AI builders are beating the AI doomers, and 418 annual reports show it

Every few months someone with a large following announces that AI is overhyped. It writes bland emails, it gets facts wrong, it saves them nothing. They are usually telling the truth about their own experience. What they are describing is themselves.

"Overhyped" is a review of your own workflow, not of the technology.

A tool that answers in proportion to what you bring it is a mirror. Bring it a vague question and you get an average answer. Bring it a well-built workflow, a sharp brief and the judgement to check its work, and it multiplies you. Calling it hype is a review of your own workflow.

That is a claim about individuals, and individuals don't publish data. Companies do. Every year each S&P 500 company files an annual report with the SEC, and nearly all of them now say something about AI. Some describe it as something they build with. Others only list it as a threat. So we read the AI passages in those reports and followed the share prices.

We call the first group AI builders and the second AI doomers. Since March 2025 the typical builder's shares have risen 24.8%, against 11.1% for the typical doomer.

Five hundred mirrors

We took the 418 S&P 500 companies that filed their annual report between January and June 2025 and found every mention of AI, machine learning, generative AI and large language models. Then we looked at where in the report it sits.

A company that writes about AI at least twice in its business description or in management's discussion of results is a builder. Chevron is one: "Chevron is applying artificial intelligence (AI) to drive productivity, efficiency and value to its global operations." A company that mentions AI only in its risk factors is risk-only. 3M is one: "The development and use of artificial intelligence may present risks to the Company's business." A company that never mentions it is silent.

The builders are not just the tech crowd. C.H. Robinson, a freight broker, writes that it is "expanding the use of generative AI in our industry, creating proprietary technology to perform work that defied automation for decades." Southwest Airlines tests its marketing copy with generative AI. Travelers has put AI into claims handling, Xcel Energy uses it for wildfire mitigation, and AES built Maximo, an AI-powered robot that installs solar panels. The risk-only side reads like one template: Chipotle, Domino's, UPS and Kimberly-Clark all wrote, almost word for word, that AI "may intensify our cybersecurity risks."

In reports filed in 2025, 48.1% were builders, 38.8% were risk-only and 13.2% were silent. A year later the builders were 61.6%, and only 4.5% stayed silent. Almost everyone now names AI as a risk (90.9%). Saying it can hurt you has become boilerplate. Saying how you use it has not.

The scoreboard

From 31 March 2025 to 12 October 2026, the typical builder's share price rose 24.8%. The typical risk-only company rose 11.1%, and the median S&P 500 stock rose 17.1%. Measured from the day after each report was filed, so no hindsight goes into the grouping, the gap is still 8.7 points.

The obvious objection is that builders are just tech companies in a tech boom. They aren't only that. Take out every chip, hardware and software company and builders still rose 21.2% against 11.0%. Within sectors, builders beat their own sector's risk-only companies in 9 of 11. Retailers and restaurants that wrote about using AI rose a median 32.6% while their risk-only peers went nowhere: Best Buy rose 28% and Ralph Lauren 72%, while Chipotle fell 39% and Domino's 30%. Industrials split 26.1% to 11.1%; C.H. Robinson rose 40% while UPS fell 4%. In household staples, Walmart (+31%) and Target (+57%) wrote about using AI; Kimberly-Clark (-26%) and Kraft Heinz (-19%) only listed it as a risk.

We also sorted the companies a second, independent way: whether the report makes a concrete business claim about AI, read clause by clause. That test gives an even wider gap, 27.2% against 11.0%, and builders won in all 11 sectors. Where both tests agree, it is 29.5% against 11.1%. A gap this size between groups this large would arise by chance about once in a hundred tries.

Where it held and where it didn't

It is a tilt, not a law. Lowe's and Tractor Supply wrote about using AI and fell 18% and 37%. General Motors and Ross Stores only mentioned it as a risk and rose more than 75%.

Not every sector played along either. Energy companies that only feared AI did better than those that wrote about using it. The builders did fine (Chevron +37%, Baker Hughes +31%, Halliburton +33%), but refiners Valero and Marathon Petroleum, which mention AI only as a risk, more than tripled on fuel margins that have nothing to do with software.

Elsewhere the pattern held with names attached. In utilities, AES (+31%), which built a solar-installing robot, beat Southern Company (-1%), which mentions AI only in its risk factors. In property, Welltower (+54%), which trains senior-housing managers with AI, beat Public Storage (+3%). In finance, Travelers (+44%), which put AI into claims handling, beat KKR (-20%), which says AI could raise risks "in ways that we cannot predict".

Healthcare was a tie, and it holds the clearest counter-examples. Merck and Eli Lilly describe real AI use, but inside their risk factors, so the keyword test marks them risk-only. Their shares rose 70% and 45%. Newmont, the gold miner, does the same with its autonomous operations and rose 147%.

Saying it is not doing it

Here is the finding that fits the mirror best. Sixty companies that were risk-only or silent in 2025 rewrote their 2026 report to describe building with AI. Home Depot now uses AI to improve search, recommendations and chat. Delta uses it to time maintenance. Duke Energy says AI "is being leveraged across the organization". PACCAR, General Motors, UPS, NextEra and PayPal made similar additions. If the market rewarded the words, those companies should have done well. They didn't, and the names show why the words predicted nothing: Delta rose 87% over the period while Home Depot fell 17% and PayPal 13%. Over the whole period they rose 13.1%, barely ahead of the companies that never changed (10.2%), and since the new reports were filed they are down 1.4% against 1.8% for those that didn't change.

Investors are not paying for the vocabulary. They are paying for evidence that the work has changed, and a paragraph added to an annual report is not that evidence. The same is true of a person who adds "AI-powered" to a CV.

The dynamo lesson

If AI were already paying off in the accounts, the builders' profits would show it. So far they don't. Builders grew revenue 7.5% in their latest financial year against 5.2% for risk-only companies, but they were already ahead the year before (6.7% against 4.5%). Risk-only companies still have higher operating margins, and widened them slightly more.

The motor was never the problem. The floor plan was.

That is what the early years of a general-purpose technology look like. Electric motors were installed in American factories from the 1890s, yet for two decades factory productivity grew slowly. Many owners had simply put a motor where the steam engine stood and kept the same shafts, belts and layout. The economic historian Paul David showed that the payoff came in the 1920s, once factories were redesigned around small motors at each machine. In 1987 Robert Solow wrote that "you can see the computer age everywhere but in the productivity statistics". The boom arrived a decade later. Today's version of the motor bolted onto the old floor plan is the 2026 annual report that adds an AI paragraph without changing the work.

Kodak is the warning on the other side. One of its engineers built the first digital camera in 1975. The company treated it as a threat to film, kept it in the risk column, and filed for bankruptcy in January 2012.

Our call

The share prices say investors already believe the builders will win. The accounts say the work is still being redesigned. We think the redesign keeps going, which fits our earlier call that Taiwan's AI exports cool in growth terms by March 2027 but stay near record levels.

Our call: between 12 October 2026 and 12 October 2027, the median S&P 500 company whose 2026 annual report describes building with AI beats the median risk-only company by at least 5 percentage points. We put that at 70%.

• Builders stay ahead by 5 points or more (70%). The redesign shows up in margins in the 2027 reports and investors keep paying for it.
• The gap shrinks below 5 points, builders still ahead (20%). Prices have run ahead of profits and the market waits for the accounts.
• Risk-only companies come out on top (10%). A 2000-style hangover in AI spending punishes the heaviest spenders first.
• Who to watch: Kimberly-Clark, Union Pacific and Waste Management still describe AI only as a risk in their 2026 reports. If they start writing about using it, and their margins follow, the laggards are catching up.
• What to watch: builders' operating margins in reports filed early in 2027, and the builder share of new reports. Above 70%, the laggards are running out of time. Below 65%, adoption has stalled.

What could make us wrong

Read these results as a strong pattern, not proof. Four caveats matter.

• Today's index. We used the companies in the S&P 500 now. Firms that rose into it since early 2025 are included and firms that fell out are not, which flatters winners, builders most.
• One cycle. Our filing data starts in late 2024, so this covers two reporting seasons and one rising market, not a full cycle.
• Words, not proof. Where AI is discussed is a proxy for use. Some builders' AI mentions are about demand or regulation rather than use, such as Caterpillar (data-centre demand) and JPMorgan (new rules), and some risk-only companies, such as Merck, Lilly and Newmont, clearly use it.
• Averages mislead. A handful of names rose several hundred percent, so we report medians and how many companies beat the median stock (56% of builders, 43% of risk-only).