Klarna saved $4M replacing 700 agents, then rehired. Which number arrived first, and which one would have warned you?

[xyzs996]

Short answer: the savings and the retraction are both real, they were published months apart, and the figure that would have predicted the second one was already sitting in the first announcement. What follows is every figure this repo has collected on the case, each kept with the sentence it was published in and the date it was read.

The figures that exist, and where each came from

• $40 million — "Klarna's AI customer service experiment, which replaced 700 human agents, initially saved $40 million in a year, but the quality of service dropped." (2026-08-10)
• $4 million — "Klarna reported $4 million a year in savings and a 99.96 percent conversation engagement rate, the kind of pair of numbers that ends an argument in a board meeting." (2026-08-16)
• 68% — "Six months after Klarna's rollout, customer satisfaction scores fell to 68%." (2026-08-16)
• 22% — "Klarna's hybrid model steered customers to human agents automatically when the AI detected emotional distress or a policy exception." (2026-08-16)
• 10% — "Klarna's AI now handles the simple issues and a human agent audits 10% of escalated cases weekly, which holds quality while cutting cost." (2026-08-16)
• 60% — "What the ladder bought Klarna was the 60% cost reduction without the service collapse, and the reason it held is that each rung was allowed to fail cheaply before the next one got built." (2026-08-16)

The ordering is the whole story. The $4 million showed up on the books months before the CSAT number showed up to argue with it. Cost savings are measured weekly and land in the quarter you make the cut; satisfaction damage is measured in churn and lands two quarters later. Any rollout judged on the first clock and not the second will look like a success at exactly the moment it is going wrong.

What none of these are. This is one company, one quarter, and Klarna had reasons to publish both the $4 million and the retraction that have nothing to do with whether the AI worked. The 99.96% engagement rate measures conversations the bot stayed in, not problems it solved — the two are opposites when a customer cannot get out.

Every figure quoted next to the name, on one page: https://xyzs996.github.io/llm-api-pricing/providers/klarna.html

Write-ups with the full context:

• https://xyzs996.github.io/llm-api-pricing/articles/when-ai-customer-service-backfired-klarna-s-case-and-the.html
• https://xyzs996.github.io/llm-api-pricing/articles/the-klarna-lesson-why-ai-implementation-needs-a-staircase.html

The one thing worth replying with: if you have replaced part of a support or ops workflow with an AI, what was the first number that told you it was going wrong — and how many weeks after the savings figure did it arrive? A one-line reply with a metric and a lag is worth more than any row above, and it goes into the table with your wording kept.