# Apollo’s AI Payoff Warning: What the Spending and Productivity Data Show

A new note questions AI’s financial returns outside the biggest technology companies, but the evidence measures several different things.

By Paras Madan | 2026-10-25


Apollo chief economist Torsten Slok published a warning on October 24: AI’s financial benefits are not showing up broadly outside the largest technology companies. His comparison of corporate margins and productivity raises a useful question for AI buyers, but it does not establish that every company’s AI investment has failed.

The distinction matters because AI adoption, spending and financial returns answer different questions. A company can buy more software without becoming more profitable. It can also use more AI while paying less for each unit of work.

## What Apollo’s AI payoff warning says

In [the new Apollo note](https://www.apollo.com/wealth/insights-news/insights/daily-spark/no-signs-of-ai-payoff-outside-the-tech-sector), Slok compares the Magnificent Seven with the remaining S&P 500 companies, which he calls the S&P 493. He argues that revenue per worker and profit margins are improving in the first group while remaining flat in the second.

He also points to productivity measures and the concentration of AI spending. His conclusion is that the broader business payoff remains difficult to see. This is a newly published interpretation of several datasets, rather than a new controlled study of companies adopting AI.

That makes it a useful challenge to broad claims about enterprise returns. It is a much weaker basis for deciding whether a particular support system, coding tool or document workflow is worth buying.

## Why aggregate margins cannot settle AI ROI

Consider a company whose margin stays unchanged after introducing an AI support tool. Perhaps the tool saved money while another cost increased. Perhaps it reduced support costs but discounts absorbed the benefit. Perhaps it made little difference at all. The final margin alone cannot distinguish these explanations.

The same problem grows when hundreds of companies are combined. Their exposure to AI, adoption dates and reasons for buying it vary. Comparing two broad groups does not isolate the effect of AI from everything else happening inside those businesses.

**A lack of visible aggregate improvement is a reason to ask for better evidence, not proof of zero returns in every deployment.** The question is what happened relative to a credible baseline, including what would have happened without the tool.

## AI spending and AI usage are separating

[Ramp’s AI Index](https://ramp.com/data/ai-index) adds an important qualification. Its displayed token data show spending increasing 23.0% and token volume increasing 52.4% in the week ending October 11 compared with the week ending September 13. Usage rose faster than the bill.

Those figures come from businesses using Ramp’s Token Spend Management product, not the whole economy or even all AI spending on Ramp. Its broader adoption measure counts observed payments and misses some free tools and employee purchases through personal accounts.

Ramp’s [October 14 analysis](https://ramp.com/data/ai-index-oct-2026) also describes lower effective prices and concentrated spending. These are earlier observations that help interpret the new warning, not fresh October 25 measurements.

For an AI vendor, slower spending growth could mean weakening demand, lower prices, a different workload mix or some combination. For the buyer, a smaller bill can be good news. Neither dollar spending nor token volume, by itself, tells us whether the work produced useful business results.

## Productivity data have their own limits

The [San Francisco Fed’s productivity series](https://www.frbsf.org/research-and-insights/data-and-indicators/total-factor-productivity-tfp/), updated October 8, reports total factor productivity growth of 1.42% over the four quarters ending in the second quarter of 2026. Its utilization-adjusted measure fell 0.07% over that period.

The adjustment accounts for changes in how intensively labor and capital are used. That helps separate some cyclical effects from underlying efficiency, but the resulting measure is still an aggregate measure for the U.S. business sector. It does not identify which part of the change came from AI, and its period ends before the latest autumn releases.

This is why a macroeconomic series should not be treated like a benchmark for a particular agent. It can test whether a broad transformation is becoming visible, while leaving individual deployments unresolved.

## What evidence should come next

For enterprise buyers, the useful response is to connect each deployment to a measurable result: cost per resolved case, completed work at an acceptable quality level, or revenue from an existing process. Include integration and review costs, and compare equivalent workloads rather than counting subscriptions.

The next convincing sign of an AI payoff will be repeatable improvements that survive that comparison across more businesses. Cheaper tokens make experimentation easier. They do not remove the need to establish what the resulting work is worth.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams explain technical products and enterprise use cases clearly, with the evidence and limits intact.

[Book a 30-minute call](https://cal.com/paras-madan/30min).