Klarna’s AI customer service assistant

In February 2024, the Swedish payments company Klarna announced that its AI assistant, built with OpenAI, had handled 2.3 million conversations in its first month, two-thirds of all its customer service chats.¹

The figures it published were striking. Klarna said the assistant was doing the equivalent work of 700 full-time customer service agents, achieved a customer satisfaction score similar to human agents, led to a 25% drop in repeat inquiries, and resolved errands in less than two minutes, compared with a previous average of eleven.¹ The announcement became one of the most cited examples of generative AI replacing human work at scale.

The change of course

In May 2025, Klarna’s chief executive, Sebastian Siemiatkowski, told Bloomberg that the company would recruit human customer service staff again, starting with a pilot of a flexible, remote working model.² Explaining the shift, he said that “cost unfortunately seems to have been a too predominant evaluation factor”, and that “really investing in the quality of the human support is the way of the future for us.” He added that “there will be always a human if you want.”²

This was not a reversal of AI. Klarna continued to use its assistant. It was a recognition that the efficiency figures from month one did not capture everything customers valued.

What the figures did and did not measure

Klarna’s launch metrics measured speed, volume and cost, and a satisfaction score. These are real benefits, and the speed gain in particular is substantial. However, they say little about the conversations that matter most to customers: complex, unusual or emotionally charged problems where the cost of a poor answer is high and where people want to feel heard.

A customer satisfaction score averaged across millions of simple queries can stay steady even while the experience of a minority with difficult problems gets worse. That minority is often where trust in a financial brand is won or lost.

Lessons for leaders

1. Judge AI on the outcomes customers value, not only on cost. In Klarna’s own words, cost became “a too predominant evaluation factor”. Decide what quality means before you launch, and measure it.
2. Look beneath the averages. Track how AI performs on difficult, high-stakes and vulnerable-customer cases separately from routine ones.
3. Keep a route to a human, and the people to staff it. Being able to reach a person is part of the service. Rebuilding human capacity after letting it go is slower and more expensive than keeping it.

Questions to ask in your organisation

1. When we report the success of AI in customer service, which measures do we use, and which customer experiences do they miss?
2. How does our AI perform for customers with complex or sensitive problems, compared with routine ones?
3. If we needed to increase human support quickly, could we, and how long would it take?