Klarna
An assistant said to do the work of 700 agents, then a reversal fifteen months later.
Klarna and OpenAI joint announcement
27 February 2024; figures company-reported
Bloomberg, interview with chief executive Sebastian Siemiatkowski
8 May 2025
In a February 2024 announcement made with OpenAI, Klarna said its AI assistant was handling the equivalent workload of 700 customer service agents, had managed 2.3 million conversations in its first month, and was expected to contribute around 40 million dollars in profit improvement that year. Those figures are company-reported and were never independently audited.
In May 2025 the chief executive told Bloomberg the company had focused too heavily on cost, that the result was lower quality, and that it was rehiring people. It is now the canonical enterprise cautionary tale, the Hollow Core's characteristic failure, and the Authority cell's failure case: the question was never whether the agent could handle volume, but who could notice quality falling and stop it.
The honest reading is not that the automation was wrong. A complicated-domain tool was pointed at a partly complex-domain problem, and nothing in the operating model was designed to catch that.
Which decisions may an agent make alone, and who is able to stop it?
For each workflow, what is the human to agent ratio, and who signs the quality baseline?
Who is qualified to overrule the machine, and how would you prove it?