For most of 2024, Klarna was the case study every AI vendor pointed to. The company replaced roughly 700 support agents with an AI chatbot and claimed it was handling 2.3 million conversations per month. By mid-2025, CEO Sebastian Siemiatkowski publicly reversed course. The company was rehiring humans. Reporting on the deployment described the initial scale similarly: the system handled roughly two-thirds of all customer chats (around 2.3 million conversations) and the company projected a $40 million improvement in profits. Inside the industry, this was treated as proof of concept. Outside it, in actual customer inboxes, something different was happening.

Customer satisfaction had fallen, service quality had become inconsistent and complaints about robotic responses and unresolved issues had grown, and the company began quietly rehiring human agents. The reversal wasn't spun as a pivot. His admission was blunt: Klarna had prioritized efficiency and cost over quality, and it wasn't sustainable. That's a rare thing for a CEO to say on the record about a flagship AI initiative, and it's worth sitting with. The bot wasn't malfunctioning: it was doing exactly what it was built to do: resolve tickets fast and cheap. The problem was that "fast and cheap" and "actually helped the customer" turned out not to be the same thing.

This isn't an isolated data point. Nearly one in five consumers who have used AI for customer service saw no benefits from the experience, a failure rate almost four times higher than for AI use in general, according to the 2026 Consumer Experience Trends Report from Qualtrics. The same report found that consumers rank AI applications for customer service among the worst for convenience, time savings, and usefulness: only "building an AI assistant" scores lower. Zoom out further and the pattern gets uglier: organizations invested $47 billion in AI initiatives during the first half of 2025, yet 89% of that spend delivered minimal returns. Klarna isn't the exception here. It's the highest-profile example of a very common outcome.

What went wrong is not a mystery: it's a measurement problem. A chatbot that closes a conversation in ninety seconds looks great on a dashboard that tracks handle time and deflection rate. It looks terrible to a customer who never got their actual issue solved and had to escalate anyway. The most common implementation mistakes include overloading a single AI model with an entire customer knowledge base instead of building use-case-specific agents, skipping quality control between AI-generated responses and customers, launching at full volume without a gradual rollout, and treating AI as plug-and-play software that requires no ongoing attention. Klarna's own leadership made close to every one of those mistakes at scale, then had to unwind them in public.

The action item for this week isn't "don't use AI in support," it's "audit what you're measuring before you scale it." Pull your last 90 days of AI-handled support tickets and cross-reference deflection rate against actual resolution rate and repeat-contact rate for the same customer. If deflection is up but repeat contacts are also up, you have a Klarna problem hiding in your dashboard, and it will surface with your customers before it surfaces in your metrics. Fix the measurement before you cut another headcount line.

Sources