Klarna. The AI business case wasn’t wrong. It was incomplete.

Klarna’s AI business case wasn’t wrong. It was incomplete, and the public record shows what it didn’t have to explain.

I’ve sat in enough AI steering meetings to know which number wins the budget. It’s the one that goes down. Klarna is the clearest public example I know of what happens next, and almost everyone tells it wrong. This note is the version you can repeat in a board meeting, built from what Klarna filed and said, with the gaps marked.

1. The annualised savings estimate rose, and the realised figure matched the projection.

   About $40m projected at launch in February 2024. About $39m delivered in 2024, per Klarna’s SEC correspondence, a realised figure rather than a projection. $58m annualised in the Q3 2025 investor presentation, $60m on the call. Satisfaction on par with human agents throughout.

2. Cost per service transaction fell 40% in two years,

   from $0.32 in Q1 2023 to $0.19 in Q1 2025. The unit got cheaper.

3. The service bill rose year on year in every filed quarter this note compares.

   Customer service and operations cost $53m in Q3 2025 against $44m, $54m in Q1 2026 against $51m, and $58m in Q2 2026 against $51m. By mid-2026 it was growing slower than revenue, 14% against 27%, but it was growing.

4. Klarna added a premium human tier, and told its regulator why.

   In May 2025 the CEO said cost had become too dominant a factor. Coverage read it as a retreat from AI. In July 2025 Klarna told the SEC the remarks concerned a pilot of highly skilled freelance consultants to raise the quality of human service, did not relate to AI quality, and were misreported. It calls the model dual-track.

Klarna’s internal business case isn’t public, so this note judges the case as Klarna presented it. That public case priced the average contact, and the average got cheaper, exactly as promised. What it didn’t have to show was the quality Klarna would stand behind for the customers who aren’t average, the human path to deliver it, and the fact that a bill can shrink per unit and grow in total at the same time, with growth alone enough to produce that combination. None of that invalidates the saving. It changes what a complete AI business case has to carry. The unit has to be the cost of resolving the customer’s problem to the quality the business will stand behind, and cost per contact doesn’t measure that.

An outside-in analysis of public filings, investor materials and reported statements, current to Klarna’s Q2 2026 results of 18 August 2026. No access to Klarna, no inside information, and no estimates of figures Klarna hasn’t disclosed. The sources section at the end lists every source and every gap.

Four things the public 2024 case didn’t have to show. All four matter now.

The record stops here and my reading starts. Klarna’s internal business case isn’t public and Klarna hasn’t explained its 2025 decision beyond what it told the SEC, so what follows is what I’ve seen happen in AI programs I’ve worked on, applied to the facts in the exhibits above. Treat it as the argument of a practitioner, not a finding about Klarna.

1. The quality the business would stand behind

   Klarna’s case said satisfaction on par with human agents, and the SEC letter says it still does. On par on average. Every program I’ve seen that was measured on the average eventually met the customer the average doesn’t describe, and discovered nobody had written down what the business owed that customer. If the standard isn’t set before automation, cost sets it afterwards. That’s the most charitable reading of what the CEO meant in May 2025, and it’s also the one Klarna gave its regulator.

   In the record: satisfaction reported in aggregate only. No segment-level figure has been published.

2. The human path as a designed tier, not a fallback

   In 2024 the human path was the thing being reduced. By mid-2025 it was a pilot of skilled freelance consultants with its own staffing model. A fallback is a sentence in a policy that says customers can still reach a person. A tier has a staffing model, a quality target, and a cost line that someone owns. Adding it later can be more expensive than designing it in, particularly if the people who knew the complex cases have already left and have to be hired back at a premium.

   In the record: the pilot described to Bloomberg and the SEC. Its cost has not been disclosed.

3. The tail, and what it does to the humans left holding it

   When the AI takes the easy two-thirds, the remaining third is likely to be harder. Average handling time for the humans can rise, the error rate on what’s left can rise, and the people doing it may need more skill than the ones who did the full mix. If so, cost per contact falls across the whole base while rising inside the human tier. None of that is inevitable, and all of it should be tested before go-live. A case that prices human minutes at the old blended rate is betting that it won’t happen.

   In the record: Klarna’s move to “highly skilled” consultants is consistent with this. Klarna has not said so.

4. The growth effect, and the two numbers it forces apart

   A unit cost can fall by two-fifths while the total bill rises year on year, if the business grows fast enough. Klarna’s did, 32% in active consumers in the year to Q3 2025. Nothing about that is a failure. It’s arithmetic, and growth on its own is enough to produce it. But a business case that reports only the unit will look like a success in month one and a mystery by month nine, when the CFO asks why the line went up. Report both, side by side, from the first quarter.

   In the record: $0.32 to $0.19 per transaction. $44m to $53m, $51m to $54m, $51m to $58m, year on year by quarter. Both Klarna’s.

None of these four needs a new kind of data. They need a place in the case. The next section is what that looks like.

Five things I’d want in the room before the next AI business case gets approved.

These are mine, not Klarna’s. They come from the pattern in the sections above and from the steering meetings I’ve sat in where the first number won.

1. The number that goes down will win the budget. Make sure it isn’t the only number in the room.

   Klarna projected about $40 million, reported roughly $39 million delivered in 2024, and later raised its annualised estimate to $58 million. The saving was also the only figure most people heard. The cost per transaction, the quarterly bill and the quality standard were all available by 2025, and almost nobody put them next to each other. Insist on the second number at the first meeting, when it’s cheap to ask for.

2. Decide the quality you’ll stand behind before you decide what to automate.

   Not the average. The standard for the customer the average doesn’t describe, written so it can be failed and measured by segment. If it isn’t written down before go-live, cost will write it afterwards, and the CEO will end up explaining the difference in an interview.

3. Price the human path on day one, at the skill level the tail needs.

   The people left handling the hard third are likely to need more skill than the people who handled the whole mix, and to cost more per hour. A case that prices them at the old blended rate is betting that they won’t. Klarna’s “highly skilled freelance consultants” is what that correction looks like when it arrives late.

4. Report cost per unit and total cost side by side, every quarter, from the first.

   They answer different questions and they can move in different directions as the business grows. Klarna’s unit cost fell 40% while its service bill rose year on year in every filed quarter compared in this note. Both were true. A board that only sees one of them will think the other is a failure.

5. When the story turns, go to the filing, not the headline.

   The version of Klarna most people repeat comes from one interview and one article. The version Klarna is accountable for is in a letter to the SEC two months later, and it says something different. Companies tell regulators things they don’t tell journalists. Read what they told the regulator.

Vijayan Seenisamy is the author of The Pilot Trap and The AI Delivery Manager Blueprint, and the creator of the AI Role Operating Framework (AI ROF™). The series lives at AI Economics.

← All case notes