The AI Study That Graded Its Own Customers

Metadata

• Topic: PwC 2026 AI Performance Study — methodology disclosure and conflict of interest
• Content Pillar: The Fine Print
• Status: revised
• Subject Line: The AI Study That Graded Its Own Customers
• Preview Text: Nearly 90% of firms say AI hasn't moved their productivity. The consulting industry billed $11B anyway.

Introduction

Nearly 90% of roughly 6,000 executives across the US, UK, Germany, and Australia told researchers last year that AI has not moved the productivity needle at their companies in three years. The AI consulting industry billed $11.07 billion in 2025 and is projected to reach $90.99 billion by 2035.

What PwC Published, and What Page 28 Admits

Last week, PwC released its 2026 AI Performance Study with a headline number built to travel: 74% of AI-driven economic gains go to just 20% of companies. The winners post an AI-driven performance score of 15.7 against a bottom four-fifths average of 2.2. The absolute bottom quintile sits at a negative score of -2.6, meaning some companies are losing ground from AI rather than gaining any.

The press release leads with that 74/20 split, and so does every downstream write-up. What nobody quotes is the methodology note on page 28 of the actual PDF: the sample is 91% publicly listed companies and 76% companies with US$1 billion or more in revenue. Fieldwork ran October and November 2025. "AI-driven performance" is self-reported by executives, not audited. Scores are measured relative to sector median, not in absolute dollars.

The study measured how already-large, publicly listed companies split AI gains among themselves, with 76% of the sample clearing $1 billion in revenue before the AI wave began, then packaged that finding as a universal diagnostic for any business.

The Firm That Wrote the Study Sells the Fix

The study is bylined by three PwC executives: Joe Atkinson, Global Chief AI Officer; Agnes Koops, Global Vice Chair and Global Chief Commercial Officer; and Matt Wood, Global CTIO. These are the business-unit heads whose sales pipelines get fed every time an enterprise reads the recommendation and calls a consultant.

PwC is not a neutral observer of the AI market. In May 2024, the firm became OpenAI's first reseller and largest enterprise customer of ChatGPT, rolling the product out to 100,000 employees. PwC has invested $1 billion in generative AI over three years and is engaged on AI with 950 of its top 1,000 U.S. client accounts. The case studies featured inside the very same PDF are PwC's own consulting engagements: John Deere's See & Spray system, Wyndham Hotels' 94% reduction in brand standards review time, Southwest Airlines' 50% backlog reduction, and two unnamed clients in retail and healthcare.

PwC authored and graded a test whose featured winners are PwC's own consulting clients, and the top-line recommendation that laggards "build AI fitness" with partners is the same line already on its sales deck.

The 6,000-Executive Survey PwC's Number Can't Survive

Look at the researchers who aren't selling a service, and the picture looks different. The National Bureau of Economic Research Working Paper 34836, co-authored with the Federal Reserve Bank of Atlanta, Bank of England, Deutsche Bundesbank, Macquarie, and Stanford, surveyed roughly 6,000 executives across the US, UK, Germany, and Australia. 89% report no impact of AI on labor productivity over the past three years, and 90% say the same about employment. Executive AI use averages 1.5 hours per week, and only 7% of those executives report more than five.

PwC's own 29th Global CEO Survey from January 2026 lands in roughly the same place if you read past the cover. Of the 4,454 CEOs sampled, 56% report getting nothing measurable from their AI investments, 22% say their costs actually went up, and only 12% report both revenue growth and cost reduction.

Line those two findings up against PwC's April 2026 AI Performance Study and the arithmetic stops making sense as growth. If nearly 90% of firms see no productivity lift while a specific 20% pull in three-quarters of the reported gains, the top quintile isn't producing new aggregate value; it's capturing share from everyone else inside its sector. The Chicago Fed's Letter 518 gets there from a different angle: AI's firm-level effects are real and big, but "not yet visible at the aggregate level" because individual wins don't roll up to sector growth.

Who Actually Gets Paid

The "20% of companies" framing hides a tier of beneficiaries the study never counts: the firms selling AI into every quintile at once. PwC is one. Accenture booked roughly $3.6 billion in generative AI contracts in fiscal 2025. BCG pulls roughly $2.7 billion a year from AI services. IBM's watsonx book of business has reportedly crossed $6 billion since 2023. Microsoft Azure blew past $75 billion in FY2025 revenue, with AI services driving 34% year-over-year growth and roughly $13 billion in annualized AI revenue as of early 2025. None of those revenue lines require the buyer to see a measurable return.

Per-seat pricing makes the math obvious. Microsoft Copilot runs $30 per user per month on top of an existing Microsoft 365 license. Salesforce Agentforce 1 Edition starts at $550 per user per month. Workday's own January 2026 research with Hanover found that 37 to 40% of the time AI supposedly saves gets spent reviewing, correcting, and verifying the AI's output. The subscription bills regardless of whether the verification overhead eats the savings.

The companies in PwC's top 20%, capital-rich and data-mature by definition, write big checks to Accenture and Microsoft to build custom pipelines. Everyone else writes smaller but steadier checks for per-seat licenses, trying to keep up. PwC, Microsoft, Accenture, IBM, and the cloud providers cash both sets of checks.

The Part That Doesn't Fit the Usual "AI Diffusion" Story

Economists who study general purpose technologies tend to point to the Solow Paradox as reassurance. Robert Solow noticed in 1987 that "you can see the computer age everywhere but in the productivity statistics." The gains eventually showed up, but not for roughly 10 to 15 years after adoption began. Brynjolfsson and collaborators formalized this as the "productivity J-curve." The San Francisco Fed's February 2026 letter leans on that history to argue the current gap between AI adoption and measurable gains is normal early-stage behavior.

The analogy breaks in one specific place. Electrification and the internet rollout sold the infrastructure once — a factory bought a Westinghouse motor; a company bought bandwidth. AI monetizes the full diffusion period at a per-seat monthly rate, charged by the same firms positioned to capture most of the gains when the curve eventually resolves.

Acemoglu's 2025 paper in Economic Policy pegs AI's ten-year boost to US total factor productivity at "at most 0.66%," a fraction of the 1.5 percentage points per year Goldman Sachs projected in 2023. Even on the optimistic end of Acemoglu's range, the cost of waiting a decade for gains to diffuse lands as quarterly enterprise invoices on companies whose bottom-quintile AI score is already running negative.

The Bottom Line

A small group of already-large firms is capturing most of AI's measurable gains. Everyone else is spending real money on subscriptions, retraining, and consulting engagements while waiting for gains that may or may not arrive. The firms running the picks-and-shovels operation (consultants, cloud vendors, platform owners, model developers) get paid either way.

The Stanford 2026 AI Index, summarized by MIT Technology Review, does document real productivity gains of 14% in customer service and 26% in software development inside specific tasks, so AI isn't useless. The question the PwC framing ducks is who captures those gains at the firm level and how much the middlemen siphon off on the way. When your board asks why the Copilot pilot hasn't paid off yet and whether you should hire a consultant to audit your "AI fitness," notice whose spreadsheet the question is really serving.

The sample methodology is on page 28 of the PDF itself, if you want to check.