Fact Brief: AI Adoption vs. Business Transformation

Why this brief exists

I wanted to say something simple in public: AI adoption is not business transformation. Before I did, I wanted to know whether the evidence backs it up or whether it only sounds right.

It mostly backs it up. Studies by McKinsey, BCG, PwC, MIT and the National Bureau of Economic Research find the same pattern. AI use is widespread, and few firms can show that it has changed their financial results.

That turned out to be the less interesting finding. None of the major studies dispute it, including the ones published by firms that sell AI transformation. The open question is what the gap means.

One reading says it is a lag. General-purpose technologies take years to show up in productivity data, and the payoff comes once firms have rebuilt how they work. The other reading says the gap does not close by waiting, because most firms never do the rebuilding.

The evidence doesn't settle that yet, but it narrows it. It shows where gains are measured and where they are only reported, which firms report value, and what the 2026 data changed.

I am arguing my side of the lag-or-failure question in a separate article. This brief is the evidence, kept apart from the argument, so anyone can check the numbers before agreeing or disagreeing with me.

Adoption is counted in users. Transformation has to show up in the results.

That is the line this brief tests. Every finding below is labeled by the kind of evidence behind it, and every source is listed with how closely it was checked.

Markus Brinsa

Founder & CEO, SEIKOURI Inc.

---

What the evidence supports

This brief was compiled with Claude, an Anthropic model, from published studies, surveys and official statistics. No company data beyond those publications was reviewed, so nothing here describes a specific firm. Figures reflect sources available as of November 16, 2026.

The brief separates four kinds of statements. Measured means an outcome taken from records, official economic statistics or a controlled experiment. Reported means an answer given in a survey, including government business surveys. Estimated means a forecast or a model-based calculation. Inference means the sources don't establish it, so read it as an argument, not a finding.

The distinction matters more than usual here. Most of the evidence on business impact is reported rather than measured, and two independent sources show reported gains running ahead of measured ones. The surveys also define success differently, from any EBIT impact to value at scale, so their figures are not directly comparable.

Several sources sell advisory services related to what they measure, including McKinsey, BCG, PwC and Gartner. Many of the academic sources are working papers that have not been peer-reviewed. The figures also move quickly, so this brief will be updated as new data appears.

What the evidence supports, in short:

• Reported: Most firms using AI report little or no measurable financial effect. Between about 5% and just under half report substantial value, depending on the study and its definition of value.
• Reported: Firms that report value are consistently more likely to report redesigned workflows and enterprise-wide AI programs.
• Measured: Gains at the task level are real but uneven, from a 14% productivity gain in customer support to a 19% slowdown for experienced developers.
• Measured: In Danish administrative records, AI chatbots changed how work is organized without measurably changing earnings or hours.
• Estimated: Effects on national productivity remain small or unconfirmed. Whether 2025 marked the start of an AI-driven payoff is disputed.

---

Key figures at a glance

The studies use different samples and different definitions of value, so their figures are not directly comparable. Newest first.

---

How widespread adoption is

Adoption looks widespread in executive surveys and far lower in government data covering all businesses. Every adoption figure here is reported, including the government's.

• Reported: McKinsey found 88% of organizations using AI in at least one business function in mid-2025, up from 78% a year earlier. The NBER survey found 78% of firms using AI in summer 2026.
• Reported: The U.S. Census Bureau's Business Trends and Outlook Survey puts AI use at 17% to 20% of all U.S. businesses between December 2025 and May 2026. Weighted by employment, a Census supplement covering November 2025 to January 2026 put use at 32%.
• Reported: Much of the gap tracks firm size. In Census data, fewer than 20% of firms with four or fewer employees use AI, against 37% of firms with 250 or more.
• Reported: Where AI is used, use is often shallow. Executives in the NBER survey averaged 2.1 hours of AI use a week, and 57% of Census adopters use AI in three or fewer business functions.

---

What firms report about business impact

Most firms report little or no measurable financial effect, and a minority reports substantial value. The answer shifts with the question asked.

• Reported: In McKinsey's 2025 survey, 39% of respondents attributed any EBIT impact to AI, and most of them put it below 5% of EBIT. About 6% met McKinsey's high-performer bar of 5% or more of EBIT plus significant value.
• Reported: BCG's 2025 study of 1,250 firms found 5% achieving AI value at scale and 60% reporting minimal revenue and cost gains despite substantial investment.
• Reported: PwC's survey of 4,454 CEOs, fielded from September to November 2025, found 56% had seen neither higher revenue nor lower costs from AI. Only 12% had seen both.
• Reported: A BCG survey of 152 CEOs at companies with at least $500 million in revenue, released in July 2026, found nearly nine in ten seeing cost or revenue benefits in targeted areas. In the same survey, 26% had embedded AI in a broader business transformation, and 14% had defined the P&L impact of all their AI initiatives.
• Reported: MIT's Project NANDA found in July 2025 that 95% of organizations were getting no measurable P&L return from generative AI. The report drew on 153 leader surveys, 52 interviews and more than 300 public implementations, and its methods are contested.
• Reported: In the NBER executive survey, executives said AI raised annual labor productivity growth by 0.2% over the past three years, with essentially no net effect on employment. In its earlier round of about 6,000 executives, fielded from November 2025 to January 2026, 89% reported no productivity impact at all.

---

What separates the firms that report value

Firms that report value tend to redesign work rather than add tools to existing processes. That pattern appears in every major survey, but only as a correlation.

• Reported: McKinsey's high performers were nearly three times as likely as others to have fundamentally redesigned individual workflows. In BCG's July 2026 CEO survey, high performers were about seven times more likely to redesign workflows and reshape the business end to end.
• Reported: BCG's 2026 leaders run AI as an enterprise-wide program, invest in governance and workforce alongside technology, and measure impact directly in the P&L. PwC found that CEOs reporting both cost and revenue gains were two to three times more likely to have embedded AI extensively across products, demand generation and strategic decisions.
• Caveat: Well-run firms may both redesign more and report more value, and surveys of successful firms are open to survivorship bias. None of the studies shows that redesign causes the results.

---

Evidence from field studies and administrative data

Controlled studies and administrative records find real effects on tasks and on how work is organized, but not yet on pay or hours.

• Measured: A study of 5,179 customer-support agents, published in the Quarterly Journal of Economics in 2025, found a 14% gain in issues resolved per hour. Novice and low-skilled agents gained 34%, while experienced agents gained little.
• Measured: In a randomized trial by the research nonprofit METR, 16 experienced open-source developers completed 246 real tasks using early-2025 tools. When AI was allowed, tasks took 19% longer, yet the developers afterward estimated that AI had made them about 20% faster.
• Measured: Economists Anders Humlum and Emilie Vestergaard linked surveys of about 25,000 Danish workers in 11 AI-exposed occupations to official labor-market records. They rule out effects of AI chatbots on earnings or hours larger than 2%.
• Reported: In the same study, employers reorganized tasks, and about 17% to 18% of users at actively adopting workplaces reported entirely new tasks. The authors conclude that technology reshapes work well before it shows up in earnings or hours.
• Reported: In a September 2025 survey of 1,150 U.S. desk workers by the Stanford Social Media Lab and BetterUp Labs, 40% had received low-quality AI-generated work, which the researchers call "workslop," in the past month. Each incident took nearly two hours to handle, an estimated $9 million a year for a 10,000-person company.
• Measured: Using ADP payroll data, Stanford economists Erik Brynjolfsson, Bharat Chandar and Ruyu Chen found a 13% relative employment decline since late 2022 for 22- to 25-year-olds in the most AI-exposed occupations. The Danish study also sees early-career declines in exposed occupations, but finds no link to firms' own chatbot adoption.

---

The macroeconomic view

The pattern has precedents. In 1987, economist Robert Solow observed that computers were visible everywhere except in productivity statistics, and Apollo's chief economist Torsten Slok applied the same line to AI in February 2026. Before AI, BCG found in 2020 that 70% of digital transformations fell short of their objectives.

The productivity J-curve, formalized by Erik Brynjolfsson, Daniel Rock and Chad Syverson, explains the delay. General-purpose technologies require large investments in processes, skills and organization that depress measured productivity before gains appear.

• Measured: Census Bureau data on industrial AI in U.S. manufacturing shows that sequence: productivity losses first, medium-term gains later.
• Estimated: MIT economist Daron Acemoglu estimates that AI will add no more than 0.66% to total factor productivity over ten years.
• Estimated: In February 2026, Brynjolfsson argued that U.S. productivity grew about 2.7% in 2025, nearly double the prior decade's average, and that the payoff phase had begun. Slok and others said the macroeconomic data did not yet show an AI effect.
• Reported: A survey of about 750 U.S. financial executives by researchers behind The CFO Survey found AI-implied labor productivity gains of 0.6% in 2025, expected to reach 1.8% in 2026. The same executives reported gains of 1.8% and 3.0%, a difference the authors attribute to delayed revenue and quality gains that revenue does not yet capture.

---

Is the gap closing?

The 2026 data moved, but how much depends on the comparison.

• Reported: BCG's September 2026 update, based on 1,330 executives, found nearly half of companies generating value from AI, against 5% achieving value at scale in 2025. On the same categories, future-built companies rose from 5% to 7.5% and scaling companies from 35% to 41%.
• Reported: BCG puts AI spending at 3.3% of revenue, up from about 1.7% in late 2025. The NBER survey's 0.2% annual productivity contribution remains modest, and executives expect 0.6% a year over the next three years.

Two readings of the same data compete. In the first, the gap is a lag: measured gains follow once complementary investments mature, as the J-curve predicts. In the second, value concentrates in firms that redesign how they operate, and BCG describes leaders pulling further ahead. The data available as of this brief does not settle which reading is right.

---

Governance as adoption moves to agents

The next phase moves from tools people use to agents that act, and controls trail the plans.

• Reported: In mid-2025, McKinsey found 23% of organizations scaling an agentic AI system somewhere, but no more than 10% in any single business function. In BCG's 2026 survey, 42% of companies expect to give AI agents genuine autonomy by 2030.
• Reported: Only 5% of companies in BCG's 2026 survey have the full set of controls BCG considers necessary, including oversight, rollback gates, security, audit and cost guardrails. In BCG's July 2026 CEO survey, 82% include technology in AI governance and 30% include HR.
• Estimated: Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. It estimated that only about 130 of the thousands of vendors marketing agentic AI offer genuine agentic capabilities.

---

What the evidence doesn't support

---

Inferences, not findings:

• The measurement gap: Two independent sources, the CFO survey and the METR trial, show perceived gains exceeding measured ones. If that pattern holds more widely, survey-based shares of firms getting value from AI overstate measured value.
• The redesign link: Workflow redesign correlates with reported value in McKinsey, BCG and PwC data. Redesign may drive the value, or well-run firms may simply do both.
• Work before results: The Danish data shows work reorganizing while pay and hours stay unchanged, which fits the early phase of the J-curve. It does not show that the reorganization will produce results later.
• The governance gap: In BCG's data, plans for agent autonomy (42% by 2030) run far ahead of controls (5% today). If transformation increasingly runs through agents that act on their own, control capacity may become a limit on it.

None of these is established by the sources. Each is a reasonable reading that further data could confirm or refute.

---

Conclusion

The evidence supports the claim that AI adoption is not business transformation. Use is broad, measured effects on firms' results are small, and the minority reporting substantial value is the minority that reports changing how it works.

The evidence does not yet show which way the gap goes. The lag reading and the persistent-gap reading both fit the data through 2026. The largest movement in 2026 came in survey-reported value, while measured effects on firm results remain small.

For leaders, the practical consequence is the same under both readings. Treat adoption figures as inputs, not results, and look for the change in the P&L, measured rather than reported.

---

Source register