What do companies say about AI?

1. U.S. Securities and Exchange Commission, EDGAR: the primary document of each annual report on Form 10-K filed from 2016 through September 2026 for the 25 companies, plus Exhibit 13 (annual report to shareholders) where a company files its MD&A there (IBM, Wells Fargo), sec.gov/edgar. Public domain.
2. U.S. Securities and Exchange Commission, XBRL company facts API, data.sec.gov/api/xbrl/companyfacts, values tagged in the same 10-K for its own fiscal year: revenue (the largest of the standard us-gaap revenue elements; for banks, RevenuesNetOfInterestExpense, or Revenues where that is not tagged), NetCashProvidedByUsedInOperatingActivities, and total capital expenditure (PaymentsToAcquirePropertyPlantAndEquipment, or PaymentsToAcquireProductiveAssets where a company uses that element). Public domain.

Method. Sections. Each 10-K is converted to text with paragraph breaks kept and split into Item 1 (Business), Item 1A (Risk Factors) and Item 7 (Management's Discussion and Analysis). For each item the heading that gives the longest span to the next item heading is used, which skips the table of contents. Where the MD&A sits outside Item 7 (JPMorgan Chase, Bank of America, Deere) or in Exhibit 13 (IBM), that text is used and labeled. Intel and Wells Fargo use layouts in which the items cannot be separated reliably, so their counts cover the whole filing and are labeled that way. Mentions. The phrase "artificial intelligence" in any capitalization and the standalone capitalized acronym "AI", divided by the word count of the section and multiplied by 10,000. Passages. Every paragraph in those sections that contains a mention (long paragraphs are trimmed around the first mention), deduplicated within a filing: 3,203 passages from 269 filings. Passages are quoted as filed, except that dashes are shown as hyphens and trims are marked with an ellipsis; the link opens the full original text. Context. Each passage was labeled by a language model (Google gemini-flash-lite-latest, temperature 0) as a product or revenue opportunity, internal operations, infrastructure investment, a risk disclosure, or other. As a check, a second model (Google gemini-3.8-flash) labeled a stratified random sample of 250 passages (50 per label) without seeing the first labels. The two agreed on 46 of 50 product labels and 45 of 50 risk labels, but on only 16 to 25 of 50 for infrastructure, internal operations and other (59.6 percent raw agreement across the stratified sample, Cohen's kappa 0.50; about 82 percent when weighted by how common each label is). Every disagreement was read by hand, and the second model's label was usually the better fit, mostly because the first model labeled product descriptions as infrastructure or internal operations. All 562 passages the first model had placed in those three rarer labels were therefore re-labeled by the second model, and that label is used. A final hand check of 60 randomly drawn passages from 2024 and later filings agreed with 55 labels; the errors were mainly risk-factor passages about products labeled as product opportunities. Context labels are estimates, not company statements. Folded labels. To keep the chart readable, the two rarer labels are folded into existing ones: passages labeled infrastructure investment are counted as a product or revenue opportunity, and passages labeled internal operations are counted as other. In the latest filings shown on this page that moves 50 of 850 passages (27 infrastructure, 23 internal operations), so the page shows three labels: product or revenue opportunity, risk disclosure, and other. Explicit AI investment. Shown only when a filing states a dollar amount attributed to AI alone; both models flagged candidate passages and each was reviewed by hand. Amounts that combine AI with other purposes are quoted in a note and not counted. Headcount. The cover page tag where filed, otherwise a company-wide headcount sentence in the filing, read and verified by hand; counts of members, subscribers, customers or covered lives are never used. Fiscal years are labeled by their end date. Capital expenditure covers all purposes, not only AI. Banks do not report capital expenditure in a standard line, and their operating cash flow reflects lending and trading and is not comparable with other companies. No combined or synthetic score is calculated.