Do AI customer service agents actually reduce total support costs?

Klarna CEO Sebastian Siemiatkowski told a Q3 2025 earnings call that the company's AI customer service agent "can now do the work of more than 853 full-time agents," up from 700 at the start of the year, and that it "has saved the company $60 million." Yet the excerpt reports that "customer service and operations cost the company $50 million in the third quarter, up from $42 million a year ago" — the savings claim and the cost trend point in opposite directions, and both figures come from the company itself, not an independent audit.

Forrester analyst Kate Leggett supplies the mechanism: AI handles "highly repetitive and fairly simple" questions well and can "deflect contacts" or speed up human reps, but Klarna "overpivoted" in 2024 by laying off experienced service reps with institutional knowledge in the run-up to its IPO. That produced the generic, nuance-blind answers customers complained about. Klarna reversed course in May 2025, rehiring for an "Uber type" customer service workforce while still maintaining — in its Q1 and Q3 2025 reports — that there was "no drop in consumer satisfaction" and that AI-assistant satisfaction is "on-par" with a human agent, alongside a company-wide NPS of 73. Leggett also floats a non-technical explanation for the savings figure itself: "I wonder how much of this cost savings was having their optics well managed for them going public," noting the overpivot's timing matched the IPO.

This sits against Does AI assistance help less experienced workers most?, which comes from an independently measured, staggered-rollout study of actual resolution rates rather than a company's own earnings-call figures — a sharper instance of the same gap that shows up in Do LinkedIn's AI hiring tools actually produce better hires?, where a vendor's efficiency metrics likewise leave quality and total cost unverified. Klarna's case adds a concrete twist: the self-reported efficiency gain ($60 million, 853 employees) and the self-reported cost trend (up $8 million year over year) are both disclosed by the same company in the same earnings materials, and still don't reconcile on their face.

The excerpt doesn't establish how much of the "work of 853 employees" figure reflects full automation versus deflection or assistance to still-employed human reps, nor does it reconcile the $60 million savings claim with the $8 million rise in total customer service and operations cost — both could be true if costs would have risen more without the AI agent, but the excerpt gives no counterfactual. The implication, at the strength the evidence allows, is that a company's self-reported AI productivity figures, especially one disclosed around an IPO, need an independent cost or quality baseline before they can be read as net savings rather than optics.