Klarna's support assistant offers two useful exercises for an AI PM: audit the automation metric, then audit the story being told about it. Public volume figures do not establish segment-level quality. A disputed account of a strategy change does not establish its cause.
What was reported
In February 2024, Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats. The company said the assistant was doing work equivalent to 700 full-time agents, that average resolution time had fallen from 11 minutes to under 2, and that it expected a $40 million improvement in profit. All of these are company-reported figures from a launch announcement.
The later quality narrative is disputed. In its 16 July 2025 response to the SEC, Klarna said media coverage had misinterpreted comments about a pilot for human customer service as criticism of AI quality. The company described continued investment in both AI and human support. That is the company's account; it does not independently establish how either service performed.
The newest figures come from Klarna's annual report for 2025, filed with the SEC on 26 February 2026. It says the assistant has had 31 million conversations since launch and handled 80% of customer service chats in 2025, "with no drop in consumer satisfaction levels since its introduction", citing its own chat logs and internal satisfaction surveys. It repeats the estimate of work equivalent to over 700 full-time agents and puts 2024 cost savings at $39 million. These are still company-reported, now in a regulatory filing; the source of each number is named, but none is independently audited. Sources checked on 6 November 2026.
Read the filing the way an interviewer would. Eighty per cent of chats is a larger share of a channel than two-thirds, but the denominator is chats, not contacts: it says nothing about customers who phoned or emailed instead. "No drop in satisfaction" compares survey scores over time, and survey response is itself selected. The useful exercise is to name the metric you would add before accepting the headline.
Keep those evidence limits in the interview answer. Do not infer that AI failed, or that it succeeded for every customer, from this record.
Why two-thirds is not the number that matters
"Handled two-thirds of chats" describes volume the assistant touched. It does not say how many of those customers had their problem solved, how many gave up, or how many came back through email or a phone call a day later. Resolution time has the same weakness. A conversation that ends in two minutes because the customer abandons it looks like a success on that chart.
None of this means the reported numbers were wrong. It means they answer a different question from the one a product owner needs answered: did the customer's problem get solved, at a cost the business accepts, without damaging the relationship?
The averages hide the segments
Support traffic is not one population. Order status, refund timelines and password resets are high-volume and well suited to automation. Disputed charges, hardship requests, fraud and anything involving a vulnerable customer are lower in volume and far more expensive to get wrong.
Consider a hypothetical queue with 600 routine requests and 300 complex ones. Automating 90% of the first group and 20% of the second handles 540 + 60 = 600 of 900 requests, or two-thirds overall. That average does not show whether the 60 complex cases were resolved correctly. Review those outcomes separately before expanding autonomy.
Read diagram description
A hypothetical support queue: 540 of 600 routine requests automated, 90 percent, and 60 of 300 complex requests automated, 20 percent, for 600 of 900 overall, two-thirds. The average does not show whether the 60 complex cases were resolved correctly, which decides whether to expand.
This example illustrates a measurement risk; it is not a reconstruction of Klarna's internal decisions. The PM owns the segment breakdown and the evidence required to expand each segment.
What to measure instead
A launch dashboard for an AI support assistant should lead with three things.
The first is resolution that holds. Track re-contact about the same issue across channels within a defined window. Use seven days as a starting hypothesis and adjust it to the task's expected completion time. Absence of re-contact alone cannot distinguish success from a customer giving up, so audit a sample and ask for explicit confirmation where practical.
The second is that rate and satisfaction broken out by category, with a floor for each. If disputed charges fall below the floor, those conversations route to people, whatever happens to the blended rate.
The third is effective cost per resolved conversation: model cost plus the cost of the cases that end with a person, divided by conversations that actually stayed resolved. This is the number that tells finance whether automation is paying for itself, and it falls naturally when quality improves rather than when deflection does.
The handoff is part of the product
The quieter lesson is about design. A support assistant that cannot hand over gracefully pushes the cost of its failures onto customers, who must find a way to reach a person and then repeat themselves. A good handoff detects the cases it should not handle (by category, by expressed frustration, by repeated failure to resolve), passes the full context so the customer is not asked again, and is measured: how many handoffs, how quickly answered, how often resolved after.
Under efficiency pressure, a team can be tempted to make the handoff harder to find, because every handoff counts against the automation rate. That trade looks good on the dashboard that leadership watches and bad on every other one.
How this comes up in interviews
Try this original practice prompt: "Our assistant now handles 70% of chats. Leadership wants 85%. What do you do?" Our rubric rewards defining "handles", splitting the 70% by category and re-contact, and checking where the remaining 30% sits. Recommend expansion only for segments that clear a quality floor; keep a staffed handoff for the rest.
If you use Klarna in that answer, cite the company's figures as company-reported and acknowledge the disputed interpretation of the later coverage. Then state which evidence you would need from your own queue before making a launch decision.
Questions about this topic
In February 2024 Klarna said its assistant handled 2.3 million conversations in its first month, two-thirds of its customer service chats, doing work it described as equivalent to 700 full-time agents, with resolution time falling from 11 minutes to under 2. These are company-reported figures.