Reading which companies use AI, not which ones talk about it
The Competition: four of us built an entry to Georgia State University’s Southeastern Hedge Fund Competition, and I owned the model. It scored how far each S&P 500 company had put AI to work inside its own business, read from its earnings calls and filings. The score used only documents through the end of 2025. Over the first eight weeks of 2026, the top quartile, the quarter of companies it ranked highest, beat the bottom quartile by 11.4%.
Self-Computed The team computed every performance figure here from Bloomberg data, and no outside party has checked them.
What this Engagement produced
-
487 companies scored
Every S&P 500 company the pipeline had documents for, each given one score between 0 and 1. -
5,548 documents read
3,800 earnings calls and 1,748 SEC filings, read sentence by sentence. -
3 years tested
2024, 2025 and the start of 2026, all three reported in the paper, including the two in which the score lost.
The Situation
A hedge fund case competition asks for a formal paper: an investable idea, with enough evidence that somebody can argue with it. Early in 2026, nearly every such idea bet on who sells AI, meaning the chips, the models and the clouds they run on. Far fewer were about who had actually put it to work.
That gap stays open because the evidence is awkward. Companies say a great deal about AI in their filings and on their earnings calls, and most of it is positioning. So we asked whether the documents alone can tell a company running AI in its operations from a company with an AI paragraph. Then we asked whether the market pays for the difference.
Why the paper looked at companies using AI rather than selling it
Sellers’ gains show up as sales right away. Users’ gains show up later, as lower costs, mixed in with everything else on the income statement.
That delay is where an outside investor can still act. Once a user’s savings reach its margins, everybody can see them, but its earnings calls and filings describe the work first. What Changed shows how that argument held up in the numbers.
What I Did
I owned the model, from how each sentence was scored to the final score for each company. We split every part of the paper into pieces and passed them around. So I also worked on the financial modeling, the index construction and the presentation, and I co-wrote the final submission. Every piece, the model included, was read by someone who had not written it. My three teammates are not named here because they have not approved a public write-up.
The team’s pipeline reads every earnings call since 2024 and every SEC filing since 2023 for 487 S&P 500 companies. It pulls out each sentence that ties AI to a financial result, such as a cost saved, a margin improved or work done with fewer people. It sorts each one into one of 12 operational categories. Claude Opus, Anthropic’s large language model, then scores each sentence from 0 to 6. A 0 is buzzwords; a 6 is a CFO citing dollar figures across several deployments.
The score weighs four things: whether a company uses AI or sells it, how often it puts a number on the result, how hard analysts question its AI claims, and whether its AI talk is speeding up. These are joined with Bloomberg figures on margins, SG&A (overhead costs) and operating leverage (how much profit grows as sales grow). An ensemble, four machine-learning models whose outputs are combined, turns them into one score between 0 and 1 for each company.
The ensemble learned from 34 labeled companies: 20 AI winners named by Goldman Sachs and 14 software companies UBS expects AI to hurt. Both banks published those lists from their own research, so the labels came from outside the team.
Exhibit 1 More AI mentions don’t raise the score. A dollar figure does.
Move the slider to write more mentions of AI into every sentence.
View the data
| Sentence | Text | Mentions of AI | What the AI is for | Dollar or percentage result | Rubric score, 0 to 6 |
|---|---|---|---|---|---|
| Sentence A | Our AI platform brings AI to every enterprise client, and AI will define the next decade of their business. | 3 | A product it sells | No | 0 |
| Sentence B | We are exploring AI across our operations and expect it to make us more efficient over time. | 1 | Its own operations | No | 1 |
| Sentence C | The AI claims tool we built now makes the first review of every claim, so our review staff handle only the exceptions. | 1 | Its own operations | No | 3 |
| Sentence D | Our three AI deployments, in routing, billing and claims, cut our cost per order by $3 this year, or 6%. | 1 | Its own operations | Yes | 6 |
Before any of that, I read every prior year’s winning entry to work out what the judges actually rewarded.
Exhibit 2 A judge has 10 minutes for one paper
Move the slider to change how many minutes the judge has.
View the data
| Part of the paper | Minutes a judge needs to check it | Built to impress | Built to be checked |
|---|---|---|---|
| Thesis, in one sentence | 1 | No | Yes |
| Thesis, argued at length | 8 | Yes | No |
| Headline backtest spread | 3 | Yes | No |
| Top-minus-bottom spread in every period tested | 3 | No | Yes |
| Spread by sector | 4 | Yes | No |
| Spread by company size | 4 | Yes | No |
| Spread at other score cut-offs | 6 | Yes | No |
| One passage scored, worked through | 3 | No | Yes |
| Where the score is weakest | 2 | No | Yes |
| Minutes a judge has for one paper | 10 |
Reading the prior winners before writing a word
The entries that had won in earlier years did not have the strongest backtests. Their thesis fit in one sentence, and a judge could check their evidence in the time a judge actually has.
So we produced fewer exhibits, each able to stand on its own, and every figure in the abstract states the period it was measured over. This was the most transferable thing I did here: finding out what the audience rewards before spending six weeks producing for it.
How the four things are measured, and the cap on AI sellers
Using or selling. “AI cut our cost per package 8%” is a company using AI. “Our AI platform serves enterprise clients” is a company selling it. A company whose AI sentences are more than 70% the selling kind has its score capped at 0.35 out of 1. That keeps chipmakers and cloud providers, whose prices already reflect the enthusiasm, out of the top of the ranking.
Numbers. Sentences with a dollar or percentage result count for more than general claims of efficiency.
Analyst pressure. Each earnings call is split into management’s prepared remarks and the analysts’ Q&A. A wide gap between management’s AI claims and how hard analysts follow up on them suggests promotion.
Speed-up. The pipeline tracks whether a company’s talk about AI is growing faster quarter over quarter. That picks out companies early in deployment, before the savings reach their margins.
Limits. Banks, insurers and REITs (real estate investment trusts) are scored on their documents alone, because the usual margin measures do not fit how they make money. The ensemble learned from only 34 labeled companies, so the paper’s main evidence is the returns that came after scoring.
What Changed
The model was built only from documents through the end of 2025. So every 2026 return is out of sample: a period the model never saw while it was being built. Across the 487 companies, the top quartile returned 9.9% from the start of January to February 27, 2026, while the bottom quartile lost 1.5%. That is a spread of 11.4%. The equal-weight S&P 500, where every company counts the same, returned 3.5% over the same weeks.
The pattern held across the whole range of scores, not only at the ends. The correlation between a company’s score and its 2026 return was +0.296.
Run on earlier years, the same score pointed the other way. The top quartile trailed the bottom by 18.4% in 2024 and by 11.9% in 2025. The paper reports both years in full.
Exhibit 3 The same score lost in 2024 and 2025, then won in 2026
Switch to the chart a paper showing only 2026 would have published.
View the data
| Period | Top-minus-bottom quartile spread | Direction |
|---|---|---|
| 2024, full year | −18.4% | Top quartile behind |
| 2025, full year | −11.9% | Top quartile behind |
| Jan 2 to Feb 27, 2026, out of sample | +11.4% | Top quartile ahead |
The paper explains those two years as a lag. Companies that put AI to work in 2022 and 2023 had already been rewarded by the market before our documents begin. Their deployments were finished, so they said little about AI in 2024 and 2025. They scored low while their stock did well. By 2026 a new group was in the middle of deploying. These companies described AI in their operations in specific terms, and the savings were not yet in their earnings. That is the group the score ranks highest.
The financials fit that explanation. Across 350 non-financial companies, a company’s score had almost no relationship with its change in operating margin over 2024 and 2025 (r = −0.070). The paper expects this, because companies deploying AI are paying for it now and have not yet banked the savings. The market is paying for them anyway, and that gap is what the strategy trades.
Two more checks point the same way. First, 453 of the 487 companies never had a label. On those, the companies the model rated Buy returned 9.2%, against a 6.6% average for all 453. So the score is finding the pattern across the index, and not memorizing the 34 it learned from.
Second, a portfolio holding Goldman’s AI winners and shorting UBS’s likely losers went from −50.4% in 2024 to +13.9% in 2025 to +32.6% in early 2026. That is the same arc as the score. The paper is plain that eight weeks is a short live window. A full year of 2026 returns is the real test.
This model is separate from the pipeline I later withdrew in The Research. It measures a different thing in different documents, and it was built months earlier.
What I Learned
First, report the periods that went against you. The score lost in two of the three years it was tested on, and the paper shows all three.
Why the losing years went into the paper
Leaving 2024 and 2025 out would have made a cleaner paper. But a judge who found the losing years on their own would have found a hole. A judge handed them with an explanation is reading a paper that already knows where it is weak.
The losing years are also what any explanation of the result has to account for. Without them, the 2026 result would have been one good window with nothing to test the lag against.
The second lesson is about how a team checks its own work. Because every piece was passed around, nobody was the last reader of their own section. That mattered most for the model. I owned it, so I was the person least able to see its gaps.
Where Else This Applies
Reporting the periods that went against you, and explaining them. Any performance claim is measured over some window, whether for a strategy, a pricing change, a marketing channel or a process redesign. That window is usually chosen after the result is known. Showing the periods that failed next to the one that worked, with the reason they failed, is the version a reader can trust.
Telling what a company does from what it says. Reading an organization from its own documents is the same problem in sustainability reporting, in risk, and in any due-diligence review of a company being bought. The useful evidence is the specific, numbered claim about its own operations. General statements written for an audience add little.
Finding out what the audience rewards before producing for it. Most organizations keep a record of what their audience has rewarded: prior winning entries, prior approved deals, board memos that got a decision. It is almost never read before the work starts.