Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
AI Adoption in S&P 500 Firms
Citation. Yu, Y., Fleming, M., Hampton, L., Combemale, C., & Thompson, N. (2026). AI Adoption in S&P 500 Firms. arXiv. https://doi.org/10.48550/arXiv.2607.08920
Summary. This preprint develops a five-level rubric for enterprise AI adoption from AI-related text in S&P 500 firms’ SEC 10-K filings, distinguishing exploratory references from production use and deep business-process integration. It combines those scores with Compustat financial data to describe adoption from 2016–2025 and examine associations with firm outcomes. In 2025, 21% of firms scored as using AI in production or deeply integrating it, while advanced adoption was concentrated in the technology sector.
Key claims
- The authors classify 4,463 AI-adoption scores for 510 firms across 4,561 firm-year observations from 2016–2025 using 10-K text, a GPT-5-mini classification procedure, and manual rubric validation and refinement.
- In 2025, 11% of S&P 500 firms scored as deeply integrating AI in business processes and another 10% as using AI in production; the combined 21% share was up from 5% in 2022.
- Firms at the production-use and deep-integration levels show a J-curve association with net profit margins, but the study finds no significant association between adoption and revenue per employee or capital expenditure.
- Technology firms accounted for 67% of firms scoring at the two most advanced adoption levels in 2025, and the authors report more advanced adoption among technology firms with more employees and higher Tobin’s q.
Evidence & limitations
- The study relies on public SEC 10-K disclosures and Compustat data; the rubric is designed to capture enterprise deployment in business processes rather than stand-alone generative-AI tool use.
- Adoption classifications depend on firms’ disclosure language and an LLM-assisted procedure, so nondisclosure and rubric decisions may affect the measured prevalence and comparisons.
- This is an arXiv preprint, not peer-reviewed; the public arXiv v1 PDF is text-identical to the ingested Drive copy but has a different binary SHA-256.