Klarna's AI-Only Support Experiment Just Got a Human Backstop
Klarna spent two years as the industry's flagship example of AI replacing customer service at scale. It cut roughly 700 human agents and let an AI system handle support on its own, a move the company touted publicly as proof that generative AI could run a real support operation without people. That story has changed.
Klarna is now piloting a hybrid model that brings human contractors back in, sourced through gig platforms rather than rehired as full-time staff. AI still handles the bulk of inquiries, roughly two-thirds of tickets, but the remaining third, the tickets that are complex or emotionally loaded, get routed to a person.
What actually broke
The reversal followed a measurable decline in customer satisfaction scores, concentrated specifically in the harder tickets: disputes, refund edge cases, anything where a customer was upset going in. Those are exactly the categories where an AI system optimized for speed and resolution volume struggles, not because it can't answer the question, but because the person on the other end needs to feel heard before they'll accept the answer. Klarna's original cost-savings claims from running the AI-only model have also been disputed since the initial rollout.
Where the ceiling actually is
The lesson isn't that AI support doesn't work. Two-thirds of Klarna's ticket volume is apparently still being handled by AI with no human involved, which is itself a significant share of a real support operation. The lesson is narrower and more useful: the ceiling for AI-only support sits somewhere below the emotionally sensitive and structurally complex tickets, and that ceiling didn't move much even at a company with Klarna's resources and motivation to prove otherwise.
What this means for anyone evaluating AI support tools
Klarna is a company with every incentive to make an all-AI model work: it's a fintech with a direct interest in cost control, and the original move generated a lot of favorable press. If the model still needed a human safety net for a third of tickets, that's a data point worth weighing against any vendor pitch that promises full automation of a support queue. The practical question isn't whether AI can handle support, it clearly can handle a majority of it. The question is whether your ticket mix looks more like the easy two-thirds or the hard third, and whether you have a plan for routing between them before customer satisfaction tells you the hard way.