How Strong Is the Evidence That AI Isn't Paying Off Outside Tech?

How Strong Is the Evidence That AI Isn't Paying Off Outside Tech?

Apollo's Torsten Slok makes a strong descriptive case: in his October 24, 2026 note, AI-linked gains show up in the Magnificent 7 and not in the rest of the S&P 500. The causal case is weaker. His charts compare aggregates, can't isolate AI from other forces, and may be measuring too early. To judge the claim properly, you have to look at each metric, its lag and what it can actually identify.

What is the claim, stated as a testable hypothesis?

Stated precisely, the claim is that AI spending has not yet produced measurable financial gains for large US companies outside the tech sector. In the October 24 Daily Spark, titled "No Signs of AI Payoff Outside the Tech Sector," Slok compares two groups: the Magnificent 7 (Apple, Microsoft, Alphabet, Amazon, Meta, Nvidia and Tesla) and the other 493 S&P 500 companies.

That makes it a null-result claim. Before accepting one, ask whether the instruments are sensitive enough and whether the sample covers enough time for an effect to appear.

Which metrics does Slok rely on, and what does each one measure?

Slok rests his case on four kinds of evidence, built from Bloomberg, Macrobond, San Francisco Fed and Ramp data:

• Inflation-adjusted revenue per worker. This is a rough output-per-employee measure. It rises for the Mag 7 and stays flat for the S&P 493. It blends pricing, product mix and headcount with real productivity.
• Profit margins. The same split shows up here. Margins also reflect competition and input costs, so a flat margin can hide efficiency gains competed away through lower prices.
• Total factor productivity. The San Francisco Fed's utilization-adjusted TFP series shows no AI boost so far. It is a national figure and subject to revision.
• Ramp spending data. Ramp's card and spending data show token spending concentrated among venture-backed firms. That measures adoption intensity, not outcomes, and Ramp's customers are not a random sample of US businesses.

Agreement across metrics strengthens the descriptive pattern, but none was designed to estimate AI's effect.

Why doesn't a Mag 7 versus S&P 493 gap prove causation?

The gap doesn't prove causation because the two groups differ in many ways besides AI. The Mag 7 sell AI as well as use it, and they had pricing power, cloud scale and cost programs before generative AI arrived. Their rising margins may reflect any of these. Slok's own August 9 note frames it that way: AI gains are "only showing up in the sellers' margins, not the buyers'."

Vendor revenue is a gain for the seller and a cost for the buyer, so the comparison partly measures money moving between the groups.

How do time lags complicate the reading?

New technologies usually depress measured productivity before they lift it, so an early null result is consistent with both "no effect" and "effect not yet visible." Goldman Sachs economist Elsie Peng, reported by Fortune, modeled a J-curve with significant gains only around year eight, similar to how slowly PCs showed up in productivity statistics. Slok's position is mainly about timing too. In a July interview with Fortune, he called the length of the ROI runway outside tech the key issue.

For a researcher, this means the observation window is part of the result. A short window can't rule out an effect that typically arrives years later.

What gets lost when you measure at the aggregate level?

Aggregates can hide variation between firms, so an index that looks flat can contain clear winners and losers. Fortune described a Ricoh claims project that ran over budget but eventually tripled the division's productivity, a gain invisible in a 493-company average. Three related issues:

1. Sector composition. The S&P 493 mixes very different industries. Slok's August note found health care margins roughly halved since 2015, which can pull down an average regardless of AI.
2. Survivorship and index churn. Index membership changes over time, so the group being measured is not a fixed panel of firms.
3. Weighting. Market-cap weighting and equal weighting can tell different stories, and the choice is rarely stated in chart-based notes.

How thin is the firm-level measurement today?

Firm-level measurement barely exists, and that is the biggest limit on the evidence. Forbes reported Slok's analysis of Q2 earnings: about 69% of S&P 500 companies actively use AI and 29% reported a quantified result, but only about 2% tracked an AI metric over time. Without it, researchers fall back on aggregates.

What would a rigorous AI-ROI study design look like?

A rigorous design compares like with like, sets the outcome in advance, and follows it long enough to see a lag. In practice:

1. Pre-register outcomes. Choose the metrics before deployment: revenue per worker, unit cost, conversion rate or cycle time. Record a baseline.
2. Build a credible counterfactual. Use randomized rollouts across teams or regions where you can. Where you can't, use difference-in-differences against matched non-adopters in the same sector.
3. Measure dose, not just adoption. Record spend, usage intensity and workflow coverage, not a yes/no flag.
4. Separate efficiency from value. Track whether time saved became more output, lower cost or freed budget.
5. Plan for the J-curve. Commit to repeated measurement on a fixed schedule over several years.
6. Report heterogeneity. Break results out by sector, firm size and use case instead of reporting one average.

How can marketing measurement apply these principles?

Marketing is a useful test case because holdout and incrementality methods already exist and can be applied to AI workflows instead of hours-saved estimates. One measurable outcome is AI visibility, meaning how a brand shows up in model-generated answers. MarkGrid's Model Share tracks that over time, matching the fixed-schedule principle. When results hold up, Budget Optimization can model reallocating budget across channels. For the reporting and charts behind the story, see MarkGrid's full breakdown of Apollo's AI payoff warning.