Most operators found Klarna the same way: through a headline that was impossible to ignore.
In February 2024, the company announced that an assistant built with OpenAI had handled 2.3 million customer service conversations in its first month, cut resolution time sharply, performed work equivalent to roughly seven hundred full-time agents, and was expected to contribute about forty million dollars in profit improvement for the year. It was exactly the kind of claim that makes an executive team ask why they are not already doing this.
Then the story got more complicated. In 2025, Klarna’s own chief executive acknowledged that some AI-driven service had been lower quality than human support, and the company began adding human representatives back into the model, even as AI continued to handle the large majority of customer-service volume.
At that point the example split into two convenient narratives. One camp used the first headline as proof that AI replaces service organizations. The other used the second as proof that it does not work. Both readings are wrong, and the second is the more expensive mistake, because it lets a leadership team dismiss something that did work.
Read the Economics, Not the Announcement
Here is what actually happened, and it is the part worth carrying into your own business case.
The operation was growing. Volume was rising, and AI absorbed a large share of the increase. Every interaction the capability handled was one the company did not have to staff for. Turnover was not backfilled. The requisitions that a year of that growth would normally have produced were never opened.
Total customer service expense still rose, because there were far more interactions to serve. What fell was the cost of serving each one, and that decline offset enough of the volume to hold the total well below what the same growth would have cost a year earlier. Notice what that means for the financial statements. No line on the income statement went down. The line that rose, rose by less than it otherwise would have. And nothing in the accounts records the difference.
Most AI Value is Cost Avoidance, and That is Still Value
This is the single most common reason credible AI value disappears inside a company. Most of what gets counted as AI value is cost avoidance rather than expense reduction, and cost avoidance is frequently the value the business actually receives.
So are the levers that never touch the income statement at all. Response time. Decision speed. Exception recovery. Quality. A leadership team that only recognizes value when a cost line falls will systematically undercount its own results, then conclude that AI did not work and quietly reduce the investment.
The opposite error is just as costly. Cost avoidance is easy to claim and hard to disprove, which makes it the favorite refuge of a weak business case. Any team can assert that headcount would have grown. Very few establish the baseline that would make the claim testable.
The discipline that separates the two is unglamorous. Set the baseline before you deploy, not after. State the growth assumption explicitly and in writing. Name the requisitions or the backfills that will not be opened, and have the operating leader and finance agree in advance that those are the measure. Then check a year later against what was actually written down, not against what the outcome now suggests should have been written down.
Why the Reversal Matters More than The Headline
The part of the Klarna story most worth studying is the correction. A company deployed at scale, discovered that quality in some interactions did not hold, and adjusted the operating design while keeping the capability that was working.
That is not a failure. That is an organization running a feedback loop, which is more than most AI programs ever do. The alternative, which we see far more often, is a pilot that gets reported as a success, never gets measured against a baseline, and quietly stops mattering while nobody says anything.
It also demonstrates the boundary that every deployment eventually finds. Generative AI predicts and produces plausible outputs from patterns learned across large bodies of human-created information. It works best when the task has many acceptable answers and the result can be reviewed quickly. It becomes riskier when the task requires exact facts, missing context, high-stakes judgment, or an answer where a confident error is more dangerous than a delay. Customer service contains both kinds of work in the same queue. Any organization that treats it as one uniform workload will eventually route the second kind to a capability designed for the first, and will discover it the way Klarna did: publicly, and after the fact.
What to Take into Your Next Review
Do not ask whether AI is impressive. Ask where its strengths match the work, and where human judgment must stay in control.
Do not import someone else’s case study as a target. Import their measurement problem, because you are going to have the same one.
And when a capability produces value that no line on the income statement will show, decide in advance how you will recognize it. Otherwise the work will be real, the benefit will be real, and the organization will conclude that neither one was.
Adapted from Feasible, Not Achieved: Why AI Transformation Stalls and the NINE Decisions That Finish It, by Weston A. Jones, MBA and Jacob C. Roxon, PhD (JRNine Press).
The Jones-Roxon Nine™ is the Operator’s System for AI Transformation.