In February 2024, Klarna CEO Sebastian Siemiatkowski confidently announced: "Our AI chatbot is doing the work of 700 CS agents." Within a month, AI was handling 75% of all customer chats, managing 2.3 million conversations per month. Then, one year later, the same CEO says: "Cost was an excessively dominant evaluation criterion. The result was low quality, and that wasn’t sustainable."
What Is This?
Klarna is a Swedish fintech company known for its "Buy Now, Pay Later (BNPL)" service. $14.6 billion valuation, tens of millions of global users. Starting in 2022, they partnered with OpenAI and went all-in on deploying AI chatbots for customer service.
At first, it really seemed to be working. The CEO froze all new hiring for a year and cut headcount from 5,000 to 3,500 — a 22% reduction. He was convinced AI would handle everything.
But over time, problems started surfacing. Customers complained that AI responses were generic, repetitive, and completely unable to pick up on nuance. When it came to complex payment issues or emotional situations, AI was helpless. Customer satisfaction visibly dropped, and the AI chatbot ended up being little more than a router that passed people through to human agents instead of actually solving problems.
As things got serious, Klarna took drastic measures. Since they couldn’t hire people back fast enough, they reassigned software engineers and marketers to the call center. Specialized talent answering phones.
"From a brand perspective, from a company perspective, I think it’s really important to make it clear to customers that you can always talk to a human whenever you want."
What Changes?
Klarna’s case isn’t an isolated one. Duolingo walked a similar path. In January 2024, Duolingo declared AI-first and laid off 10% of its contract translators. CEO Luis von Ahn said, "We will no longer use contractors for work AI can handle." The result? Lessons became repetitive, explanations got shallow, and what was once fresh content started feeling mass-produced. The AI-first memo posted on LinkedIn received over 1,000 comments, most of them from people saying they’d delete the app. Von Ahn eventually admitted he "didn’t write the memo well" and walked things back.
This isn’t just a two-company problem. The data speaks for itself.
| Full AI Replacement (Klarna Before) | Human+AI Hybrid (Klarna After) | |
|---|---|---|
| Simple Inquiries | AI handles (fast) | AI handles (fast) — same |
| Complex Issues | AI tries → fails → customer churns | AI summarizes context → human resolves |
| Emotional Situations | Canned response → frustration explodes | Human empathizes → trust restored |
| Escalation | Customer re-explains from scratch | AI auto-transfers chat history, intent, and attempted solutions |
| Cost | Short-term savings ✓ | Long-term customer retention + brand value preserved |
| Customer Satisfaction | 60% (AI alone) | 88% (human agent benchmark, Verizon) |
Gartner predicted in a February 2026 report: "50% of companies that cut CS staff because of AI will rehire humans by 2027." Klarna became the first company to prove that prediction right.
The Essentials: How to Get Started
If you want to innovate CS with AI without repeating Klarna’s mistakes, follow these 5 steps.
- Set AI’s role as "assist," not "replace"
Design AI to handle simple, repetitive inquiries (shipping status, password resets, etc.) and route complex ones to humans from the start. Klarna’s mistake was setting up AI not as "instead of humans" but "without humans." - Make "context handoff" mandatory for escalations
When a customer transitions from AI to a human, the chat history, inferred intent, and attempted solutions should transfer automatically. Making customers re-explain everything from scratch is a CX failure. - Always keep the "talk to a human" option visible
This is the key insight Siemiatkowski learned the hard way. Give customers the choice. Don’t hide it behind menus. 50% of consumers worry that AI will block them from reaching a human agent. - Track AI quality metrics separately from cost savings
The trap Klarna fell into was equating cost reduction with success. Track CSAT (Customer Satisfaction), FCR (First Contact Resolution), and escalation rates separately, and review them with the same weight as cost metrics. - Roll out gradually and keep a rollback plan
Cutting 700 people at once isn’t an experiment — it’s a gamble. Increase AI’s share in 10–20% increments while checking quality metrics at each stage. Maintain a structure where you can immediately bring humans back if problems arise.
Core Principle
The goal of AI adoption isn’t "reducing headcount" — it’s "letting people focus on what matters more." This lesson Klarna learned the expensive way applies to every company. If AI handles 80% of simple tasks, humans can fully focus on the remaining 20% of high-value interactions.
