In one major survey, 98% of leaders report a positive return on AI. In another, 1% of the S&P 500 can show it in earnings. Both numbers are true, and the distance between them is where most AI business cases quietly fail.
At a glance
We reviewed 179 studies from the world's leading firms and experts, published from 2024 to 2026, and 37 company disclosures.
Ask whether AI is delivering value and the median answer is 87%. Ask whether it has moved the P&L and the median falls to 15%.
No source we scored traces an AI practice to a specific financial line with a worked example.
Four in ten sources count no AI cost at all.
AI value is measured relentlessly and almost never shown. Read together, the leading studies describe how AI feels far more precisely than what it earned, and none gives a sponsor a number finance would sign. This page covers the headline findings; the full report sets out the tests, the scorecard and the approach.
The question sets the number
The closer a study's question gets to the P&L, the less value it finds. Self-reported value and positive ROI cluster between 67% and 98%. Questions about EBIT, EBITDA or both revenue and cost fall to between 6% and 37%.
Source: 25 study findings from 18 publishers, 2024–2026 · Outperforma analysis
The headline moved faster than the results
BCG's widely quoted swing, from "5%" of companies capturing AI value to "almost half," reflects which tier led the press release. Its top tier moved from 4% to 7.5% over two years.
Source: BCG, Where's the Value in AI? (Oct 2024); The Widening AI Value Gap (2025); Applied AI Index (2026)
A stricter question erases most of the headline
Ask the same respondents a harder question and the number shrinks, in most cases by two-thirds or more. FTI's private equity survey moves from 95% "meeting expectations" to 17% "significantly exceeding" them.
Source: Six 2026 studies; each pair comes from a single survey · Outperforma analysis
AI feels faster than it measures
In METR's randomized trial, experienced developers believed AI had made them 20% faster. Measured, they were 19% slower. In the U.S. CFO Survey, firms reported productivity gains three times larger than their own revenue and headcount imply.
Source: METR (July 2025) randomized controlled trial; Federal Reserve Bank of Atlanta / Duke CFO Survey (March 2026)
The advice stops where the money starts
We scored every source on how far its guidance goes, from a slogan to a practice tied to a specific P&L or balance-sheet line with a worked example. More than half stop at slogans or named practices, and none reaches the financial line.
Four in ten sources count no AI cost at all, and only one in eight counts the human review that makes AI output usable. Where cost is counted, the token invoice is the smallest part of the bill.
Falling prices will not close the gap. Token unit costs have dropped by more than 95% while AI bills at many companies have more than tripled.
Source: McKinsey, The AI Operating System (October 13, 2026)
Companies talk about AI on the P&L, then stop short of proof
Of 37 company disclosures, 15 put a dollar figure on AI and one explains how it was measured. None appears in audited financial statements as a benefit attributed to AI.
Source: 37 public-company disclosures, 2024–2026 · Outperforma analysis of earnings calls, filings and investor-day materials
What the full report adds
The full 20-page report goes further. It tests the leading frameworks against nine standards finance already uses and shows which come closest. It maps the six questions leadership teams ask most often against the financial levers they should move. And it sets out Outperforma's AI Economics approach, the RTI tree, with a worked scenario that traces AI cost and value from the token to the P&L. The report is free; download it below.
