Klarna, the AI poster child, and the productivity paradox

"You can see the computer age everywhere but in the productivity statistics."

— Robert Solow, New York Times Book Review, 12 July 1987

The paradox's favourite counter-example

Almost forty years after Solow's remark, the productivity paradox has a new name and old numbers. A January 2026 National Bureau of Economic Research survey of roughly 6,000 senior executives at firms in the US, UK, Germany and Australia found that 69% of firms actively use AI, that even among regular users executives average only about 1.5 hours a week with it, and that nine in ten firms report no effect on employment or productivity over the past three years — while the same executives forecast a 1.4% productivity gain over the next three (NBER, "Firm Data on AI", Working Paper 34836). Apollo's chief economist puts the gap the other way: "AI is everywhere except in the incoming macroeconomic data" (Fortune, 17 February 2026). The Federal Reserve Bank of St. Louis measures 1.9% of excess cumulative productivity growth since ChatGPT launched; Daron Acemoglu's own estimate is "no more than a 0.66% increase in total factor productivity over 10 years" (NBER Working Paper 32487, 2024). Federal Reserve Governor Michael Barr, surveying the same ground in February 2026, concluded generative AI has had "relatively modest penetration" and that the evidence so far is "consistent with AI being a normal early-stage general-purpose technology" (Federal Reserve speech, 17 February 2026).

Every serious macro treatment comes with the same caveat: these are averages, and averages hide the firm that cracked it. The firm everyone reaches for is Klarna.

Klarna is the canonical case that AI makes a company, as a whole, more productive. In February 2024 the Swedish payments firm announced an OpenAI-powered assistant that had, in a single month, handled 2.3 million conversations, two-thirds of its customer-service chats — "the equivalent work of 700 full-time agents" — with resolution times cut from eleven minutes to under two, a 25% drop in repeat inquiries, and "an estimated $40 million dollar profit improvement in 2024" (Klarna press release, 27 February 2024). Three months later it announced that 87% of staff now use generative AI daily, under the headline "game changer for productivity" (Klarna press release, 14 May 2024). In December 2024 it revealed it had hired no human workers for the previous year, and in February 2025 its chief executive told Bloomberg what the market wanted to hear: "AI can already do all of the jobs that we, as humans, do" (Fortune, 9 May 2025). Klarna went public on the NYSE in September 2025 at $40 per share, a valuation of roughly $15 billion (Klarna Group Form 20-F; StockAnalysis). Here was the refutation of Solow in a single income statement: output per employee rising, headcount falling, costs flat.

The same company now offers the surprise ending. As of the close on 2 October 2026 KLAR traded at $12.51, its market capitalisation about $4.7 billion — roughly 69% below its IPO price and 72% off the 52-week high of $44.75 (StockAnalysis, 2 October 2026). Its most recent quarter, reported in August 2026, beat its own guidance on revenue, margins and income and was still greeted with a near-20% one-day selloff on a trimmed volume outlook (MoveSurge, 18 August 2026; Klarna Q2 2026 results). Whether Klarna refutes the productivity paradox or is the paradox's best exhibit is the subject of this article. Both readings have serious support; they are not mutually exclusive.

What Klarna claimed

The primary documents are unusually specific. On 27 February 2024 Klarna claimed its assistant had, within a month of launch, handled 2.3 million conversations; delivered an on-par customer-satisfaction score with human agents; cut repeat inquiries by 25% between December 2023 and January 2024; reduced average resolution time from 11 minutes to under 2; operated in 23 markets and more than 35 languages; and was "estimated to drive a $40 million USD profit improvement to Klarna in 2024" (Klarna press release, 27 February 2024; OpenAI case study, 27 February 2024 — a marketed portrait of a partner's product, treated accordingly).

The audited-looking numbers followed a year later, in the company's 2025 annual report on Form 20-F. For the year ended 31 December 2025:

Our AI assistant handled 80% of customer service chats in the year ended December 31, 2025, according to our service chat log data, doing the work equivalent of over 850 full-time agents (estimated based on the average monthly reduction in chat and telephone conversations handled by full-time agents in 2025 following the launch of our AI assistant), and in 2025 delivered approximately $59 million in cost savings.

The same document reports full-time headcount falling from 4,352 (2023) to 3,422 (2024) to 2,831 (2025), characterises the reduction as following from a "strategic decision to reduce our overall headcount and drive operational efficiency by leveraging AI", and expects further decreases. Employee count is not revenue, but the ratio moved hard: revenue per full-time employee rose from about $0.52 million in 2023 to $1.24 million in 2025 (my division of the 20-F's own figures), a 2.4-fold increase.

And the cost line behaved as the company claims it should. Customer service and operations expenses fell from $240 million in 2023 to $203 million in 2024 — the year the assistant began operating — and then rose just $4 million, or 2%, in 2025 while volumes rose 32% and transactions 25%, which the company presents as "continued operating leverage" (Form 20-F, FY2025). This is the heart of the firm-level claim, and it is real.

The fine print

That the expense line moved does not mean the headlines should be swallowed whole. Three features of the disclosure limit what the figures can prove.

First, the savings are estimates of avoided work, not audited results. The "equivalence of over 850 full-time agents" and the "$59 million" are the company's own computations from chat-volume data, resting on the average monthly reduction in conversations that, in the counterfactual, "full-time agents" would have handled. A reduction in resolved chats is not the same as value created, and the arithmetic of the original claim was looser still: 2.3 million conversations in a month divided across 700 "equivalent" agents is about 3,300 conversations per agent — roughly 150 per working day, plausible only if nearly all were short, routine tier-one queries, which is precisely what such an estimate assumes. None of these numbers is audited; they appear in the same filing that sells the stock.

Second, Klarna's support was largely outsourced, and always had been. The 20-F itself lists dependence on "third-party service providers for various critical functions, including ... customer service" among its risk factors, and the company's 2025 communications describe the re-hiring of human agents specifically "for the company's outsourcing partners" (Fortune, 9 May 2025; Form 20-F). Historically customer support was delivered through outsourcing firms, so "the equivalent of 700 agents" was mostly a reduction in invoices to third parties: a genuine cost saving to Klarna, but a transfer in and out of labour, not an internal headcount action. The largest pre-AI headcount cut — about 700 people, 10% of a 6,500 workforce — happened in May 2022, before the assistant existed, in response to the funding and credit crunch (CNBC, 23 May 2022). Most of the rest of the decline came from attrition and a hiring freeze that began in 2024, not from the assistant (Fortune, 9 May 2025).

Third, revenue per employee is not a productivity instrument. Klarna's revenue growth in 2025 was driven disproportionately by interest income, which rose 39% to $937 million and supplied $262 million of the $698 million total revenue increase — a product of scale and of higher funding yields, available to any lender regardless of AI (Form 20-F). Compounding that, headcount was already falling before the assistant launched and fell in every year of the period. A ratio that rises because revenue grew faster than an already-shrinking payroll is a fact; it is not a clean measure of AI's contribution.

The quality it had to buy back

In May 2025 the chief executive reversed the publicity machine. "It's so critical that you are clear to your customer that there will be always a human if you want," Siemiatkowski told Bloomberg, adding: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality. Really investing in the quality of the human support is the way of the future for us" (Fortune, 9 May 2025). Klarna began recruiting freelance human agents. Gergely Orosz, the software engineer who tested the assistant a year earlier, described it as "underwhelming" — it resolves only routine queries and "acts basically as a filter to get to customer support" (Fortune, 9 May 2025).

The walk-back is compatible with the company's claims, and this matters for a fair reading. The assistant never claimed to handle the hard tail of support — complaints, edge cases, financially distressed customers — and the "lower quality" admission concerns that tail, which had been left to an automated filter it was never good at. The routine majority stayed automated; in FY2025 the assistant still handled 80% of chats (Form 20-F). What the episode shows is that "productivity" as the company measured it — volume of chats resolved per agent-equivalent — structurally omitted the quality dimension that had to be bought back with human labour. A productivity measure that required a quality fix is incomplete, not fraudulent.

Did it make Klarna more productive?

Nothing in the fine print is a reason to doubt that the assistant works, and read one way Klarna is the cleanest demonstration that generative AI pays for itself in the right kind of work. The expense line is the argument. Customer service and operations cost Klarna $240 million in 2023, $203 million in 2024 — the year the assistant was switched on — and just $4 million more in 2025, while chat and transaction volumes grew by about a third (Form 20-F). A kink like that, at the right moment and of roughly the right size — a first-year saving of $37–40 million against the company's own estimates of $40–59 million — does not appear in an audited cost line by accident. The direction matches the best controlled evidence available: in Brynjolfsson, Li and Raymond's randomised trial, AI assistance raised contact-centre agents' output by about 14% (Quarterly Journal of Economics, 2025). Klarna's reported effect was larger, which is plausible: it replaced agents outright on a high-volume, well-specified task rather than simply augmenting them. And on the oldest definition of all, output per unit of labour, the books moved the same way — revenue per employee up 2.4-fold, the same work done by a much smaller payroll. Put in the most sympathetic light, Klarna did what it said it would.

The reservations are not about whether that happened; they are about how far the demonstration reaches. Every headline figure is the firm's own arithmetic, run on its own chat logs, measured against agents who were never hired. The saving is worth about 1.7% of revenue. And in the first full year of the assistant the profit that was meant to follow did not arrive: net income went from a $21 million profit in 2024 to a $273 million loss in 2025, adjusted operating profit fell from $181 million to $65 million, and Klarna added $807 million of operating expenses while its transaction margin slid eight points (Form 20-F). A $59 million saving that cannot be found in a bottom line that lost $273 million is not nothing — it is just too small to be seen from a distance, and the market has said so in its own noisy way: the two most credible ratings actions since the listing were downgrades to Neutral and Hold — "a show-me story" — and the stock, which listed above $40, now changes hands at $12.51 (MoveSurge, 18 August 2026; StockAnalysis, 2 October 2026).

Putting a number on all of this is the honest part of the exercise. That the assistant genuinely took over most routine chats and saved real money, I would put at about 85% — the expense-line kink is corroboration of a kind no press-release arithmetic can fake. That the figures as published — the $59 million, the 850-plus agents — are accurate is a weaker bet, closer to six in ten: the direction is solid, the magnitude is an estimate, and the attribution is muddied by interest income and prior layoffs. And that any of it registers in Klarna's profit, its market value or the industry's statistics — the test by which the paradox is actually judged — I would put well below two in ten. That last number is the one worth keeping.

So here is the answer, in plain terms. The AI at Klarna worked — about as well as AI has worked anywhere — and the honest size of what it bought was tens of millions of dollars a year, roughly 1.7% of revenue. That is a real win, and it is a small one. It did not show up in profit, it did not move the share price, and it left the productivity statistics untouched. Klarna shows that AI can make one part of a company more productive. It does not show that AI can make a company as a whole more productive — and the gap between the two is precisely the productivity paradox. The paradox never claimed that AI does nothing; it claimed that you cannot see it in the numbers. Even the strongest case on record — Klarna's — leaves that claim standing.

Sources

Klarna (primary, all self-reported)

Academics and official bodies

Reporting

Market data

Vendor content (used for leads only, not as evidence) — usefini.com, twig.so, aiintelreport, gteams.ai, promptlayer.com, valueaddvc.com and similar blogs re-state Klarna's Q3 2025 figures ("853 full-time agents", "$60 million", NPS 73) and market surveys; they are not independent sources and are cited here only to flag that re-reporting of self-reported figures is not verification.

In the first full year of the assistant, Klarna's profit, its share price and the industry statistics all failed to move.