By This Hour AI Desk
Artificial intelligence may be reshaping corporate plans, investment priorities and executive rhetoric, but it has not yet translated into an observable financial payoff for most large US companies outside the technology sector, according to Torsten Slok, chief economist at Apollo Global Management.
Slok’s argument, as summarized by TechCrunch, centers on a blunt distinction between widespread enthusiasm for AI and evidence of improved business performance. He examined a group he calls the S&P 493: the S&P 500 excluding Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla. Across that remaining group, the analysis found profit margins and revenue per worker appearing flat, rather than showing a clear productivity improvement attributable to AI.
The implication is not merely that companies have been slow to deploy a new tool. Slok’s stated position is more demanding: outside the major technology businesses, corporate America has not demonstrated a financial return from AI in the measures he reviewed. That challenges the expectation that increasingly available AI systems will straightforwardly lift revenue growth, cut costs or widen margins across a broad range of large companies.
The argument turns on results outside the biggest technology groups
The construction of the S&P 493 is central to the claim. By removing seven companies associated with major technology platforms, AI development, semiconductors, cloud infrastructure or closely related businesses, Slok sought to look beyond the firms most directly positioned to benefit from the AI buildout. His conclusion concerns the remaining large listed companies, not the S&P 500 as a single undifferentiated group.
That distinction matters because the excluded companies can shape the broader conversation about AI’s commercial value. They sit close to the products, computing capacity, platforms and services associated with the technology. A strong financial showing among those businesses would not, by itself, establish that a manufacturer, retailer, insurer or other non-tech corporation has achieved comparable gains. Slok’s framing separates the prospects of AI suppliers and major technology operators from the realized results of the companies using, testing or considering the technology elsewhere.
His assessment also uses company-level outcomes rather than announcements of adoption. Corporate discussion of AI can cover many activities: experimenting with software, placing it into workflows, purchasing related services, training employees or exploring possible uses. Slok’s argument, as described in the report, asks a narrower question. Have these efforts produced higher revenue per worker or better profit margins across the non-tech group? The analysis says the available pattern did not show that they had.
That is a consequential standard. A tool can be available and actively used without yet changing a company’s reported financial profile. Conversely, a change in a financial metric does not automatically prove that AI caused it. Slok’s conclusion rests on the absence of a visible improvement in the measures he chose, not on a claim that no company has experimented with AI or that the technology cannot perform useful tasks.
Flat margins and revenue per worker form the core evidence
The two indicators described in the TechCrunch account are profit margins and revenue per worker. Profit margins speak directly to the portion of revenue left after costs, while revenue per worker offers a broad measure that can be used to consider output relative to employment. In Slok’s account, both appeared flat for the S&P 493. He saw no evidence in that analysis that AI was improving productivity for the group.
Those findings lead to the report’s stark formulation of “zero financial payoff.” But the phrase should be read in the context of the analysis rather than as a precise accounting statement about every AI project at every large company. The supplied account does not describe company-by-company AI spending, project-level results, or a separate calculation of returns for individual deployments. It instead describes a broad reading of aggregate corporate outcomes among the companies remaining after the seven major technology names were excluded.
That scope leaves room for an important difference between an aggregate finding and the experience of any particular business. Flat results across a large basket can coexist with uneven results inside it. Some companies may report local operational uses, while others may be early in experimentation, or may see costs and benefits that do not move company-wide measures. The report does not provide evidence sufficient to identify winners, laggards, industries, applications or the reasons any individual company’s results may differ.
Nor does the supplied material establish a timetable for when a benefit should become visible. Slok’s conclusion is about the absence of evidence of an improvement in the measures he examined. It does not specify how long companies have used particular AI tools, whether their investments were intended to generate an immediate return, or whether later financial statements could show a different pattern. These omissions do not negate his observation; they set limits on what it can demonstrate.
Slok questions whether cheap intelligence can become business value
Slok’s skepticism goes beyond the claim that adoption is incomplete. As summarized in the source material, he questions whether lower-cost access to intelligence can readily become faster revenue growth and higher margins. The point is that business performance involves more than a system’s ability to calculate, summarize or generate an answer.
His reasoning emphasizes judgment and the elements of organizational life that are not captured by an abstract notion of intelligence. Accountability, trust, culture, unwritten rules, ethical trade-offs and fairness all affect reputational risk, he argues. In that view, they also shape the commercial outcomes that determine revenue and profitability.
That argument provides a possible explanation for the gap Slok sees between intense interest in AI and flat company-level indicators outside tech. If the work that determines durable commercial outcomes depends heavily on judgment and organizational responsibility, then making a form of intelligence cheaper may not automatically improve those outcomes. AI could be capable of handling a task while still leaving businesses to decide how to use its output, who bears responsibility for decisions, and how customers or other stakeholders respond.
The supplied reporting does not test that explanation against alternatives. It does not show, for example, whether the flat indicators reflect AI’s limitations, a delay between deployment and impact, the way companies measure performance, or factors unrelated to AI. Slok presents a skeptical interpretation, not a detailed causal account that resolves each of those possibilities.
Still, the separation between capability and payoff is the heart of the economist’s warning. Business leaders can view AI as useful without treating usefulness as proof of a measurable return. Under Slok’s approach, the relevant evidence is whether broad financial indicators improve among companies outside the technology sector, and he says they have not.
A broad conclusion with significant boundaries
The report raises a question for companies and investors assessing AI initiatives: what would count as convincing evidence of financial payoff? Slok’s answer, inferred from the measures highlighted in the account, points toward revenue per worker and profit margins. Those are consequential outcomes, but they are high-level ones. They may not capture every reason a company chooses to use a tool, and the supplied material does not argue that they do.
Equally, the S&P 493 label defines a particular boundary. The claim does not say that every company in the group has obtained exactly the same result. It does not purport to measure smaller businesses, private companies, firms outside the United States, or the seven excluded technology companies. It also does not present evidence about AI’s financial impact in the economy as a whole. The conclusion is specifically directed at large US corporations outside the named technology cohort and is based on the broad measures Slok reviewed.
Readers should therefore distinguish three propositions that can easily be collapsed into one. First, AI has received substantial corporate attention and investment. Second, certain major technology companies are excluded from Slok’s comparison group because of their close connection to the sector. Third, Slok says the remaining group has not shown a financial benefit in flat margins and revenue per worker. Only the third proposition is the central empirical claim of this report, and it is limited by the scope and method described.
The report has not been independently corroborated. The available material is a secondary summary of Slok’s analysis, and it does not include the underlying data, full methodology, time period, company-level breakdowns or an independent assessment of whether AI was isolated from other influences on profits and revenue per worker. Those gaps mean the conclusion should be treated as a significant but unverified interpretation of broad corporate performance, rather than a settled finding on AI’s economic value.
What can be said from the supplied account is narrower and clearer. Apollo’s chief economist argues that, after setting aside seven major technology companies, the financial indicators he examined do not show AI-driven productivity gains among the rest of the S&P 500. Whether subsequent results, more detailed disclosures or other analyses support, refine or challenge that judgment is not established by the material provided.
For further context on this subject, see Labour immigration argument centres on care workforce, article says.
Reporting notes
What is confirmed: The reported analysis found profit margins and revenue per worker appearing flat across that group.
Why this matters: The claim separates widespread AI use and investment from evidence of higher margins or productivity.
What remains unclear: The underlying data, period, methodology and company-level results were not supplied. This report is based on one source and has not been independently corroborated.
