The five lines

1. AI investment. Annual AI spend. Disclosed when the company states it, derived when built from figures it states, modeled from indirect evidence (AI headcount, share of R&D or IT, capex) with a wide range. Acquisitions are amortized over five years.
2. Claimed value. Management's quantified AI value claims, in annual dollars, that point at a reported line (a cost line or a revenue line). Claims the company does not attribute to AI are excluded.
3. Verified value. How far the line beat its expected path, capped at the claim and never below zero.
4. Net AI return. Verified value minus AI investment for the same year. Negative when spend outweighs what the statements show.
5. Impact. Net return in basis points of revenue (margin impact), as a share of operating income, and verified revenue uplift as a share of revenue growth.

Four baselines, one ensemble

Each claim is tested against four expected paths. The published verified value is the average of the two central ones; the range is the lowest to highest of all four, combined with the low-to-high investment estimate.

The peer baseline asks whether the company improved more than similar companies over the same years. If a whole sector's cost ratio fell, the company does not get credit for the sector's tailwind. Peer groups: Financials, Health care, Industrials, Technology & services, Consumer & retail, and Telecom. When the two central methods disagree by more than 2×, or one finds value and the other none, the company loses a confidence point.

AI attribution

Each claim gets a probability that management attributes it to AI rather than to a broad program (restructuring, "efficiency", automation in general). At 0.70 or above it counts in full; 0.40 to 0.70 counts and is flagged for analyst review; below 0.40 it is excluded. An attribution-weighted verified value is published alongside the base figure. Since NAIR v0.5 the probability comes from Jev, an independent decision model (TypeSafe, version pinned), asked whether AI is the main stated driver of the result: revenue from a product that is itself AI, or a result credited mainly to the company's own use of AI. A transparent keyword rule is used only if Jev is unavailable, and every API response names which backend produced each probability.

Productivity signal

Revenue per employee, change over the covered years, compared with the peer median. It is context for the verdict (the Jevons question: does AI make each employee more productive, or does it just add spend?). It is never added to verified value. Gaps over 40% usually reflect spin-offs, divestitures or pandemic swings and are flagged as not comparable.

Verdict and confidence are separate

• Verdict says what the statements show: Paying off, Partly visible, Not visible, or Not testable.
• Confidence says how sure we are of that verdict, scored 0 to 5: disclosed investment (2 points, 1 if derived), a testable claim (1), methods agree (1), clear AI attribution (1). 4 or more is High, 2 to 3 Medium, 0 to 1 Low.

Disclosure grade (since NAIR v0.6)

Most large companies describe AI value in operational percentages (contacts handled, cycle time, share of code written by AI) with no line in the financial statements behind them, so most of the S&P 500 is Not testable under this model. The disclosure grade says how clearly each company reports AI value, so every company on the leaderboard carries a reading even when it cannot be tested. It is scored out of 100 and shown as a letter: A from 80, B from 65, C from 50, D from 35, otherwise F.

• Specificity (30). Share of logged claims that are quantified (15), at least one dollar-denominated value claim such as revenue from AI or costs saved (10), and AI spend disclosed or derivable from disclosed figures (5).
• Testability (25). At least one claim tested against a statement line (15), and the share of claims the Jev decision model rates as dollar-testable (10).
• Sourcing (20). Share of claims whose source is the company's own: SEC filing, press release, investor materials or a company-hosted transcript. A third-party transcript of a company call counts half; news and analyst commentary count nothing. For companies researched before September 2026 the source type is read from the source URL.
• Consistency (15). Starts at 15; minus 5 for every claim later contradicted by the company's own disclosures, minus 3 for every claim management stopped reporting.
• AI attribution (10). Mean probability, across the company's claims, that AI is the stated driver rather than one lever of a broader program (Jev, the same model used for verification).

The grade is about disclosure, not results: a company can report AI value with great clarity and still show no verified return, and a Paying off company can grade poorly if its claims are vague or sourced only through the press. It never enters verified value, net return, the verdict or confidence. Companies with fewer than three logged claims carry a dagger (†) because the grade rests on little; companies with no logged claims are not graded.

Built on open methods

The approach follows published, open-source causal-inference practice so any customer can reproduce it:

• diff-diff and CausalPy: difference-in-differences and synthetic-control estimators, the basis of the peer baseline.
• CausalImpact (TensorFlow Probability): Bayesian structural time-series counterfactuals, next once quarterly lines are loaded.
• MAPIE: conformal prediction intervals, to replace the min-max range with calibrated 80% and 95% intervals.
• Jev (TypeSafe System One): typed, probability-scored decisions for claim attribution, category and testability; Open-Jev is an open implementation of the same idea.

What this does not claim

Financial statements cannot prove causation. Verified value means the numbers are consistent with the claim beyond trend and beyond peers. Zero means the claim is not visible in reported results yet, which can be true of real but small or early benefits. This site publishes data and method only. It is not investment advice.