressed two weeks of a lawyer's labor into a day, and a hospital's AI cut physician burnout by 13 percentage points. Yet a year later Klarna was hiring people back. Draw real corporate cases out as step-by-step workflows, and you can see exactly what an agent takes and where it stops.
"AI eliminates jobs" is far too blunt a sentence. Look closely at how it actually gets deployed, and an agent rarely swallows a person's whole job. Instead, of the several layers that make up that job, it absorbs one specific layer — the repetitive, standardized processing of information. The support rep's "check the order status," the lawyer's "first draft," the developer's "simple migration," the doctor's "charting" are all that kind of layer.
So the outcome splits by role. Where standardized processing is close to the whole job (frontline support, simple document work), headcount shrinks. Where standardized processing is only part of the job (strategy, disputes, in-person care), people move up to the layer above it. This piece draws four real cases as workflows, showing at a glance which stage hands off to the agent and which stage the human keeps. And what happens when you draw that boundary wrong.
In February 2024 the fintech Klarna switched on an OpenAI-based AI support agent worldwide. In the first month alone it handled 2.3 million chats — work the company said was equivalent to roughly 700 full-time agents. Resolution time dropped from 11 minutes to 2 minutes, repeat inquiries fell by 25%, and customer satisfaction matched human levels. The company projected a $40 million profit improvement in 2024 alone (against a build cost of $2-3 million). Total headcount fell from about 5,000 to 3,500 (mostly through natural attrition).
Seen as a workflow, the frontline-support role was relocated wholesale.
That's the half that gets quoted everywhere. The other half matters more. In May 2025, CEO Sebastian Siemiatkowski publicly admitted the company "cut too deep" on people and reopened hiring for premium support staff. Over six months customer satisfaction had slipped: on simple inquiries the AI matched humans, but on complex disputes and hardship cases its resolution quality was noticeably lower. In the end, layer ④ of the workflow above turned out to be thicker than expected.
What Klarna proved was not "AI replaces 700 people," but a more precise proposition: "the frontline-support layer gets replaced, but the layer above it still belongs to humans." — The lesson of Case 01
The legal AI Harvey took hold fast in law, where standardized document labor makes up a large share of the job. The asset manager Bridgewater saw more than 95% time savings on large-scale contract review, cutting vendor-contract review from an average of 2 days to 2 hours. The firm A&O Shearman rolled it out across the company to 4,000 people in 43 jurisdictions, saving 2-3 hours a week and trimming contract-review time by 30%. The most striking figure is on the litigation side. In one matter, deposition summaries and theme analysis that several junior lawyers would take two weeks to do were finished in under a day.
What's worth noting is the place of the junior lawyer. At one firm (Lynn Pinker), associates reported that instead of spending time on first drafts and bulk document review, they engage in case strategy earlier and more deeply — stress-testing arguments and preparing depositions. As layers ① and ② moved to the agent, people were pushed up into layers ③ and ④. The work didn't disappear; its center of gravity shifted upward.
Coding is the area agents fought over first, because grading the answer is automatic (tests, compilation). The payments company Stripe compressed the migration of a 50-million-line Ruby codebase from the months it was scheduled to take down to days (per Anthropic). The telecom TELUS adopted agentic coding tools internally and reported shipping engineering code 30% faster, saving a cumulative 500,000 hours and more.
The pattern Anthropic's agentic-coding report observed is two-sided. Time spent per task fell (automation), but output per person rose by far more (amplification). It means the same headcount produces more, which feeds both the fear that "juniors disappear" and the optimism that "the engineer is promoted from code author to the conductor of an agent legion." But strip out layer ④ — review and accountability — and the agent becomes a machine for mass-producing plausible errors at speed.
Outside the office, on the ground where people collide with each other directly, the picture differs. Medicine is the prime example. The AI ambient scribe in the exam room (Abridge and others) doesn't replace the doctor. It listens to the conversation the doctor has with the patient and automatically writes the chart and visit notes. It peels off only the most draining administrative layer of the doctor's work.
The numbers are clear. Abridge has signed contracts with more than 150 healthcare providers, and a study of 1,800 clinicians across 5 academic centers found savings of 16 min on documentation and 13 min on the electronic record per 8-hour clinical day. In a study of 263 physicians, burnout fell from 51.9% to 38.8% in 30 days, and at St. Luke's after-hours documentation dropped 35% while face time with patients rose 15%.
Put the four cases on the same table and the boundary between the layer the agent took and the layer people kept comes into sharp relief.
| Role | Layer the agent took | Layer the human kept | Measured |
|---|---|---|---|
| Customer support | Lookup · simple resolution · frontline reply | Disputes · fraud · hardship · empathy | 700 people's work · 11→2 min |
| Legal | Doc review · summary · first draft | Strategy · stress-testing · advocacy · accountability | 2 weeks → a day |
| Engineering | Migration · repetitive fixes | Design · review · merge · accountability | Months→days · +30% |
| Healthcare (field) | Chart · note writing (admin) | Diagnosis · exam · empathy · signature | Burnout 51.9→38.8% |
Sources: each company's announcements and studies (notes below). The common thread is sharp. The agent takes the "standardized, repetitive, information-processing" layer and leaves the "judgment, exception, relationship, accountability" layer to the human. Which roles shed headcount comes down to how large the front layer's share is in that role.
The most accurate name for today's agents is not "replacement" but "a capable but supervision-hungry, infinitely scalable legion of juniors." That legion devours the routine-labor layer fast and pushes people up into the layer above — judgment, relationships, accountability. The good news and the bad news come from the same fact. Roles where the front layer was most of the job lose headcount, and for anyone ready to climb to the upper layer, a lever appears that lets one person conduct the work of ten.
The practical lessons are three. First, drawing the boundary wrong is expensive — Klarna underestimated layer ④ and ended up calling people back. Second, the review layer is non-negotiable — Harvey, Stripe, and Abridge all leave the human's signature, merge, and verification for the end. Third, value moves upward — what disappears is the first draft, not the advocacy; the charting, not the diagnosis. The career strategy of the agent era is simple. What can you do on top of the layer the agent eats.