The GenAI Divide: State of AI in Business 2025 (MIT NANDA)
TL;DR
The source of the widely quoted “95% of organizations get zero return” figure. Four authors from Project NANDA at MIT label it “Preliminary Findings” (p. 2). The research period was January to June 2025 (p. 2). The cover says July 2025, and the report reached the press in mid-August 2025.
- The headline. “Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return” (p. 3). The same paragraph says “Just 5% of integrated AI pilots are extracting millions in value” (p. 3).
- The funnel. For general-purpose LLMs, 80% of organisations investigated, 50% piloted and 40% “successfully implemented”. For embedded or task-specific GenAI it was 60%, 20% and 5% (exhibit, p. 6; text, p. 3).
- The success definition. “We define successfully implemented for task-specific GenAI tools as ones users or executives have remarked as causing a marked and sustained productivity and/or P&L impact” (p. 7). The appendix gives a second definition: “deployment beyond pilot phase with measurable KPIs. ROI impact measured 6 months post-pilot” (p. 24).
- The authors’ own caveat. “These figures are directionally accurate based on individual interviews rather than official company reporting. Sample sizes vary by category, and success definitions may differ across organizations” (p. 6).
- Buy beats build. External partnerships reached deployment “~67% of the time, compared to ~33% for internally built tools” (p. 19). The authors add that the gap “may reflect organizational capabilities rather than implementation approach alone” (p. 19).
- Shadow AI. Only 40% of companies bought an official LLM subscription, yet workers at “over 90%” of the companies surveyed use personal AI tools for work (p. 8).
Key data
| Figure | Exact wording or definition | Page |
|---|---|---|
| 95% | “of organizations are getting zero return” | 3 |
| $30–40 billion | ”enterprise investment into GenAI” (no source given) | 3 |
| 5% | “of integrated AI pilots are extracting millions in value” | 3 |
| 60% → 20% → 5% | organisations that evaluated, piloted, and put into production enterprise-grade (custom or vendor) GenAI systems | 3, 6 |
| 80% → 50% → 40% | the same funnel for general-purpose LLMs (ChatGPT, Copilot); text says “nearly 40 percent report deployment” | 3, 6 |
| 5% | “of enterprises have AI tools integrated in workflows at scale” (Myth 2) | 7 |
| ~83% | pilot-to-implementation rate for generic LLM chatbots | 7 |
| 2 of 8 / 7 of 9 | sectors with structural disruption vs sectors without (the report uses both counts; the bar chart shows seven industries, the table eight) | 3, 4, 5 |
| 0.5 of 5 | Healthcare & Pharma on the disruption index: “Documentation/transcription pilots; clinical models unchanged” | 5 |
| ~67% vs ~33% | deployment rate, external partnerships vs internal builds | 19 |
| 40% vs 90%+ | companies with an official LLM subscription vs companies whose workers use personal AI tools | 8 |
| ~50% (text) / ~70% (survey) | share of a hypothetical $100 GenAI budget going to sales and marketing | 9, 10, 20 |
| 90 days vs 9+ months | pilot-to-implementation time, top mid-market firms vs enterprises (>$100M revenue) | 7 |
| 5–20% | workforce reductions in customer support and administrative processing at affected companies | 21 |
Method and definitions
- Design (p. 2): “a systematic review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected across four major industry conferences.” The appendix repeats this as “52 structured interviews … 300+ public AI initiatives … surveys with 153 leaders” (p. 24). The four conferences are not named.
- Who answered. Executive sponsors and frontline users (p. 11). A “lightweight survey supplemented these interviews” (p. 25). The interview guides are printed (pp. 25–26).
- Unit. The 95% is a share of organizations. The 5% figures are shares of pilots, of organisations reaching production, and of enterprises with tools integrated at scale. The report moves between these without reconciling them.
- Enterprise means a firm with more than $100 million in annual revenue (p. 7).
- Disruption index. Each industry is scored 0–5 on five indicators, “triangulated from public indicators and interview-derived assessments” (p. 4).
- Reviewers. “Pradyumna Chari, Project NANDA” (p. 2), who is also a co-author. There is no external peer review.
Scope and reliability
- Small, self-selected sample. 153 survey respondents met at conferences and 52 interviewed organisations. The authors list selection bias themselves (p. 24): firms willing to talk about AI problems “may systematically differ” from those that decline.
- Self-report, not company data. The authors call the figures “directionally accurate” (p. 6). Several exhibits are coded from interview notes (p. 15), and budget sub-categories “should be treated as directional at best” (p. 10).
- Two success definitions. The 95% rests on “zero return”. The p. 7 definition requires a “marked and sustained” impact remarked on by users or executives. The p. 24 definition requires measurable KPIs six months after the pilot. A six-month window “may be insufficient”, “potentially understating success rates” (p. 24).
- A stricter bar than other surveys. Stefano Puntoni, a co-author of the GBK 2025 report, told Big Technology: “Our definition is very different from the MIT report. Theirs is much more stringent” (Kantrowitz, 31 Oct 2025; raw at
raw/articles/2025-10-31-kantrowitz-big-technology-are-74-percent-of-businesses-seeing-genai-roi.md). - Method reported differently in the press. Fortune’s 18 August 2025 story (read with WebFetch) described “150 interviews with leaders, a survey of 350 employees, and an analysis of 300 public AI deployments.” The PDF says 52 organisations and 153 senior leaders. This page follows the PDF.
- Interest in the conclusion. Project NANDA builds agent-interoperability infrastructure. The report’s prescription ends with “Systems like NANDA, MCP, and A2A represent early infrastructure for this web” (p. 23).
- Provenance. No copy on an MIT domain was found on 7 November 2026; nanda.media.mit.edu redirects to the Media Lab group page. The file used is a third-party copy whose metadata names Aditya Challapally as author (created 14 July 2025).
Dynamic-capabilities reading
digital-seizing/balancing-digital-portfolios: the report is about scaling up, from pilot to production. Its main contrast, external partnerships against internal builds, is the cell’s “balancing internal and external options.”digital-transforming/navigating-innovation-ecosystems: the successful buyers “act like BPO clients, not SaaS customers”, co-develop with vendors, and find tools through referrals and existing partners (pp. 17–19).contextual/internal-barriers: the rated barriers to scale are change management, lack of executive sponsorship, poor user experience, output quality and unwillingness to adopt new tools (p. 11). The report’s own diagnosis is a “learning gap”: tools that do not keep feedback or adapt (p. 3).
How it connects
- enterprise-ai-adoption: a value-capture figure, measured on a much stricter bar than adoption surveys.
- ai-maturity-measurement-comparison: a strict-definition instrument with a small conference sample. Its 95% and 5% figures use different denominators.
- micro-productivity-trap: the report’s split between individual productivity (ChatGPT, Copilot) and P&L impact (p. 3) is the trap stated as data.
- ai-knowledge-hiding: the “shadow AI economy” (p. 8), workers using personal accounts “often without IT knowledge or approval.”
- agentic-web: the closing section names the “Agentic Web” (pp. 22–23).
- Allen at the AWS London Executive Forum quotes the headline. His three failure modes (workflow alignment, wrong targets, missing integration) follow the report’s text (pp. 3, 6, 9–10).
- Allen, Brozovich and Davis at re:Invent 2025: the earlier edition of the same talk with the same citation. The Sydney edition repeats it as “95% of AI pilots fail.”
- BBC AI Decoded: Peter Grant prefers the PwC CEO survey to this report on sample-size grounds.
- GBK 2025, BCG 2025 and Accenture 2025: other 2025 measures of AI value, each with its own definition and sample.
What was actually ingested
All 26 pages of the PDF, converted with pdftotext -layout. Exhibits whose values are drawn rather than typed (pp. 11, 12, 13, 15) lost their numbers in conversion; this page quotes only figures stated in the text. The copy comes from mlq.ai, a third-party host. Fortune’s 18 August 2025 article was read only for its description of the method. The Big Technology article is kept in raw/ as secondary commentary.
Linked entities and concepts
- Jonathan Allen: cites the report in three editions of his AWS talk.
- PwC: the survey the BBC panel prefers to this one.
- enterprise-ai-adoption, micro-productivity-trap, ai-knowledge-hiding, agentic-web: concepts.
- ai-maturity-measurement-comparison: thread.
Dangling (single-source mention, deferred): Aditya Challapally, Chris Pease, Ramesh Raskar, Pradyumna Chari (authors); MIT NANDA / Project NANDA (organisation; also named in the bodies of the three AWS sources and the BBC source); MIT Project Iceberg (cited p. 22); Stefano Puntoni (quoted via Big Technology; author of the Wharton/GBK report); Alex Kantrowitz (Big Technology); Fortune.