45% of traditional US public companies mentioned AI in their 2025 annual reports

45% of traditional-sector US public companies named artificial intelligence in their 2025 annual reports. 14.5% described a specific AI capability in operation. Fewer than 1 in 200 attached a number to a result.

We read every 2025 annual report (Form 10-K) from a traditional-sector, non-tech US public company that contains the phrase “artificial intelligence”, 2,112 companies in all, and graded each mention on a 0-to-5 scale from an incidental reference up to a quantified result. This page reports what we found, with the full coded data below. It measures how companies describe AI to investors, not how much AI they run.

What the mentions contain

Most sit at the low end of the scale, and the production numbers depend on where you set the bar, so we report both settings. Under the strict definition, where “in production” requires a specific described capability, 14.5% of companies qualify. Another 17% claimed to use, or plan to use, AI without naming a single system or function; count those bare claims as production and the figure rises to 19.3%. Either way, a measured outcome is the rarest thing in the census: eight companies, 0.38%.

Boilerplate or incidental Aspiration or unspecified use Named pilot In production Quantified result

View this chart as a table

companies	Boilerplate or incidental	Aspiration or unspecified use	Named pilot	In production	Quantified result
Strict definition (n=2112)	62%	21%	2.6%	14%	0.4%
Original definition (n=2112)	57%	20%	3.1%	19%	0.4%

To make the scale concrete, here is each end of it and the middle. At the low end, boilerplate: The St. Joe Company notes only that artificial intelligence “may be used to identify vulnerabilities and craft increasingly sophisticated cybersecurity attacks”, a risk, not something the filer runs. In the unspecified middle: American Express says it utilizes AI and machine learning models “for a variety of business purposes”, a real claim of use that names nothing. In production under either definition: Northern Trust “uses a variety of machine learning and artificial intelligence solutions to process transactional activity more efficiently and to mitigate risk”, a described function with no number attached. Quantified, the rarest tier: Hologic reports its imaging AI “saving an estimated average of one hour per eight hours of daily image interpretation time”.

Not all production claims are the same kind of claim. Of the 306 companies describing a specific capability in operation, 49% put the AI in their own operations, 38% in products they sell, and 13% in both. AI running inside a traditional company’s own operations, specifically described, comes to about 9% of all the AI-mentioning companies in the census.

Own operations Product they sell Both Third party

View this chart as a table

group	Own operations	Product they sell	Both	Third party
Companies with a specific AI capability in operation (n=306)	49%	38%	13%	0.3%

Two more features of the language. AI turned up in the risk-factors section (Item 1A) in 76.5% of the filings that mentioned it. And 11.0% named a specific system, tool, or vendor; the rest kept it general.

The language also repeats from filer to filer. About one in seven filings share at least three-quarters of their AI wording with another company’s: the recycled risk-factor paragraphs that move through law firms and filing templates. Most companies still write their own, but a sizable minority lean on shared boilerplate.

How many mention it, by sector

Across every traditional-sector 10-K filed in 2025, 45.0% named AI. For tech-sector filers the figure was 77.4%. Within the traditional group the range is wide, from insurance at the top to banks and lenders at the bottom.

View this chart as a table

sector	share
Insurance	83%
Tech (comparison group)	77%
Transport & logistics	71%
Telecom & media	66%
Retail & wholesale	55%
Manufacturing & industrials	55%
Healthcare	54%
Real estate & REITs	49%
Energy & utilities	48%
Pharma & life sciences	48%
Banks & lenders	23%

What this shows and what it does not

It measures disclosure. A 10-K is what a company chose to tell investors under legal constraint, graded for how concrete the AI language is. It is not an audit of what companies run, and three things keep it from being read that way.

A small spot-check suggests a 10-K can understate real activity. Of 12 companies whose filings describe AI in production, seven had a specific AI result in public sources (an investor day, a newsroom, trade coverage) that never appeared in the 10-K. Take two. JPMorgan told its 2025 investor day that AI and machine learning “delivered a 35% increase in value last year” and cut manual exceptions by more than half. Medtronic’s newsroom reported a 9% reduction in false positives from an AI-assisted polyp-detection tool. Neither number is in the company’s 10-K. Both are the company’s own claim, not independently verified. Twelve companies cannot support a rate, but they show that absence of production language is not evidence of absence.

The sparse quantification is not specific to AI. In the same filings, the outcomes of cybersecurity, cloud, automation, and cost programs were quantified about as rarely. The near-zero rate is a fact about how annual reports are written.

And the frame is public companies, which skew large. The Census Bureau’s Business Trends and Outlook Survey, which samples US employer firms of all sizes, found 37% of firms with 250 or more employees currently use AI, against under 20% economy-wide. So this set likely runs ahead of the wider economy, not behind it. It also sits well below survey figures like McKinsey’s 88%, because a 10-K mention and a survey’s “do you use AI anywhere” are not the same measure.

How we coded it

Two independent models, Claude Sonnet 5 and GPT-5.5, graded each filing’s AI passages against a fixed rubric, and Gemini broke the ties. They agreed 86% of the time (Cohen’s kappa of 0.79). A human review of 67 stratified filings confirmed 94% of the panel’s labels, and it also exposed the census’s one unstable boundary: filings that say “we use AI” without naming anything sat between tiers depending on the reading. So we drew the line explicitly. The whole census was re-coded under a sharpened rubric where “in production” requires a specific described capability, every filing where the two rubrics disagreed about production status was resolved by a fresh three-model majority, each model coding blind, and the report gives both numbers. A second human review of 30 of those resolved boundary cases confirmed all 30. Both the original codebook and the sharpened one are downloadable, and the dataset carries every label, for every filing, under both, so anyone can check the counts or draw the line differently.