The Hook
In February 2024, Klarna published a press release that became one of the most-cited AI-displacement claims of the year: its OpenAI-powered assistant was handling the work of 700 customer service agents, resolving queries in under two minutes, and on track to deliver a significant profit impact. The story spread because Klarna named a number and stood behind it.
The Question
If the AI assistant was genuinely doing the work of 700 people — and the company was heading toward a U.S. IPO — why was Klarna, by mid-2025, quietly hiring human agents back?
The Paper Trail
Klarna’s February 2024 announcement stated the AI assistant was handling roughly two-thirds of customer service chats, achieving resolution in under two minutes versus eleven for human agents, and producing customer satisfaction scores comparable to human performance. The company projected an annual profit impact of approximately $40 million from the deployment.
By mid-2025, CEO Sebastian Siemiatkowski conceded the calculus had shifted. His own words: “We went too far,” and “We focused too much on cost. The result was lower quality.” The rehiring of human agents began in May 2025 ahead of the company’s U.S. IPO, which priced on the NYSE in September 2025. Klarna subsequently moderated its public claims about the deployment.
The Synthesis
The original number was real: the assistant did handle a large share of chats. The problem was a human decision to extend that deployment beyond the interactions it handled well — into the escalations, the unusual cases, the moments where a customer relationship is actually on the line. Quality degraded. Customers noticed. The company pulled back.
The walk-back didn’t get the same promotional treatment as the original announcement. That asymmetry is a disclosure choice, not a machine behavior.