Hype Index: 'Klarna Replaced 700 Agents With AI, Then Hired Humans Back'

TL;DR: No, Klarna's AI customer service bet did not fail, and it also never fired 700 people. Klarna's 2025 Form 20-F says the assistant handled 80% of all customer service chats in 2025, up from two-thirds in month one, while full-time headcount fell every year (4,352 to 3,422 to 2,831) and the filing still says it expects headcount "to continue to decrease in future periods." The May 2025 rehiring that made global headlines was, in Klarna's own written words, "just two new agents in a flexible, remote setup." We score the two-sided claim 80% noise / 20% signal.

Klarna is the most-cited case study in the AI-replaces-support argument, and it gets cited for both sides. Slide one: an AI did the work of 700 agents. Slide two: it went so badly they hired the humans back. Both circulate as settled fact.

What is the claim?

Two claims, stapled together:

Klarna replaced 700 customer service agents with an AI assistant, then reversed course and rehired humans after quality collapsed.

What did Klarna actually say, and when?

I pulled every Klarna primary I could open. The record, in order:

Date	Source	What Klarna said
2024-02-27	Press release	"It is doing the equivalent work of 700 full-time agents." 2.3m conversations in month one, two-thirds of chats, "estimated to drive a $40 million USD in profit improvement to Klarna in 2024."
2025-05-19	Q1'25 earnings release	Customer service cost per transaction $0.32 (Q1'23) to $0.25 to $0.19, "with no drop in consumer satisfaction." Period-end employees 5,276 to 4,075 to 3,225.
2025-09-08	F-1/A registration statement	AI handled 69% of chats in the twelve months to 30 June 2025. "Equivalent work of over 700 full-time agents (based on the average monthly reduction in chat and telephone conversations in 2024)" and "$39 million in cost savings in 2024."
2025-11-18	Q3'25 investor presentation	"28m annual conversations, solving 81% of customer service chats." "AI does the equivalent work of 853 full time agents." "Delivering $58m of annual cost savings."
2026-02-26	Form 20-F	80% of chats in 2025, 31 million conversations since launch, "over 850 full-time agents" and "approximately $59 million in cost savings" for 2025. Headcount 2,831.

Two details there matter more than the headline numbers. First, the 700 was never a headcount action. Both filings define it as an estimate derived from the drop in conversations handled by human agents, and those humans were mostly not Klarna staff: the 20-F still books "outsourced assistance" inside the customer service line. Second, the $40 million was a projection. The filed figure came in at $39 million, so Klarna hit its own forecast almost exactly, which is the least-quoted fact in the saga.

What is actually true?

The concession is real. On 8 May 2025 Siemiatkowski told Bloomberg, in remarks Fortune reported the next day, that "as cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality," and that "really investing in the quality of the human support is the way of the future for us." That is a CEO conceding he over-optimized for cost.

What he did about it is where the story falls apart. In the same Fortune piece, a Klarna spokesperson said "we're very much still AI-first," and Siemiatkowski put the size of the hiring at two new agents acting as freelance customer support. Two. A month later at London SXSW he told TechCrunch that headcount had gone from 5,500 to around 3,000, that "two things can be true at the same time," and that "offering human customer service is always going to be a VIP thing."

Then Klarna wrote it down. Forbes ran "Klarna Reverses On AI" on 18 May 2025, then appended the company's response: "Klarna is not reversing on AI... Our AI assistant now performs the work of over 800 full-time roles (not 700)... Klarna never eliminated human support. We still work with several thousand outsourced agents. The current pilot involves just two new agents in a flexible, remote setup, it's an addition, not a rehire or reversal."

The phrase Klarna used there, a "dual-track approach" combining scalable AI with high-quality human support, is now the exact wording in both its F-1 and its 20-F. A PR line became registration-statement language, a much stronger commitment than a press quote.

What is not true?

That anyone reversed anything. Every measurable line kept moving the same way after May 2025. AI share of chats: 66%, then 69%, then 81%, then 80% for the full year. Conversations since launch: 25.3 million by mid-2025, 31 million by year end. Headcount: down again to 2,831, with the filing still projecting further decline. Customer service and operations spend was $240m in 2023, $203m in 2024 and $207m in 2025, up 2% in the year of the supposed rehiring while revenue rose 25%.

Nor is the origin story right. Klarna's filings do credit AI for shrinking the company, but through a hiring freeze and "normal course employee attrition," not layoffs pinned to the assistant. Headcount was already falling before it launched, from 5,527 at the end of 2022 to 4,352 at the end of 2023. The 700-agent estimate and Klarna's separate job cuts are two numbers that got fused, which is the mechanism behind most "we're cutting jobs because of AI" announcements.

What did my count of 20 articles show?

On 15 August 2026 I took the non-Klarna results from four searches (the February 2024 press release, the rehiring, Q1 2025, and the 853 figure) and classified each against the filings above. Of 23 URLs, 20 opened: fastcompany.com returned 403, marketbeat.com 404, and flexos.work 301s to leadwithai.co, so I substituted the next results from the same searches.

Five of 20 accurately represented what Klarna said: hrgrapevine.com, forbes.com (Jack Kelly), fortune.com, techcrunch.com, zenml.io. Fortune is the only one of the twenty that reports the number two in its own voice.

Five overstated the success, mostly by converting work-equivalence into fired people in the headline while the body said otherwise (getperspective.ai, leadwithai.co, asisteclick.com, one Forbes column that stretched a one-month result to "over the past year", and a customerexperiencedive.com piece asserting Klarna "laid off workers in favor of going all in on AI").

Nine overstated the reversal: two twig.so pages, bigeye.com, entrepreneur.com, emarketer.com, usefini.com, two more customerexperiencedive.com stories, and the Forbes column that carries Klarna's denial inside it.

The twentieth, a Medium post, rested on a "22% CSAT decline" and a "$10 million" savings figure I cannot find in any Klarna document, in a year when three filings report no drop in satisfaction.

Untraceable numbers worth naming: the "$60 million" AI savings figure repeated by four of these articles appears in none of Klarna's published materials, which say $58m and then $59 million. Entrepreneur's "3,500 employees" in May 2025 conflicts with the 3,225 Klarna reported ten days later. And the count of roles outsourced to Foundever is 250 in one article and 750 in another, with no Klarna document behind either.

What would change my mind?

Every operational number here is Klarna's own, unaudited and self-serving: satisfaction "on par with human agents" rests on internal surveys, and a company mid-IPO had every reason to keep the story clean. If a regulator or a churn analysis showed satisfaction did fall, the reversal camp would gain its missing evidence. It has not appeared. Klarna's own 11-minute human baseline also became 12 minutes by the time it reached the filings, which tells you how firm these figures are.

What should you do instead?

When a company tells you AI does the work of N people, ask what N counts. At Klarna it counted conversation volume that stopped reaching outsourced agents, not badges collected. And when the reversal story arrives, ask for the hiring number. Two is a number. "Rehiring humans" is not. It is the same discipline that survives contact with every famous one-person AI company.

Verdict: 80% noise / 20% signal. The signal is a real CEO admission that cost was weighted too heavily and quality suffered. The noise is everything built on it in both directions, from a stat shaped like the 95% pilot-failure number: a narrow measurement quoted as proof of something enormous.