Apollo's AI payoff claim hinges on how productivity is measured

Apollo's latest AI payoff argument gives businesses a reason to demand better evidence before expanding software spending. It does not establish the return on a particular company's AI workflow. Revenue per employee, profit margins and economy-wide productivity answer related but different questions.

In an October 24, 2026 Daily Spark note, Apollo chief economist Torsten Slok argues that financial benefits have not spread beyond the largest technology-related companies. He describes rising revenue per worker and margins for the Magnificent Seven, against flat measures for the remaining S&P 493. This is a new commentary on the evidence, not a newly reported controlled deployment experiment.

Start with the population behind the claim

The S&P 493 label refers to the rest of that stock index after the seven-company group is removed. It should not be read as a census of every nontechnology employer or every AI user.

Apollo notes that its revenue-per-worker comparison is inflation-adjusted to the fourth quarter of 2019. That helps define the measure, but it remains revenue divided by workers. It is not the same quantity as real output per hour or the cost of completing an individual business task.

A group-level pattern can coexist with different results among its members. Establishing an AI effect would require a comparison that distinguishes the deployment from other changes in the business. That is a limit of the inference available from these aggregate measures, not a claim that TechKili has found hidden productivity gains.

Keep the productivity measures separate

The Bureau of Labor Statistics' September 3 release defines labor productivity as real output per hour. It reports nonfarm-business growth of 1.4% at an annualized quarterly rate in the second quarter of 2026, and 2.2% from the same quarter a year earlier. Neither number measures the contribution of AI on its own.

Total factor productivity takes a broader accounting approach to labor and capital. The San Francisco Fed's methodology, updated October 8, describes adjustments for changes in how intensively those inputs are used. This matters because running existing equipment harder and improving the underlying production process can both raise output per worker.

In her September 3 analysis, San Francisco Fed researcher Huiyu Li explains how capital investment, workforce composition and utilization can contribute to labor-productivity growth. She treats the economy-wide AI contribution as uncertain. Rising output per hour is therefore insufficient by itself to credit AI; an inconclusive aggregate signal is also insufficient to assign zero value to every deployment.

Define the evidence before increasing the budget

For a business, the useful response is a measurement plan tied to the same work before and after a change. Define the accepted output, turnaround time, rework and human review required. Include token charges, integration and operating labor when calculating the cost per completed task.

A faster first draft may still require enough correction to erase the saving. Conversely, a workflow can be useful without moving a large public company's overall revenue-per-worker measure visibly. These are evaluation possibilities, not outcomes observed by TechKili.

Expand only when the measured result meets the business's quality and cost requirements. Slok's caution is relevant to that decision, but neither a broad pessimistic claim nor a vendor's productivity promise can replace evidence from the work being funded.

Sources

• Apollo: Torsten Slok's October 24 AI payoff commentary

• BLS: September 3 revised second-quarter productivity release

• San Francisco Fed: total factor productivity methodology

• San Francisco Fed: Huiyu Li's September 3 productivity analysis