The job
Every day, millions of Klarna shoppers have questions. A payment didn't go through. A refund looks wrong. A late fee feels unfair. Each one of those messages used to land on a human agent's screen, and there were hundreds of people answering them around the clock. That's expensive, it's slow, and during busy seasons the lines pile up. Klarna wanted something that could answer the easy questions instantly, in any of their 35-plus languages, without hiring a small army every holiday season. So basically, they needed a machine that could do the boring-but-critical job of answering "where's my money?" two million times a month.
The agent
Klarna teamed up with OpenAI to build a custom AI assistant that plugs straight into their chat support. It's not a dumb FAQ bot picking from a script. The agent reads your question, pulls up your account details, and handles the actual task — refund lookups, payment issues, order status — across 23 markets. When it launched in February 2024, it took over two-thirds of all customer service chats. The company said it was doing the work of around 700 full-time support agents. That's not a typo. One system, roughly 700 people's worth of chat volume.
Results
The launch numbers were huge. In its first month the assistant handled 2.3 million conversations, and average resolution time fell from 11 minutes to under two minutes (Klarna's own announcement, Feb 2024, reported by TechGig May 2025). Repeat inquiries dropped about 25%, which suggests the AI was actually getting simple stuff right the first time. Klarna projected $40 million in annual savings. And here's the twist: in May 2025, CEO Sebastian Siemiatkowski told Bloomberg they'd gone too far, saying cost had been "a too predominant evaluation factor" and quality dropped. They started hiring human agents again on a flexible gig-style model, with AI still covering the simple two-thirds.
The catch
This is the famous cautionary tale, and you should actually pay attention to it. The AI crushed simple questions but fell apart on complex disputes, emotional customers, and edge cases — wrong answers about fees or contract terms aren't just annoying in fintech, they're compliance problems. Satisfaction on complicated cases dipped. Reports surfaced of engineers being pulled off their real jobs to cover support when the AI failed. The lesson isn't that support agents don't work. It's that "replace everyone" was the wrong goal. Klarna's current setup is the honest version: AI handles the easy volume, humans handle everything that matters.
Who it's for — who should skip
If you run support at scale — hundreds of thousands of chats a month with a heavy share of routine questions — this is the model to copy: Klarna's AI-first-plus-human-backup hybrid. Skip the "fire your whole team" fantasy, though. Any company with regulated answers, emotional customers, or complex multi-step problems should skip full replacement. Build your AI agent to do the simple two-thirds brilliantly and keep humans for the hard third. That combination is where the real money gets saved.