Deep dive: Where Is AI's Payoff Outside Big Tech?

Why today

Torsten Slok is chief economist at Apollo Global Management, a large private-markets investor. He just published a short note in his Daily Spark series with a blunt headline: no signs of AI payoff outside the tech sector. In his words, the technology "has delivered zero financial payoff to corporate America outside of tech." His evidence is a set of charts. For the seven giant tech firms, revenue per worker and profit margins are rising. For the other 493 companies in the S&P 500, both are flat. Gary Marcus, a long-time critic of generative AI, picked it up quickly. He said Slok had "made a brutal case" and used it to support his argument that the technology may be a net harm. The timing matters because a lot of startup valuations rest on enterprise AI revenue. If the customers aren't getting paid back, that revenue comes from experimental budgets, and experimental budgets get cut. So today's question is this. Does the gap show AI failing its customers, or a lag the charts can't see yet?

Plainly

Start with what the charts measure. Revenue per worker is sales divided by headcount. Margin is profit as a share of sales. Suppose AI made a bank or a retailer more efficient. You would expect one of those numbers to move: more sales from the same staff, or the same sales at lower cost. Slok's point is that for the S&P 493, neither has moved. Now look at the other line. In this picture the Mag 7 are mostly not AI buyers. They are AI sellers. When the 493 spend on cloud, chips and model access, that money shows up as revenue at Microsoft, Nvidia and Amazon. So the rising Mag 7 line partly measures everyone else's AI bill. The two lines are linked. Here's a worked example. Picture a regional insurer with 10,000 staff that buys AI tools for claims handling this year. The licences and the integration work are costs today. Staff get retrained, not let go. Claims move faster, but demand hasn't changed, so revenue doesn't change either. On Slok's chart, this insurer shows flat revenue per worker and a slightly thinner margin. Notice that both stories predict exactly this picture. If the AI is useless, you see it. You also see it if the AI is useful but the firm hasn't yet redesigned workflows, slowed hiring or launched new products. That is the old productivity paradox. In the late 1980s Robert Solow observed that computers were everywhere except in the productivity statistics. The economic historian Paul David answered in 1990 with electricity. Factories that swapped steam engines for electric motors saw little gain at first. The payoff arrived decades later, once they rebuilt their layouts around a motor on each machine. Erik Brynjolfsson, Daniel Rock and Chad Syverson formalised this as a productivity J-curve. Firms first spend on intangibles like process redesign and training, which the national accounts barely record. Measured productivity dips. Later the harvest shows up, and can even be overstated. So a flat line fits failure, and it also fits lag. The chart alone can't tell them apart. For that you need data closer to the work.

What the evidence shows

Here is what's established. For individual tasks, there is careful evidence that generative AI helps. Erik Brynjolfsson and two co-authors studied 5,179 customer support agents while an AI assistant was rolled out in stages. Issues resolved per hour rose 14% on average. Novice and low-skilled workers gained 34%, while experienced agents gained little. Because the rollout was staggered, the study is close to a natural experiment. It is now published in the Quarterly Journal of Economics. Zoom out and the effect shrinks. In a working paper from the University of Chicago's Becker Friedman Institute, Humlum and Vestergaard linked surveys of 25,000 Danish workers in 11 exposed occupations to employer payroll records. They found no significant effect on earnings or recorded hours. Executives tell a similar story. An NBER working paper surveyed nearly 6,000 senior executives in the US, UK, Germany and Australia. Adoption is also thinner than the hype suggests. The Census Bureau's business survey puts overall AI use at between 17% and 20% of US firms. That covers December 2025 through May 2026. Among firms with 250 or more employees, the figure was 37%. Now for what is claimed but not shown. Slok's note shows aggregate charts. As published, it doesn't separate AI from the other forces that move margins, such as interest rates, input costs and pricing power. Nor does it compare heavy adopters with firms that barely adopted. The word zero claims more than a flat line can support. Marcus goes further, to net harm for society. That rests on a list of harms with no numbers attached, plus Slok's charts. Both are commentary, not measurement. On the other side, the J-curve is a theory of why gains might be hidden. It predicts a lag. It doesn't prove the gains are coming.

How people read it

Here the readings genuinely differ. The AGI-forward reading says the task-level gains are real and the tools have improved since the surveys were run. The Danish data come from late 2023 and 2024, before today's agentic tools. Aggregates always trail. Executives are forecasting gains, and electricity took decades where this may take years. On this view, today's flat lines are the bottom of the J. The AI-realist reading accepts the task gains but calls them small and narrow. The Danish authors also find that chatbots create new job tasks, so some of the time saved gets used up elsewhere. And every AI-drafted output still needs someone to check it. On this view, gains will be modest and may never show up in the aggregates. The market reading points out that margins are the wrong place to look. If every insurer gets the same tools, competition pushes the savings into lower prices. Customers win, while shareholders of the 493 see nothing. That would make Slok right about margins and wrong about payoff. For AI vendors, it means selling into a cost line that buyers will watch very closely. All three readings fit the current facts. What separates them is evidence nobody has published yet: a comparison of adopters and non-adopters within the same industry.

What it means for builders and investors

Our read is that the evidence leans one way, and it is narrow. Firm-level and economy-wide studies so far show small measured effects. Task-level studies show real gains in narrow work that is measured closely. Slok's charts add little to this, because they can't separate lag from failure and the Mag 7 line partly counts the 493's own spending. Still, the stronger studies point roughly where he does: so far, small. For builders and investors, this is a question about revenue quality. A lot of enterprise AI revenue today probably comes from pilot and innovation budgets. That money is real, but it's discretionary. The products that survive a budget cut tend to be tied to a metric the customer already tracks, like tickets resolved, claims closed or error rates. The support study worked partly because resolutions per hour were already being counted. One question worth exploring in diligence might be this: which line in the customer's books does the product move, and has a customer measured that change against their own baseline? A renewal with expansion is strong evidence. A logo on a slide is weak. Our view would change if firm-level data began showing adopters pulling clearly ahead of their peers.

What to watch

Three things would settle this. The first is a firm-level split: studies that compare heavy AI adopters with peers in the same industry on margins and output. If the J-curve is right, adopters should start pulling ahead within a few reporting cycles. The second is the executive survey's own forecast. The third is the Census adoption rate. If use among large firms climbs well past 37% while margins stay flat, the lag story gets weaker. If adoption stalls and pilots are cut in the next budget cycle, the experimental-budget reading is confirmed. That cycle is the near-term test.