Deep AI Adoption Reaches 11% of S&P 500

Large firms are central to the economic case for AI because their internal process changes can affect investment, productivity, and labor demand at scale. This paper studies whether that adoption is actually happening inside S&P 500 firms, rather than being inferred from product announcements, executive rhetoric, or employee use of consumer tools. The authors’ answer is cautious: adoption is rising, but deep enterprise integration is still concentrated and uneven.

Core Contribution

The paper’s main contribution is a firm-year measure of AI adoption built from SEC 10-K filings from 2016 through 2025. The authors use these filings because public companies face legal constraints against materially false or misleading statements, making the text a cleaner signal than marketing pages or earnings-call language. The measure separates firms with no adoption from firms using AI in products or service delivery, and from firms where AI appears deeply integrated into business processes.

That distinction matters. A company can mention AI because it sells an AI-enabled product, because its employees use AI tools, or because machine learning has become part of core operations. The paper tries to isolate the last category, which is the one most relevant for enterprise productivity and labor-market effects.

Technical Approach

The authors classify adoption from 10-K text using AI-related keyword and phrase patterns, then validate the resulting industry-level measures against external AI exposure and adoption benchmarks. The appendix shows the keyword set and firm-level samples, while the main figures compare adoption intensity across industries and over time. Figure 1 summarizes the industry-level validation exercise: the constructed measure lines up with sectors that external sources also identify as more AI-exposed, while avoiding the claim that every AI mention implies operational transformation.

At the firm-year level, the paper tracks adoption across the S&P 500 universe and distinguishes technology firms from non-technology firms. Figure 2 shows the time pattern: AI adoption was low and fairly flat before the generative-AI boom, then accelerated after 2022. By 2025, 11% of S&P 500 firms are classified as having deep AI integration, with another 10% using AI in production of goods or delivery of services. The paper reports that overall adoption has more than quadrupled from 5% in 2022.

Results and Analysis

The strongest descriptive result is concentration. Technology firms account for roughly two-thirds of deeply integrated enterprise adoption, despite being only a subset of the index. Non-technology adoption is rising, but more slowly. That pattern supports a narrower reading than many macro-level AI forecasts imply: deep AI use is spreading through large firms, but the current diffusion frontier still sits heavily inside the technology sector.

The firm-outcome evidence is more mixed. The paper reports a profitability “J-curve” as firms move from no adoption toward deeper adoption: profitability initially weakens, then improves at deeper adoption levels. That is consistent with an implementation-cost story, where firms absorb organizational and integration costs before seeing gains. It is not, by itself, proof that AI causes higher profitability, because adoption is not randomly assigned and higher-capability firms may be more likely to adopt.

The capital-expenditure and productivity results are more restrained. Figure 4 plots mean capex-to-revenue ratios by AI adoption level from 2016 to 2025 and finds no clear monotonic relationship in either technology or non-technology firms. The authors interpret this as evidence against the view that most firms are building large in-house AI infrastructure. A more plausible reading is that many enterprises are consumers of externally supplied AI infrastructure rather than owners of it.

Regression results add a second layer. Among technology firms, adoption is higher for firms with more employees and higher Tobin’s q, but these relationships do not appear in the same way outside technology. The paper also reports no clear productivity differences across adoption groups. For practitioners, that is the central takeaway: the measurement strategy gives a useful lower-hype view of enterprise AI diffusion, but the outcome evidence is still associational and early.

Limitations

The paper is strongest as a measurement and descriptive-diffusion study. Its use of 10-K filings reduces marketing noise, but it also misses internal AI use that firms do not disclose as material. The adoption categories depend on text signals, so classification quality is tied to wording choices in corporate filings. The outcome analysis is informative, but it does not establish causal effects of AI adoption on profitability, capex, or productivity. The evidence supports the claim that deep adoption remains limited and concentrated; it does not yet support a broad productivity-impact claim for the full S&P 500.