SEC Filings Show Enterprise AI Adoption Remains Concentrated

Enterprise AI adoption is easy to overcount when every chatbot rollout is treated as transformation. This paper takes a narrower view: AI counts as deep adoption only when firms describe integration into business processes, production, or service delivery in SEC 10-K filings, where materially false or misleading statements carry legal risk. The result is a descriptive map of how far AI has moved inside large public companies from 2016 to 2025, and where that movement has not yet shown up in firm-level financial measures.

Core Contribution

The main contribution is a firm-year measure of AI adoption for S&P 500 enterprises. The authors score filings on a five-point scale: 1 for no mention, 2 for exploration, 3 for pilot phase, 4 for use in production, and 5 for deep integration. That distinction matters because the paper is not measuring employee-level use of generative AI tools. It is trying to identify AI embedded in the execution of business processes, such as production systems, service delivery, supply-chain operations, or sector-specific workflows.

Figure 2 shows the core adoption pattern: AI mentions and pilots spread widely after 2022, but deep integration remains concentrated, especially in technology firms.

Technical Approach

The dataset covers S&P 500 firms over 2016–2025 and links the AI score to industry group, headcount, net profit margin, capex-to-revenue, Tobin's q, and revenue per employee. The appendix lists AI-relevant keywords used to identify filings, including terms such as artificial intelligence, machine learning, computer vision, natural language processing, neural networks, transformers, recommender systems, support vector machines, random forests, XGBoost, reinforcement learning, and object detection.

The empirical design is mostly descriptive. The tables use firm-year observations and regress outcomes on adoption levels, with specifications that add firm fixed effects, year fixed effects, and sector-by-year fixed effects. That helps separate persistent firm differences and common time shocks from adoption-stage patterns, but the authors explicitly avoid causal claims. The paper reads the coefficients as associations between adoption stage and firm characteristics or outcomes.

Results and Analysis

By 2025, 11.4% of S&P 500 enterprises were scored as deep integration and another 10.0% as used in production. The broader adoption funnel is much larger: 45.2% were in pilot phase, 15.3% were exploring, and only 18.1% had no AI mention. Technology firms are the outlier. In 2025, 50.0% of technology firms were in deep integration, versus 3.8% for financial firms and 5.7% for all other firms. The paper also reports that deep adoption among non-technology firms rose from 1% in 2022 to 5.7% in 2025, so diffusion outside technology is real but still early.

The financial results are more mixed than the adoption curve. Profitability follows a J-shaped pattern: firms in early integration show 1% to 3% lower profitability than non-adopters, while firms deeply embedding AI show around 6% higher profitability. Figure 3 makes the sector split clearer. Among technology firms, those with no or very limited adoption average 14.4% net profit margin, pilot-phase firms average 12.4%, and deep-adoption firms average 17.4%. For non-technology firms, the deep-adoption estimate has a wide confidence interval because the category contains only 19 firms, so that comparison should not carry much weight.

The negative result is as useful as the positive one. The authors do not find a clear relationship between AI adoption and capex intensity across the full S&P 500, and they report no meaningful productivity gains when productivity is measured as log revenue per employee. Their interpretation is plausible: many firms buy AI model services through APIs or cloud platforms rather than building capital-intensive AI infrastructure, and task-level productivity gains may be blocked by organizational frictions, non-AI bottlenecks, or lagged workforce adjustment.

Caveats

The paper is strongest as a measurement exercise, not as proof that AI causes higher margins. SEC filings are a disciplined source, but they can lag internal deployment and they reflect what firms choose to disclose. The sample is also limited to large public firms, with technology firms carrying much of the deep-adoption signal. The most defensible takeaway is therefore narrower than a productivity story: by 2025, enterprise AI had moved beyond experimentation in a minority of S&P 500 firms, but the measurable gains were concentrated in profitability associations, not in capex or revenue-per-employee productivity.