Klarna: The AI Customer-Service Round Trip
Featuring Sebastian Siemiatkowski
In early 2024, Klarna made headlines by announcing its OpenAI-powered assistant was handling two-thirds of customer service chats, doing in a month what had taken roughly 700 full-time agents. It read like a textbook AI cost-savings win, and the press treated it that way. Then, quietly, the metrics started to move in the wrong direction, and by 2025 CEO Sebastian Siemiatkowski was admitting the company had gone too far and reversing course. The round trip cost more than standing still.
This case is not about AI failing, because two-thirds of chats handled by a bot is a real achievement. It's about a threshold that's easy to miss and expensive to cross. For founders and operators racing to automate, it sharpens the discipline of measuring the right outcomes, because the cheapest number to track is rarely the one that determines whether the economics actually work.
Frequently asked questions
What happened with Klarna's AI customer service?
In early 2024 Klarna announced its OpenAI-powered assistant was handling two-thirds of customer service chats, doing in a month what had taken roughly 700 full-time agents. It read as a textbook AI cost-savings win. But the metrics later moved the wrong way, and by 2025 the company reversed course.
Did Klarna admit it went too far with AI automation?
Yes. By 2025, CEO Sebastian Siemiatkowski admitted the company had gone too far with AI customer service and reversed course. The round trip of automating, then walking it back, cost more than standing still. It became a cautionary tale about over-automating.
Why is the Klarna case not really about AI failing?
Two-thirds of chats handled by a bot is a genuine achievement, so the case is not about the technology failing. It is about a threshold that is easy to miss and expensive to cross when you optimize the wrong metric. The lesson is about measurement discipline, not AI capability.
What can operators learn from Klarna's AI round trip?
When racing to automate, measure the outcomes that actually determine whether the economics work, because the cheapest number to track is rarely the right one. The lesson is to find the threshold where more automation starts costing you. CaseBook turns this into a move you apply to your own company, with an AI coach that reads your answer.