The missing number: AI's return in 498 annual reports

What did the study find?

Not one of 498 S&P 500 annual reports states an achieved, quantified benefit from the company's own use of AI. Eleven companies do it in their earnings releases, and none of those eleven gives a dollar figure for AI alone.

The question was narrow. Does a company put a number on a benefit from its own AI use, and say it has already happened? AI products it sells don't count, and neither do forecasts.

	Annual report	Earnings release
Achieved and quantified	0 (0.0%)	11 (2.2%)
Quantified but forecast or target	1 (0.2%)	12 (2.4%)
Any quantified AI value figure	1 (0.2%)	20 (4.0%)
Achieved dollar figure attributed to AI alone	0	0

A company can have both an achieved figure and a forecast, so the first two rows don't add up to the third. The one annual-report figure is a forecast: Sysco's roughly $100 million of expected fiscal 2027 efficiency improvements.

The coding was done by a language model, Claude, working to a written rubric under the direction of Roiva's founder, Luke Wilcox, who settled disagreements and ruled on every rubric change. To test the model, Wilcox blind-coded 30 companies' 10-Ks himself, and 18 of those companies' earnings releases. The later recall passes used independent model instances that could not see the existing codes.

What counts as attributing a figure to AI?

The filing has to credit the figure to AI itself, by definition or by a causal phrase. Being nearby doesn't count.

In this study, a figure is attributed to AI only when the thing it is credited to names AI in its own definition, or when the text joins AI to the figure with a causal phrase such as "driven by" or "enabled by."

Quest Diagnostics shows why. Its 10-K says its Invigorate program "has consistently delivered 3% of annual cost savings and productivity improvements." The next sentence says, "We are leveraging automation and AI to improve quality and productivity across our entire value chain, not just in the laboratory." But the filing defines Invigorate by patient services, logistics, laboratory operations, revenue services, IT and procurement. No AI, and no phrase linking AI to the 3%. The MD&A and the financial-statement note both say the program "aims to deliver" 3%. Quest had been the only achieved case in the annual reports. This rule removed it and took the count to zero.

The same rule removes the best near-miss. Hasbro's 10-K reports almost $800 million of gross cost savings through 2025 from a program called Operational Excellence. The nearest AI term is 1,266 characters away, in a growth paragraph about AI-enabled entertainment. Nothing ties the two.

Can companies count AI at all?

Yes, when AI is something they sell. Companies report AI product revenue in both documents, more often in releases. That coding is provisional, and no finding here rests on it. The gap is in the return on their own use.

Can a 10-K carry a savings figure at all?

Yes. A screen for quantified savings tied to a named program, with AI not required, found them in more than forty of the 498 annual reports, and none of them named AI in the same sentence.

That screen was looser than the AI test. It counted targets and realized amounts together, while AI figures had to be achieved. But some are achieved. PPG reports about $75 million of restructuring savings realized in 2025. Hasbro's $800 million is achieved too. None of the forty-plus credits AI.

What kind of number do companies report when they report one?

When AI gets the credit alone, a percentage. Five of the eleven credit AI alone, and all five figures are percentages. The other six bundle AI with other levers.

Company	Figure	Document
Amazon	"improving robot travel efficiency by 10%"	Q2 2025 release
Airbnb	reduced the need to contact a human agent by "around 15%"	Q3 2025 shareholder letter
Bank of America	"helping to reduce calls to service desk by ~50%"	Q2 2025 earnings exhibit
Broadridge	"25% efficiency gains on headcount through Agentic AI"	April 2026 earnings presentation
Revvity	"~10% reduction in software development timelines"	JPMorgan 2026 conference deck, filed with a results 8-K

These eleven are not general practice. AI terms appear in 483 annual reports and only 286 companies' release sets. Releases discuss AI less overall, and the quantified claims sit in a small group.

What do companies say AI costs?

They put numbers on it. 38 companies quantify what AI costs them in at least one document: capital spending, AI research and development, acquisition prices, restructuring charges.

	Quantify an AI cost	Quantify an achieved AI return
Annual report	26	0
Earnings release	19	11

Seven do it in both. The cost count is a floor.

Ecolab's release puts both side by side: about $425 million of expected pre-tax charges for the program, mostly severance, and "total annualized savings of approximately $325 million by 2027 as the leverage of AI and digital technologies drive stronger benefits." That is a forecast. Its annual report attaches no savings figure to AI.

Why is the number missing from the annual report?

The data can't say. It narrows the explanations but can't separate the ones left.

Three were open at the start. The value isn't measured. It's measured and not put in the annual report. Or disclosure habits produce savings figures only around restructurings, and most AI spending isn't one.

Four companies ran an AI program through a restructuring, the vehicle that usually carries a savings figure, and none of them gave a savings figure for it in the annual report. That counts against the third explanation, but only a little: three of the four give no savings figure for anything in their annual reports, so their silence is partly house style. The third explanation is weakened, not ruled out. The first two remain, and this data can't tell them apart.

How was the study done, and where is it weak?

The population is the S&P 500 as of September 30, 2026: 498 companies. The filings were collected from EDGAR programmatically: each company's latest 10-K filed by then, and every results 8-K for the same fiscal year with all exhibits.

Claude coded every company on a five-tier ladder. The rubric changed 14 times, each change logged. Wilcox's blind coding came in two draws. The first gave a kappa of 0.707 on tier. In the second, agreement on companies' own AI use across both documents was 26 of 36, a kappa of 0.553, which misses the 0.6 threshold set in advance. At the achieved-and-quantified line, where every finding here sits, agreement was complete: 72 of 72 judgments. Six independent model instances then re-read every long annual-report passage blind and found no achieved AI benefit.

The study covers one fiscal year and the written record only. Earnings calls aren't in it.

The strongest objection is PPL. Its earnings release reports $170 million of cumulative annual O&M savings in 2025 against a 2021 baseline, credited to grid hardening and to smart grid technology, automation and data science. It is excluded only because "automation" was set aside as a pre-wave term before any data was seen, and "data science" sits on the same side of that line. The release's only AI term is in its forward-looking boilerplate. Count those terms and PPL becomes a release case, and other filings that credit automation might too; the study didn't code them. PPL's own figure isn't in its 10-K and is bundled, so it doesn't change the annual-report zero or the bottom row of the table.

Roiva is AI portfolio management software that tracks every AI initiative's budget, approvals, cost and ROI. This is a finding about a problem Roiva sells into, which is a reason to check the work. The methods note has the detail, and the per-company data is available on request.