The best-documented support rollout also produced the best-documented correction.
These are other companies' numbers, not ours. Each one is linked to its source, and each one comes with what happened afterwards, because that is where the useful part sits.
Klarna automated two thirds of support, then walked part of it back.
Klarna
In February 2024 Klarna put an OpenAI-built assistant in front of customer service. Within a month it was handling two thirds of all chats, and the company reported resolution times falling from 11 minutes to under 2.
conversations in the first month, which Klarna estimated at a $40M profit effect for 2024 and the workload of 700 agents.
In May 2025 Klarna started hiring customer service agents again. CEO Sebastian Siemiatkowski said the company had cut too far and that what you end up with is lower quality. Klarna kept the AI for high-volume routine work and put people back on the complex and emotional cases.
The correction is the part worth building for.
Automate the repeat, not the exception
Escalation is a feature, not a fallback
Measure the hard cases separately
What we do not promise
- That your numbers will look like Klarna's. They are a payments company with enormous, highly repetitive volume.
- That headcount goes down. In the one publicly documented case at this scale, it went back up.
- That the system handles complex cases. It should recognise them and hand them over quickly.