What Klarna's AI Assistant Teaches Us About Scaling Support — and Where It Broke

Why This Case Study Matters

Klarna's AI assistant is one of the most cited — and most re-litigated — examples of enterprise generative AI in production. It's worth studying precisely because the story doesn't end at the impressive launch numbers. Read as a two-act case study, it tells you far more about deploying AI agents responsibly than either act alone.

Act One: The Launch (Feb 2024)

According to Klarna's own press release, the OpenAI-powered assistant handled 2.3 million conversations in its first month — two-thirds of all customer service chats — doing the equivalent work of 700 full-time agents. Resolution time dropped from 11 minutes to under 2. Customer satisfaction scores reportedly reached parity with human agents, and repeat inquiries fell 25%. Klarna estimated the assistant would drive $40M in profit improvement for 2024.

The architecture, as analyzed in Efi Pylarinou's Medium write-up on Klarna's "AI Native" framework, wasn't a single chatbot bolted onto a help desk — it was a systems-level rebuild: the assistant plugged into order data, refund logic, and policy documents so it could take action, not just answer FAQs.

Act Two: The Correction (2025)

This is the part most case studies skip. In May 2025, Klarna's CEO Sebastian Siemiatkowski publicly acknowledged the company had cut human support capacity too aggressively, and began rehiring agents after customers complained about generic answers on complex, nuanced cases. By Q3 2025, Klarna reported $60M in documented savings — a real number, but arrived at only after rebuilding some of the human-in-the-loop capacity it had removed.

The Research Takeaway

The lesson isn't "AI replaces support teams" — it's "AI absorbs tier-1 volume; humans move up the value chain." The failure mode wasn't the AI's accuracy on routine tickets (order status, refunds, policy questions) — it was removing the escalation path and institutional knowledge needed for the 10–15% of cases that are genuinely ambiguous.

For any team automating support, Klarna's experience argues for three design principles from day one:

• Keep a real escalation lane. Don't treat human agents as a cost to eliminate — treat them as the tier for the cases your confidence scoring flags as uncertain.
• Measure resolution quality, not just resolution speed. A 2-minute wrong answer is worse than an 11-minute right one; CSAT parity metrics can mask degradation on the long tail.
• Plan for reversibility. Klarna could rehire and recover because the AI layer was additive to their systems, not a replacement of the underlying support infrastructure.

This is the same architecture RudraAI uses for support automation projects: autonomous resolution for deterministic categories, explicit confidence-based routing to humans for everything else — see our Customer Support AI Agent case study for how that plays out at a smaller scale.