How Klarna went all-in on AI — and learned where the limits are

A fintech under pressure makes a very public bet

Klarna, the Swedish buy-now-pay-later giant, went into its AI push under real cost pressure ahead of a planned U.S. listing. Its CEO, Sebastian Siemiatkowski, became one of the loudest enterprise voices arguing that generative AI could do most knowledge work — and he picked customer service as the flagship proof point.

The pitch was simple and bold: if AI could handle the bulk of support, Klarna could run leaner, faster, and cheaper. For a while, the numbers seemed to back him up completely.

An AI assistant that did the work of hundreds

In February 2024, Klarna and OpenAI announced an assistant that, in its first month, handled 2.3 million conversations — about two-thirds of Klarna's customer-service chats — across more than 20 markets and 35+ languages. Klarna estimated it was doing the equivalent workload of around 700 full-time agents (a workload comparison, not 700 people let go), resolving issues in roughly two minutes versus eleven, with about a quarter fewer repeat inquiries. It projected a $40 million profit improvement for the year.

Around the same time, Klarna effectively froze hiring for roughly a year and let its headcount shrink through attrition. The story became a global headline and a rallying cry for “AI-first.”

This AI breakthrough in customer interaction means superior experiences for our customers at better prices, more interesting challenges for our employees, and better returns for our investors.

Sebastian Siemiatkowski, CEO, Klarna — company press release, February 2024

Then Klarna started hiring people back

By 2025, the tone had changed. Klarna began re-recruiting human agents — in a more flexible, on-demand model — after concluding that pushing automation primarily to cut cost had let service quality slip. The company that had been the face of replacing people with AI was now, very publicly, putting some of them back.

It wasn't a repudiation of AI. The automation gains were real and stayed. It was an admission that the original framing — optimize for cost, automate as much as possible — had been the wrong objective function.

Automation is real. Cost-only automation isn't a strategy.

Klarna's arc lands on a distinction every leadership team is now wrestling with. AI can absorb enormous volume — that part was never in doubt. But deciding what AI should and shouldn't do, and keeping humans on the moments that actually shape trust, turns out to matter more than maximizing the automation rate.

The durable model that emerged is human-in-the-loop: let AI carry the volume, let people own the high-stakes and high-empathy moments, and measure quality alongside cost. Adoption, done well, is not “replace everyone.” It's a deliberate division of labor — which is precisely the harder, more valuable work.

The cautionary tale every board now cites

Klarna has become the reference case for evaluating AI ROI honestly: dramatic early automation wins, followed by a quieter rebalancing once quality and customer experience were weighed properly. Expect more companies to skip straight to the hybrid model rather than learning the lesson the expensive way.

The takeaway isn't “AI doesn't work.” It's that the goal was never the lowest headcount — it was the best outcome per customer. Cost is one input, not the scoreboard.