A July 2025 report from MIT’s Project NANDA stated that 95 per cent of organisations were getting zero return from generative AI. The figure rests on 52 interviews and a six-month window, while the larger surveys of executives published since then find small measured gains alongside large expectations. The evidence places the binding constraint in the organisation where models already handle the task, and in the models themselves where agents must carry long work alone.
In July 2025 four researchers at MIT’s Project NANDA published “The GenAI Divide: State of AI in Business 2025”. Its executive summary states that “95% of organizations are getting zero return” on generative AI, and its authors locate the cause in how tools are adopted: “The core barrier to scaling is not infrastructure, regulation, or talent. It is learning.”1 Fifteen months later, a survey of nearly 9,000 senior executives in seven countries, revised in October 2026, reports 78 per cent of their firms using AI and that executives credit it with adding about 0.2 per cent a year to labour productivity growth over the past three years.2
Measured returns on generative AI in companies have so far been small, and the 95 per cent figure measures something narrower than an organisational failure: the share of custom tools that reached production with measurable effects within six months, drawn chiefly from interviews with 52 organisations. Where a model can already do the task, how the company deploys it governs the return, though at company level the evidence for this is largely correlational. Where agents must carry long, multi-step work on their own, the limits of the models still bind first.
Three tests can falsify this reading.
- Size. In large surveys of firms, most firms using AI must report no measurable productivity effect so far, with an average effect below one per cent. If such surveys show large measured gains, the premise fails.
- Mechanism. Where the tool is held constant, outcomes must vary with what employers do. If the same tools fail regardless of deployment, model capability is the better explanation.
- Scope. Measured gains must appear on bounded, high-volume tasks and weaken on long, multi-step work done by agents. If agents complete most realistic multi-step tasks, the limit fails.
The first test passes. The second passes within firms and remains correlational across them. The third passes on the evidence available in November 2026.
What the 95 per cent measures
The NANDA report draws on three bodies of evidence, which its methodology section lists: a systematic review of over 300 publicly disclosed AI initiatives, structured interviews with representatives of 52 organisations, and survey responses from 153 senior leaders collected at four industry conferences. The research period ran from January to June 2025. The report defines success as “deployment beyond pilot phase with measurable KPIs”, with return on investment “measured 6 months post-pilot, adjusted for department size.”1
The figure therefore describes a specific population and a specific window. The organisations are those that disclosed initiatives publicly or agreed to be interviewed, and the leaders are those who attended the conferences. The outcome is a pilot that reaches production and shows measurable effects within half a year. The report states its own limits: its figures are “directionally accurate based on individual interviews rather than official company reporting,” its percentages “reflect our interview sample of 52 organizations and may not represent broader market patterns,” and its “six-month observation period may be insufficient” to assess outcomes fully.1
The report also separates two kinds of tool. For general-purpose assistants such as ChatGPT and Copilot, over 80 per cent of organisations had explored or piloted them and nearly 40 per cent reported deployment. For custom or enterprise-grade systems, 60 per cent had evaluated tools, 20 per cent had reached the pilot stage and 5 per cent had reached production.1 The 95 per cent figure belongs to the second funnel. Read as a sequence, the same numbers say that one organisation in four that ran a pilot of a custom tool reached production in the period studied. Elsewhere the report describes its 5 per cent as a share of “integrated AI pilots”, a different base, and it does not reconcile the two uses.1
| Element | What the report states |
|---|---|
| Evidence base | Systematic review of over 300 publicly disclosed AI initiatives; structured interviews with representatives of 52 organisations; survey responses from 153 senior leaders at four industry conferences |
| Definition of success | Deployment beyond pilot phase with measurable KPIs |
| Timing of return | Measured 6 months post-pilot, adjusted for department size |
| Stated limits | Figures “directionally accurate based on individual interviews rather than official company reporting”; percentages reflect the 52-organisation interview sample |
| General-purpose tools (ChatGPT, Copilot) | Over 80% explored or piloted; nearly 40% report deployment |
| Custom or enterprise-grade tools | 60% evaluated; 20% reached pilot; 5% reached production |
| Pilots of custom tools that reached production | 25% (5 of 20), computed by Titanica from the funnel above, which the report presents on a common base of organisations |
| Evaluations of custom tools that reached production | 8% (5 of 60), computed by Titanica from the funnel above |
The report’s diagnosis also assigns part of the problem to the tools. Its authors write that “most GenAI systems do not retain feedback, adapt to context, or improve over time,” and they describe the failing tools as ones “that don’t learn, integrate poorly, or match workflows.”1 The first of those properties belongs to the software, the third to the fit between software and work, and integration to both. The report thus supports an organisational reading only in part, and its own explanation is mixed.
Two further findings in the report bear on the organisation directly. Only 40 per cent of companies said they had purchased an official subscription to a large language model, while workers from over 90 per cent of the companies surveyed reported regular use of personal AI tools for work.1 And tools built with external partners reached deployment about 67 per cent of the time, against about 33 per cent for tools built internally, a correlation that the authors note “does not necessarily prove causation.”1 Both findings describe employees and firms using the same class of technology with different results, which is the pattern an organisational explanation predicts. Neither isolates it.
Six surveys, six denominators
The share of firms said to be using AI depends on who is asked and how answers are weighted. In May 2026 the US Census Bureau reported that 19.8 per cent of American businesses used AI in the period ending 3 May, in its Business Trends and Outlook Survey. Among firms with at least 250 employees the share was 37 per cent.3 The Bureau changed its core question in November 2025 from use “in producing goods or services” to use “in any business function”, so later figures are not directly comparable with the earlier series.3 Expectations differ by sector as much as by size. About 42 per cent of firms in the information sector and roughly 39 per cent in finance and insurance expected to use AI within six months, against about 17 per cent in retail trade, and none of these rates had changed significantly from six months earlier.3
A note by the Federal Reserve Board in April 2026 set the Census survey, which weights each firm equally and so reflects the many small firms, beside the Atlanta Fed’s Survey of Business Uncertainty, which asks executives and weights responses by employment. That survey estimated that 78 per cent of the US labour force worked at firms that had adopted AI in November 2025, and 69.4 per cent of firms on an equal-weighted basis.4 The note attributes the gap to weighting, sampling, question wording and the type of respondent, and observes that surveys aimed at senior leaders may be pushed upward because leaders face pressure to report AI use as an efficiency initiative.4
Surveys run by consultancies and business schools report higher adoption and more varied returns. McKinsey’s survey of 1,993 participants in 105 countries, fielded from 25 June to 29 July 2025, found that 88 per cent of organisations used AI regularly in at least one business function, that 39 per cent reported any effect on earnings before interest and taxes at enterprise level, and that about 6 per cent qualified as “AI high performers”, attributing 5 per cent or more of EBIT to AI.5 The Wharton School’s report with GBK Collective, published in October 2025, found that 82 per cent of leaders used generative AI at least weekly and that “three out of four leaders see positive returns” on their investments, with 72 per cent formally measuring return.6
| Survey | Respondents and period | Measure | Result |
|---|---|---|---|
| US Census Bureau, Business Trends and Outlook Survey | US firms, equal-weighted; period ending 3 May 2026 | Firms using AI in any business function | 19.8% of all firms; 37% of firms with 250+ employees |
| Atlanta Fed, Survey of Business Uncertainty (via Federal Reserve Board note) | US executives; November 2025 | Adoption of AI | 78% of the labour force at adopting firms (employment-weighted); 69.4% of firms (equal-weighted) |
| Yotzov et al., “Firm Data on AI” (March 2026 version) | Central-bank and university business panels, US, UK, Germany, Australia; November 2025 to January 2026 | Firms using AI; no productivity effect over past three years | 69% use AI; 89% report no effect on labour productivity |
| McKinsey, “The state of AI in 2025” | 1,993 participants, 105 countries; 25 June–29 July 2025 | Regular use; enterprise EBIT effect | 88% use in at least one function; 39% report any EBIT effect; about 6% attribute 5%+ of EBIT to AI |
| Wharton and GBK Collective, “Accountable Acceleration” | Enterprise leaders; published October 2025 | Weekly use; positive return | 82% use weekly; three in four see positive returns; 72% formally measure return |
| MIT Project NANDA, “The GenAI Divide” | 52 interviews, 153 conference respondents, 300+ public initiatives; January–June 2025 | Custom tools in production with measurable KPIs | 5% of organisations; “95% of organizations are getting zero return” |
The surveys measure different things. The Census counts firms, most of them small; the Atlanta Fed counts workers inside adopting firms; McKinsey and Wharton count respondents who chose to answer surveys about AI; NANDA counts custom projects that crossed a defined line within six months. A return of “zero” in one and “positive” in another can describe the same economy. For the first test, the most useful of them is the one built on established business panels run by central banks and universities, which asks a fixed question about productivity.
Small returns, large expectations
The paper “Firm Data on AI”, by Ivan Yotzov, Nicholas Bloom, Steven Davis and fourteen co-authors, combines AI questions placed in four business panels: the Atlanta Fed’s Survey of Business Uncertainty in the United States, the Bank of England’s Decision Maker Panel, the Bundesbank’s online panel of firms, and Macquarie University’s Business Outlook Scenarios Survey in Australia. The questions were fielded between November 2025 and January 2026.7 In the March 2026 version, 69 per cent of firms on average used AI, and 89 per cent reported no impact on labour productivity, measured as sales per employee, over the past three years.7
The average reported effect on productivity over the past three years was 0.29 per cent across the four countries. The same executives expected an average effect of 1.44 per cent over the next three years, with the largest expected gains in the United States (2.25 per cent) and the United Kingdom (1.86 per cent).7 The October 2026 revision, which adds Canada, Mexico and South Africa and brings the sample to nearly 9,000 executives, states the results as an effect of 0.2 per cent a year on productivity growth so far and an expected 0.6 per cent a year over the next three years.2
| Country | Firms using AI | No productivity effect, past 3 years | Average effect, past 3 years (%) | Expected effect, next 3 years (%) | Expected ÷ past |
|---|---|---|---|---|---|
| United States | 78% | 91% | 0.24 | 2.25 | 9.4 |
| United Kingdom | 71% | 89% | 0.29 | 1.86 | 6.4 |
| Germany | 65% | 91% | 0.24 | 0.87 | 3.6 |
| Australia | 59% | 79% | 0.49 | 0.92 | 1.9 |
| All four | 69% | 89% | 0.29 | 1.44 | 5.0 |
The first test passes on these data. Most firms that use AI report no measured productivity effect, and the average is a fraction of one per cent. The result does not depend on the NANDA sample. It comes from established business panels run by central banks and a university, though the respondents are executives and their answers are estimates rather than accounts. The March version finds that executives at larger and higher-paying firms expect greater gains,7 and the October revision reports that about 2 per cent of firms account for more than half of total AI spending.2 Measured returns are small on average, and the spending that might change them is concentrated in few firms.
The employment results follow the same pattern. More than 90 per cent of firms in the March version reported no effect of AI on employment over the past three years, and the average effect was zero; the same executives expected a fall of 0.68 per cent over the next three years, with the largest expected reductions in the United Kingdom and the United States.7 The October revision reports that more than 40 per cent of firms anticipate positive or negative employment effects that largely offset each other in aggregate.2 What executives describe is a technology whose effects they expect, in the main, to arrive later.
Workers’ own data point the same way. Anders Humlum and Emilie Vestergaard linked two surveys of about 25,000 Danish workers in 7,000 workplaces and 11 occupations exposed to AI, conducted in late 2023 and 2024, to administrative records of earnings and hours. Users reported average time savings of 2.8 per cent of work hours, and the authors found “no significant impact on earnings or recorded hours”, with confidence intervals that rule out effects larger than 1 per cent.8 They estimate that only 3 to 7 per cent of the productivity gains reported by workers passed through to higher earnings, and that chatbots had created new job tasks for 8.4 per cent of workers.8
Where the organisation binds
The second test asks whether outcomes vary with what employers do when the technology is the same. The strongest evidence that they do comes from studies within firms. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer-support agents at one company as access to a generative AI assistant was introduced. Access raised the number of issues resolved per hour by 14 per cent on average and by 34 per cent for novice and low-skilled workers, with minimal effect on the most experienced.9 The study, published in the Quarterly Journal of Economics in 2025 and based on the staggered introduction of the assistant across agents, shows a tool that works at the task when an employer has built it into the job, gives it to agents and measures the result.
The Danish data isolate the employer’s role more directly. Where employers encouraged the use of chatbots, take-up for the typical worker rose from 47 to 83 per cent, the gap between men and women in take-up narrowed from 11.9 to 5 percentage points, and the reported benefits were 10 to 40 per cent greater.8 The authors conclude that “firm-driven investments and workplace reorganizations are critical” to the technology’s potential, and they report that 30 per cent of employees had received training.8 Employers chose whether to encourage use, so the comparison can reflect other differences between the firms that encourage it and those that do not.
Across firms, the evidence is correlational. McKinsey’s high performers were nearly three times as likely as other respondents to say their organisations had “fundamentally redesigned individual workflows”, and the firm reports that redesign “has one of the strongest contributions to achieving meaningful business impact of all the factors tested.”5 The NANDA report’s partnership finding and its observation that workers use personal tools where companies have bought none belong to the same class of evidence.1
| Study | Setting | Design | Finding |
|---|---|---|---|
| Brynjolfsson, Li & Raymond, Quarterly Journal of Economics (2025) | 5,179 customer-support agents at one firm | Within-firm comparison as access to an AI assistant was introduced | Issues resolved per hour +14% on average; +34% for novice and low-skilled agents; minimal for the most experienced |
| Humlum & Vestergaard, BFI Working Paper 2025-56 | About 25,000 workers, 7,000 workplaces, 11 occupations, Denmark, 2023–2024 | Surveys linked to administrative data; employer encouragement observed | With encouragement: take-up 47% to 83%; benefits 10–40% greater. Overall: time saved 2.8% of hours; no significant effect on earnings or hours |
| McKinsey, “The state of AI in 2025” | 1,993 participants, 105 countries, 2025 | Cross-sectional survey; correlational | High performers (about 6%) nearly three times as likely to have fundamentally redesigned workflows |
| MIT Project NANDA (2025) | 52 organisations interviewed, 153 leaders surveyed | Interviews and survey; correlational | External partnerships reached deployment about 67% of the time, internal builds about 33%; personal AI use reported at over 90% of companies, official subscriptions at 40% |
The second test therefore passes in its within-firm form. The same customer-support tool raised output when an employer deployed it, and the same chatbots delivered more to workers whose employers encouraged them. The company-level claim, that firms which redesign work earn more from AI because they redesign it, rests on surveys in which redesign and success are reported together. Those surveys are consistent with the claim and do not establish its direction.
A longer literature describes why the organisational step takes time. Brynjolfsson, Daniel Rock and Chad Syverson argued in 2021 that general-purpose technologies require large complementary investments, mostly intangible, which national accounts capture poorly, so that measured productivity growth is understated early in a technology’s life and overstated later. Adjusting for intangibles related to computer hardware and software, they estimated total factor productivity in the United States at the end of 2017 to be 15.9 per cent higher than official measures.10 On that account, reorganisation is itself an investment whose costs appear before its returns, and a six-month window would capture mainly the costs.
Where the model binds
The alternative explanation holds that the models fail at the work companies ask of them. On long, multi-step work the evidence for it is direct. In a randomised trial published by METR in July 2025, 16 experienced open-source developers completed 246 real issues in repositories they had worked on for years, using mainly Cursor Pro with Claude 3.5 and 3.7 Sonnet. They took 19 per cent longer when AI tools were allowed. Before the study they had expected a speed-up of 24 per cent, and afterwards they still believed AI had made them 20 per cent faster.11 METR states that the result is a snapshot of early-2025 tools in one setting, that it does not show AI failing to speed up most developers, and that it has since published results current to early 2026 in a continuation of the study.11
Benchmarks of autonomous agents point in the same direction. TheAgentCompany, a benchmark built by researchers led by Frank Xu and Graham Neubig that simulates a small software company, reported in May 2025 that the most competitive agent completed 30 per cent of its tasks autonomously, and that “more difficult long-horizon tasks are still beyond the reach of current systems.”12 In June 2025 the research firm Gartner predicted that over 40 per cent of agentic AI projects would be cancelled by the end of 2027 because of rising costs and unclear business value, and estimated that only about 130 of the thousands of vendors offering agentic AI were genuine. Its analyst Anushree Verma said that “current models do not have the maturity and agency to autonomously achieve complex business goals.”13
| Type of task | Evidence | Nature | Result |
|---|---|---|---|
| Bounded, high-volume conversations with a human in the loop | Customer support, one firm (Brynjolfsson, Li & Raymond) | Peer-reviewed, measured | Issues resolved per hour +14% |
| Assistance across tasks in AI-exposed occupations | Danish workers, 2023–2024 (Humlum & Vestergaard) | Working paper, self-reported time and administrative earnings | Time saved 2.8% of hours; earnings and hours unchanged |
| Expert maintenance of mature software | 16 developers, 246 issues, early 2025 (METR) | Randomised trial | Tasks took 19% longer with AI allowed |
| Multi-step office and software tasks done by an agent alone | TheAgentCompany benchmark, May 2025 version | Benchmark | Best agent completed 30% of tasks autonomously |
| Enterprise agentic AI projects | Gartner, June 2025 (reported by Reuters) | Analyst forecast | Over 40% expected to be cancelled by end-2027 |
The third test passes on these data. The largest measured gain comes from the most bounded task, a support conversation with a human answering, and the weakest results come from long work in which the system must act on its own or assist experts in complex code. The NANDA report’s emphasis on tools that “don’t learn” points to the same boundary, since its authors describe systems that do not retain feedback or adapt to context.1 Where the task falls within what the model can do, the evidence shows deployment shaping the return. Where it falls outside, the evidence reviewed here contains no case in which reorganisation closed the gap.
The test has an expiry date. The METR result describes tools from early 2025, and its authors have flagged it as out of date.11 The agent benchmark is from May 2025. Model capabilities move faster than company structures, so the boundary between the two regimes moves too, and a claim about where it lies in November 2026 needs fresh measurement to hold in 2027.
The case that the models are the bottleneck
The strongest objection a specialist would raise runs as follows. The firms that reported no return in 2025 and 2026 bought tools that, by benchmarks and controlled trials of the period, completed under a third of realistic multi-step tasks and slowed expert developers, so the simplest explanation of weak returns is weak tools. The correlation between workflow redesign and profit in consultancy surveys may run backwards, because firms redesign work around a tool after it has proved itself. An organisational explanation that attributes every failure to insufficient redesign cannot be falsified in practice.
Much of this holds. The company-level evidence for the organisational mechanism is correlational, and the direction of the McKinsey association is not established by the survey that reports it.5 The NANDA report attributes the divide in part to the tools themselves.1 For agentic work in particular, the trial, the benchmark and the Gartner analyst’s statement all place the constraint in the model.111213
The objection weakens where the tool is held constant. In the customer-support study, a tool the employer had built into the job raised output by 14 per cent on average, and its effect ranged from 34 per cent for novice workers to minimal for the most experienced, so the return on the same tool depended on where in the workforce it was placed.9 In the Danish data, the same class of chatbots delivered 10 to 40 per cent more to workers whose employers encouraged them.8 The objection also leaves the first finding untouched: most firms report no measured effect, including firms that use tools whose task-level value is established in peer-reviewed work. The organisational explanation can be tested where the tool is fixed and the deployment varies, and in those settings it has so far held.
The window that has not closed
The questions that would settle the matter are being measured now. The Census Bureau’s revised question has run since November 2025 and gives a consistent series of firm-level adoption from that date.3 The panels behind “Firm Data on AI” ask the same productivity question repeatedly, and the October 2026 revision already shows executives’ weekly use of AI rising from 1.5 to 2.1 hours since winter 2026.2 The executives surveyed expect the effect on productivity to grow over the next three years, and Gartner’s forecast for agentic projects falls due at the end of 2027.713
Those dates give the thesis its terms. If firms that report organising around AI show measured productivity effects in the panel data while firms that do not continue to report none, the organisational reading will have passed in the form that matters. If measured effects stay near zero across firms whatever their practices, the constraint will lie elsewhere. The 95 per cent figure, drawn from 52 interviews and a six-month window in the first half of 2025, records an early moment in that sequence, at a point where, on the J-curve account, the costs of reorganisation would have arrived before its returns could be measured.110
- Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025”, July 2025 (research report, not peer-reviewed): “95% of organizations are getting zero return”; methodology of a review of over 300 publicly disclosed initiatives, interviews with 52 organisations and 153 survey responses, January–June 2025; success defined as deployment beyond pilot with measurable KPIs, return measured six months post-pilot; adoption funnel; stated limitations. Copy of the report hosted by Webber Wentzel: https://webberwentzel.com/News/Documents/2025/MIT1757411281972.pdf (accessed 10 November 2026) ↩︎
- Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom, Steven J. Davis et al., “Firm Data on AI”, NBER Working Paper 34836, February 2026, revised October 2026 (working paper, not peer-reviewed): abstract of the revised version; nearly 9,000 executives in seven countries; 78% of firms use AI; productivity growth raised by 0.2% a year over the last three years and an expected 0.6% a year over the next three; 2% of firms account for over 50% of spending; executive use of 2.1 hours a week, up from 1.5. https://www.nber.org/papers/w34836 (accessed 10 November 2026) ↩︎
- Adam Grundy, Cory Breaux and Dhanapati Khatiwoda, US Census Bureau, “Large Firms With at Least 20 Employees Biggest AI Users”, 26 May 2026 (official): Business Trends and Outlook Survey; 19.8% of businesses using AI as of 3 May 2026; 37% of firms with 250 or more employees; revision of the core AI question from November 2025; expected use by sector. https://census.gov/library/stories/2026/05/ai-use-businesses.html (accessed 10 November 2026) ↩︎
- Jeffrey S. Allen, FEDS Notes, Board of Governors of the Federal Reserve System, note on monitoring AI adoption in the US economy, 3 April 2026 (official): comparison of the Census BTOS, the Atlanta Fed Survey of Business Uncertainty and worker surveys; 78% of the labour force at adopting firms (employment-weighted), 69.4% equal-weighted; reasons the estimates differ. https://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html (accessed 10 November 2026) ↩︎
- Alex Singla, Alexander Sukharevsky, Bryce Hall, Lareina Yee and Michael Chui, McKinsey & Company, “The state of AI in 2025: Agents, innovation, and transformation”, 5 November 2025 (consultancy survey): 1,993 participants in 105 countries, 25 June to 29 July 2025; 88% regular use; 39% report enterprise EBIT impact; about 6% high performers; workflow redesign. https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-state-of-ai-2025 (accessed 10 November 2026) ↩︎
- Jeremy Korst, Stefano Puntoni and Prasanna Tambe, Wharton Human-AI Research and GBK Collective, “Accountable Acceleration: Gen AI Fast-Tracks Into the Enterprise”, 28 October 2025 (business-school survey report): 82% weekly use; three in four leaders see positive returns; 72% formally measure return. https://knowledge.wharton.upenn.edu/special-report/2025-ai-adoption-report (accessed 10 November 2026) ↩︎
- Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom, Steven J. Davis et al., “Firm Data on AI”, March 2026 version (working paper, not peer-reviewed): surveys in the US, UK, Germany and Australia fielded November 2025 to January 2026; shares of firms using AI; shares reporting no effect on productivity and employment; average past and expected effects by country. https://wfhresearch.com/wp-content/uploads/2026/03/Firm-Data-on-AI-March-2026.pdf (accessed 10 November 2026) ↩︎
- Anders Humlum and Emilie Vestergaard, “Large Language Models, Small Labor Market Effects”, Becker Friedman Institute Working Paper 2025-56, May 2025 (working paper, not peer-reviewed): about 25,000 workers in 7,000 Danish workplaces; time savings of 2.8% of hours; no significant effect on earnings or recorded hours; effects of employer encouragement; pass-through of 3–7%; new tasks for 8.4% of workers. https://bfi.uchicago.edu/wp-content/uploads/2025/04/BFI_WP_2025-56-3.pdf (accessed 10 November 2026) ↩︎
- Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, “Generative AI at Work”, Quarterly Journal of Economics 140(2): 889–942, 2025, doi:10.1093/qje/qjae044 (peer-reviewed): 5,179 customer-support agents; issues resolved per hour +14% on average, +34% for novice and low-skilled workers. Working-paper version: https://www.nber.org/papers/w31161 (accessed 10 November 2026) ↩︎
- Erik Brynjolfsson, Daniel Rock and Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies”, American Economic Journal: Macroeconomics 13(1): 333–72, January 2021 (peer-reviewed): TFP 15.9% higher than official measures at end-2017 after adjusting for hardware- and software-related intangibles. https://www.aeaweb.org/doi/10.1257/mac.20180386 (accessed 10 November 2026) ↩︎
- METR, study of the impact of early-2025 AI tools on experienced open-source developer productivity, 10 July 2025 (randomised controlled trial, not peer-reviewed): 16 developers, 246 issues; 19% longer with AI allowed; forecast and perceived speed-ups; authors’ caveats and notice of a continuation with results current to early 2026. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ (accessed 10 November 2026) ↩︎
- Frank F. Xu, Graham Neubig et al., “TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks”, arXiv:2412.14161, version 2 of 19 May 2025 (preprint): most competitive agent completes 30% of tasks autonomously; long-horizon tasks beyond current systems. https://arxiv.org/abs/2412.14161v2 (accessed 10 November 2026) ↩︎
- Zaheer Kachwala, Reuters, “Over 40% of agentic AI projects will be scrapped by 2027, Gartner says”, 25 June 2025, as published by The Star (press report of an analyst estimate): Gartner forecast; about 130 genuine agentic AI vendors; statements by Anushree Verma. https://www.thestar.com.my/tech/tech-news/2025/06/25/over-40-of-agentic-ai-projects-will-be-scrapped-by-2027-gartner-says (accessed 10 November 2026) ↩︎
- Context
- The MIT Project NANDA report "The GenAI Divide" (July 2025) and its figure that 95% of organisations get zero return from generative AI, set against executive panel surveys run with central banks (2025–2026), US Census Bureau adoption data, consultancy and business-school surveys, and controlled studies of AI at work, as of 10 November 2026.
- Causes
- The complementary investment and reorganisation that general-purpose technologies require; differences in how employers deploy, encourage and integrate AI tools; the limits of current models on long, multi-step and autonomous tasks; differences in survey design, weighting and definitions of return.
- Actors
- MIT Project NANDA; the US Census Bureau; the Federal Reserve Board and the Federal Reserve Bank of Atlanta; the Bank of England, the Deutsche Bundesbank and Macquarie University business panels; McKinsey & Company; the Wharton School and GBK Collective; Gartner; METR; the researchers whose studies are reviewed; firms and workers using generative AI.
- Perspectives
- Separates what each survey measures from how its headline figure is used; distinguishes peer-reviewed, working-paper, survey, benchmark and analyst evidence; separates within-firm causal evidence from cross-firm correlation; records the strongest counter-argument; makes no judgement of companies, vendors or researchers.
- Relationships
- Second Inference Dispatch. Links to Inference (Nvidia’s Board Now Investigates a National Security Question, https://titani.ca/dispatch/nvidias-board-now-investigates-a-national-security-question/) and Dry Powder (AI’s Money Loop Concentrates Risk on Few Balance Sheets, https://titani.ca/dispatch/ais-money-loop-concentrates-risk-on-few-balance-sheets/).

