Apollo’s chief economist says companies outside tech have seen no financial gains from AI. His evidence is a set of charts built from public company reports: outside the biggest tech firms, profit margins stay flat. I build AI systems for a living, so I wanted to know whether that is even true. I did the math. Public reports can’t settle this question in either direction. To know, you have to look inside a company.
A flat line can hide $105 million
A bathroom scale shows a slightly different number every time you step on it. Before you trust it to tell you whether you gained a pound, you check how much it jumps around. I did the same with company accounts.
I took 180 large US companies outside tech, finance and real estate, and compared industries where more of the work is the kind AI can do with the rest. The gain I looked for is those industries pulling ahead of the others by one point of operating margin, say from 10% to 11%. A typical company here had about $10.5 billion in revenue, so for it one point is roughly $105 million a year. For comparison, Klarna says its AI assistant saves it about $58 million a year. Then I looked at how much margins moved on their own in 2015-2019, before generative AI, and asked how often a test on public accounts would spot a one-point gain inside that movement.
The answer is 16 times out of 100. To be spotted 80 times out of 100, a gain would have to be about $313 million a year. Before collecting any results, I had decided that if the test could not see a one-point gain, I would not run it on the AI years, 2024 and 2025, because a “no effect” from it would mean very little. So I stopped there.
Apollo’s charts are blunter than my test, because they average every company together. If firms with a tenth of all revenue gained one point of margin, the average would move by a tenth of a point.
The test has limits: one sample of 180 companies, AI exposure estimated from the jobs people hold in each industry rather than from observed use, and one pre-AI period.
The big numbers measure something else
The optimistic numbers fail in a different way. Forrester’s 132-353% ROI is a projection for a modeled company. McKinsey’s 39% (2025) and 37% (2026) of firms reporting an AI contribution to profit are what respondents say about themselves. BCG finds that companies further along with AI have better financial results, but comparing leaders with everyone else does not separate what AI earned from everything else those companies did. I checked 33 numbers like these from 17 studies and reports, and none of them shows what was left after all costs, compared with not using AI. The full list is at the end of the full version of this article.
Where the money goes in one project
To see where the money goes, I took one project. In a large experiment on an international shopping platform, 44,614 users got either the usual “support unavailable” reply or an AI chat when they had a question before buying. Per 1,000 users, purchases rose by $274. Note what this compares: an AI chat against no help at all, so it says nothing about replacing a good human service.
Suppose the shop keeps 20% of a sale after product and running costs. Then the $274 becomes about $55. Take away $4.50 for the model and $20 for the rest of the service, and about $30 per 1,000 users is left to pay for the launch. With a $10,000 launch, the project needs about 330,000 users to break even. The 20% margin and the costs are my assumptions, not the shop’s real budget.
Model prices keep falling, and people expect that to fix this math. With a model ten times cheaper, the project still needs 291,000 users.
The model was a small part of the bill from the start. And if the real sales effect sits at the low end of the experiment’s range, about $84 per 1,000 users, the project loses money with either model.
Klarna’s savings
Klarna is the best-known public case. In February 2024 it forecast a $40 million annual profit improvement from its AI assistant. In May 2025 its CEO told Bloomberg that the focus on cutting costs had gone too far, and described a pilot to hire human agents so customers would keep a human option. Later that year it reported $58 million in annual savings, calculated as fewer customer contacts priced at what human agents would have cost. That is avoided cost, and it leaves out what it took to build and run the assistant. In the second quarter of 2026 its customer service and operations costs rose from $48 million to $57 million while the business grew. A rising total in a growing company proves nothing on its own, which is the point: even at Klarna, the assistant’s profit cannot be read from the public numbers.
How I would check it inside a company
Roll the AI out to some stores, teams or regions picked at random, and keep the old process in the rest. Over the same months, compare the extra money each side brings in and the quality of the work. Then subtract everything the project costs: launch, models, integration, training, the people who handle escalations and the ongoing support. Hours saved count as money only when they let you skip a hire or serve more customers. A negative balance is an answer too, and it costs less to learn it from a pilot than from a full rollout.
What to ask before calling AI profitable
- Compared with what? Was there a real working alternative without AI, rather than an easy invented comparison?
- Which money, and when? Is the financial outcome named, with the time allowed to achieve it?
- What is in the bill? Does it include launch and all extra costs, including rework and human involvement?
- How was it checked? Were the comparison, quality standards and stopping conditions set in advance?
- How much demand is there? Does real demand cover the investment at the quality that was tested?