Klarna AI Customer Service
The CEO who fired 700 agents, then quietly hired humans back
What They Said
In February 2024, Klarna issued one of the most quoted corporate AI announcements of the year. The buy-now-pay-later company said its OpenAI-powered chatbot was handling two-thirds of customer service conversations — work previously done by 700 full-time agents. The bot, Klarna claimed, resolved issues in under two minutes versus eleven for humans, drove a 25 percent reduction in repeat inquiries, and was on track to add $40M in profit in 2024.
CEO Sebastian Siemiatkowski went further. He said Klarna had stopped hiring entirely, told staff that AI could already do most of his employees’ jobs, and pitched the company’s headcount freeze as a model for what every services business should be doing. Tech press treated the announcement as proof that the AI productivity dividend had arrived.
What Actually Happened
By May 2025, Siemiatkowski reversed himself in public. Speaking to Bloomberg, he conceded that the cost-cutting strategy had gone too far and that Klarna had “ended up with lower quality” customer service. The company announced it was reinstating human agents and piloting a model where remote workers — initially students and rural residents — could log in on demand to handle customer conversations.
The reversal followed months of customer complaints that the bot resolved simple cases well but handled disputes, fraud claims, and account escalations badly. Internal data showed customers in complex financial situations were being routed in circles. For a company whose product is consumer credit, the inability to handle escalations wasn’t a UX problem. It was a regulatory and trust problem.
The headcount freeze also fractured. Klarna had cut its workforce from roughly 5,000 to 3,500 between 2022 and 2024, attributing much of the reduction to AI. By 2025, Siemiatkowski was acknowledging that “investing in the quality of human support is the way of the future” and that Klarna would always offer customers the option to reach a human.
The strategic narrative had been built on the idea that AI was a substitute for support staff. The walk-back made it clear it was a complement, and an unreliable one in the highest-stakes conversations.
The Root Cause
Klarna optimized for the average ticket, not the dangerous one. The bot performed well on the broad middle of customer inquiries — order status, return windows, payment dates — and that was the data Siemiatkowski cited publicly. But customer service value isn’t distributed evenly. The 5 percent of conversations involving fraud, hardship, or disputed charges drive almost all of the regulatory risk and churn. The bot was weakest exactly where the stakes were highest.
The second failure was a CEO who made the strategy too public to course-correct quietly. Once Siemiatkowski had told the world AI could replace agents, every quality issue became a referendum on his judgment, not a tunable product problem.
The Pattern to Watch For
When a vendor or executive cites aggregate AI performance metrics — “handles 70 percent of cases,” “two-minute resolution time” — ask for the distribution. Specifically: what is the model’s accuracy on the highest-stakes 5 percent of cases, and what is the cost of getting those wrong? Averages hide the conversations that determine whether you keep your customers and your license.
What You Should Steal
Segment your customer interactions by stakes before you deploy AI on any of them. Tier 1 (informational) is fair game for full automation. Tier 2 (transactional) needs human review on edge cases. Tier 3 (disputes, hardship, fraud) needs a human in the seat with AI assist, not the reverse. Klarna’s reversal cost it credibility because it deployed the same automation logic across all three tiers. Stratify the work, then automate the layer where the stakes match the model’s actual reliability.