Klarna's AI customer service: the 700-agent claim and the U-turn

If an AI transformation deck has one case study in it, it is usually Klarna. In February 2024 the company said its AI assistant was doing the work of 700 customer service agents. In May 2025 the headlines said Klarna was going back to humans. Both stories get quoted, often in the same meeting, to support opposite conclusions.

Klarna is now a listed company, so there is a better source than headlines: its filings with the US Securities and Exchange Commission. This post puts the announcement, the correction and the filings side by side.

The short answer: the volume claims held up and grew. By 2025 the assistant handled about 80% of customer service chats, and Klarna estimated it did the work of more than 850 agents. What did not hold was the idea that cost could be the main measure: the CEO said cost had been too dominant and quality suffered. Klarna did not abandon AI; it moved to what it calls a dual-track model, with AI for volume and a human always available. Every number here is Klarna's own estimate.

The announcement, February 2024

After one month live, Klarna's press release reported (Klarna, 27 February 2024):

• 2.3 million conversations, two thirds of all customer service chats, in 23 markets and over 35 languages;
• the equivalent work of 700 full-time agents;
• customer satisfaction on par with human agents, and 25% fewer repeat inquiries;
• errands resolved in under 2 minutes, against 11 minutes before;
• an estimated $40 million profit improvement in 2024, a projection at the time.

It also said customers could still choose a human. That sentence got much less attention than the 700.

The correction, 2025

In May 2025 CEO Sebastian Siemiatkowski told Bloomberg that cost had been "a too predominant evaluation factor" and that the result was lower quality; investing in the quality of human support was the way forward (Bloomberg, 8 May 2025). He described a pilot of remote agents who work flexible hours, an Uber type of setup in his words, and said customers must always know a human is available.

Two weeks later he clarified that the assistant was taking on more work, not less: about 1.3 million errands a month, the equivalent of about 800 people, and moving into harder second-level cases (CMSWire, 20 May 2025). In June he said human service would always be offered, as a VIP option (TechCrunch, June 2025). From August 2025 Klarna's IPO filings describe a dual-track approach: scalable AI combined with high-quality human support, with a human option for every customer.

What the filings show

The numbers moved as Klarna updated them, which is itself a lesson:

The company also got much smaller over the same period, which is the context the AI story sits in:

The filings add details the press release did not have. Human agents averaged 12 minutes per chat in 2024, not 11. Satisfaction parity is based on chat logs and internal surveys, with no published scores or sample sizes (prospectus). And in 2025, customer service and operations costs rose 2% while volumes grew 32% (20-F). That last figure is the strongest evidence that the assistant changed Klarna's economics, and it comes with the usual caveat: the company decides how to classify those costs.

What the case teaches

1. Cost is an outcome, not the goal. Klarna's own CEO named the mistake. Measure quality alongside cost from the start: resolution, repeat contacts, escalations and satisfaction on hard cases, not only averages.
2. Keep the human path, and make it visible. Klarna said so in its first press release and made it central after the correction.
3. "Agent equivalent" is an estimate. It moved from 700 to 800 to 700 to 853 in Klarna's own documents. Use it as a rough scale, never as a target or a headcount plan.
4. Volume is the easy part. Routine errands moved to AI quickly. The harder, second-level cases are where the gains now come from and where quality is tested.
5. Read the filings, not the headlines. Both the 2024 triumph and the 2025 U-turn were overstated. The SEC documents tell a steadier story.

For an agent in your own customer service, the measurement side is covered in Agentic AI from demo to production, including escalation and undo rates. For why company-level AI success claims need the same scrutiny, see "95% of AI pilots fail": what the MIT report actually measured; for the same gap between claimed and measured gains in software teams, Does AI make developers faster?; and to set a threshold before you start, A GraphRAG business case, done the evidence-based way.

How much weight this evidence deserves

Sources

• Klarna (27 February 2024). Klarna AI assistant handles two-thirds of customer service chats in its first month.
• Bloomberg (8 May 2025). Klarna turns from AI to real person customer service.
• CMSWire (20 May 2025). Klarna CEO: we're giving AI more customer service work, not less.
• TechCrunch (4 June 2025). Klarna CEO says company will use humans to offer VIP customer service.
• Klarna Group plc (March 2025). Form F-1 registration statement.
• Klarna Group plc (September 2025). Prospectus (Form 424B4).
• Klarna Group plc (November 2025). Q3 2025 investor presentation (Form 6-K).
• Klarna Group plc (February 2026). Annual report for 2025 (Form 20-F).