Apollo’s AI payoff claim exceeds what its charts show

Apollo chief economist Torsten Slok’s October 24 commentary argues that AI has delivered no financial payoff outside technology. His charts cannot establish that causal conclusion.

For a manager deciding whether AI expenditure is worthwhile, the missing link is the outcome at businesses using it. A comparison needs to connect adoption and costs to results, while accounting for what would have happened without the technology. An aggregate financial ratio cannot supply that comparison on its own.

The big change

• What changed: Apollo’s challenge to AI returns exposes a measurement gap: sales to adopters do not establish gains for those adopters.
• Why it matters: A payment is revenue for one business and a cost for another. Neither side of that transaction establishes whether the purchaser’s work became more valuable after all costs.
• What to watch: The useful evidence connects an adopter’s expenditure to its own outcomes over a defined period, with a credible comparison. That would let managers assess a return without treating an economy-wide average as their business case.

Company ratios leave the cause unresolved

Apollo compares the Magnificent Seven with the S&P 500’s other 493 companies using inflation-adjusted revenue per employee and profit margins. The margin series moved substantially before 2023. These groups neither distinguish AI adopters nor represent every business outside technology.

Revenue per employee divides sales by headcount. It can change with either number, without identifying the contribution of a particular tool. Profit margin relates profits to sales, so it also incorporates costs and pricing beyond AI. Even an unchanged margin leaves unanswered whether AI prevented a decline, had no effect or added costs. The charts offer no comparison that chooses among those explanations.

This limits the positive claim too. Higher ratios at a technology company do not reveal how much value its customers obtained from buying AI. Establishing that would require evidence from the customers and their work.

Productivity measures answer different questions

The Bureau of Labor Statistics’ revised second-quarter release, published September 3, reports that US nonfarm business output per hour rose 2.2% from a year earlier in the second quarter of 2026. That is growth in measured labor productivity; the release does not attribute it to AI.

The San Francisco Fed’s October 8 update reports a 0.07% decline in utilization-adjusted total factor productivity over the four quarters ending in the same quarter. Its series covers the US business sector and accounts for labor and capital inputs, with an adjustment for how intensively they are used. The Fed cautions that some corrections, including effects from reallocation and imperfect competition, are absent from the quarterly series.

These results can coexist because their definitions and coverage differ. More output for each hour worked does not automatically imply the same increase in output relative to combined labor and capital inputs. Neither measure isolates AI adoption. Reading the figures together supports a narrower conclusion: aggregate productivity evidence has to be interpreted through its measurement method before it can inform a claim about a specific technology.

A spending slowdown can hide more usage

Ramp’s October 14 analysis reports virtually no token-spending growth in August and September despite rising token volumes. Economist Ara Kharazian attributes the pattern mainly to falling prices, alongside model efficiency and quality improvements.

Ramp’s methodology says token metrics come from businesses using its Token Spend Management product, which imports connected providers’ usage and billing data. They do not represent all AI spending across the platform’s broader sample of more than 85,000 businesses. The same page’s weekly comparison, October 11 against September 13, shows token spending rising. That later interval should not be substituted for the earlier monthly period.

Dollar spending combines the amount bought with its price. It therefore cannot, by itself, show whether a purchaser used less AI or obtained less benefit. Token counts add a measure of consumption, but they still do not measure completed useful work or the cost of checking and correcting it.

For the wider economic context, BIG CHANGE has covered the IMF’s warning about uneven gains from AI investment. The management question here is more specific: which costs and outcomes belong to the same adopting business? Until those are connected, a supplier’s revenue, a national productivity series and a customer’s AI bill remain incomplete evidence of that customer’s return.