MIT "GenAI Divide" report (2025)
Last reviewed
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,481 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,481 words
Add missing citations, update stale details, or suggest a clearer explanation.
The "GenAI Divide" report, formally titled The GenAI Divide: State of AI in Business 2025, is a research report published in July 2025 by MIT NANDA, an initiative based at the MIT Media Lab. The report examines why most enterprise generative AI initiatives fail to reach production or show financial impact, despite heavy corporate spending. It became widely known for a single statistic: that roughly 95 percent of organizations studied were getting "zero return" from their generative-AI pilots, while only about 5 percent were capturing meaningful value. The report contrasted this small group of successes against the large majority that stalled, a split the authors named the "GenAI Divide." [1][2]
Reported by Fortune on August 18, 2025, and amplified across business and technology media, the figure became one of the most-cited data points in the 2025 debate over enterprise return on investment on AI. It was credited with helping to trigger a brief stock-market wobble in major AI-linked equities later that month and feeding wider speculation about an AI bubble. At the same time, several analysts criticized the report's methodology, its data transparency, and the way the 95 percent figure was framed. [1][3][4]
The report was produced by Project NANDA (an acronym for "Networked Agents and Decentralized AI"), a research effort led by MIT Media Lab professor Ramesh Raskar. NANDA's broader mission is to build infrastructure for an "Internet of AI Agents," a decentralized network in which autonomous AI agents can discover, authenticate, and transact with one another. The State of AI in Business 2025 report is a separate output examining the current enterprise adoption of generative AI rather than NANDA's core agent-protocol work. [4][5]
The report lists four authors: Aditya Challapally (the lead author and the figure most quoted in press coverage), Chris Pease, Ramesh Raskar, and Pradyumna Chari. The document describes itself as "Preliminary Findings from AI Implementation Research from Project NANDA" and carries a disclaimer that the views expressed are those of the authors and do not reflect the positions of any affiliated employers. [2]
According to the report's own methodology note, the findings draw on a multi-method research design covering a research period of January to June 2025: [2]
| Method | Scope |
|---|---|
| Systematic review of public AI initiatives | More than 300 publicly disclosed deployments |
| Structured interviews | Representatives from 52 organizations |
| Survey responses | 153 senior leaders, collected across four major industry conferences |
Some early news coverage cited different figures (for example, "150 interviews" and a "survey of 350 employees"), but the numbers above are those stated in the published report. The discrepancy between press accounts and the source document later became one focus of methodological criticism. [1][3]
The report's executive summary states that "despite $30 to 40 billion in enterprise investment into GenAI," about 95 percent of organizations were "getting zero return," while "just 5% of integrated AI pilots are extracting millions in value." The authors argued that this divide did "not seem to be driven by model quality or regulation, but seems to be determined by approach." [2]
The report drew a sharp distinction between adoption and transformation. It found that general-purpose tools such as ChatGPT and Microsoft Copilot were widely used, with more than 80 percent of organizations having explored or piloted them and nearly 40 percent reporting deployment. But these tools mainly improved individual productivity rather than profit and loss. By contrast, custom or vendor-built enterprise systems fared poorly: roughly 60 percent of organizations evaluated such tools, about 20 percent reached a pilot stage, and only around 5 percent reached production. [2]
The report also argued that industry-level disruption remained limited, finding that only two of eight major sectors it scored (technology and media/telecom) showed clear signs of structural change, while the others, including financial services, healthcare, and consumer/retail, remained on what it called the "wrong side" of the divide. The report summarized four patterns: limited disruption across sectors, an "enterprise paradox" in which large firms led in pilot volume but lagged in scaling, an "investment bias" toward visible top-line functions, and an "implementation advantage" for external partnerships. [2]
The report's central explanation for the divide was what it called the "learning gap." It argued that the main barrier to scaling generative AI was "not infrastructure, regulation, or talent" but learning: most generative-AI systems "do not retain feedback, adapt to context, or improve over time." Generic chatbots succeeded for individuals because of their flexibility, the authors wrote, but stalled in enterprise settings because they did not learn from or adapt to specific workflows, producing brittle integrations misaligned with day-to-day operations. [1][2]
Two practical findings followed from this thesis. First, the report concluded that buying tools from specialized vendors and forming partnerships succeeded far more often than building in-house: it stated that external partnerships saw roughly twice the success rate of internal builds. (In press interviews, Challapally put the purchased-tool success rate at about 67 percent, against roughly one-third as often for internal builds.) Second, the report identified a budget-allocation problem: organizations directed more than half of generative-AI budgets toward sales and marketing, yet the report found the largest measurable returns in back-office automation, such as reducing business-process-outsourcing and external-agency costs. [1][2]
The report described this misallocation as a strategic error rather than a technology failure, suggesting that "learning-capable systems, when targeted at specific processes, can deliver real value, even without major organizational restructuring." [2]
After Fortune publicized the report on August 18, 2025, the 95 percent statistic spread rapidly. In the days that followed, shares of several AI-linked companies, including Nvidia, Microsoft, Alphabet, Palantir, and CoreWeave, fell, and commentators asked whether an "AI bubble" was deflating. Coverage noted that the sell-off coincided with separate remarks by OpenAI chief executive Sam Altman warning of a bubble in privately held AI startups; some observers argued that traders conflated the two stories and that the market reaction said more about investor sentiment than about the report's actual conclusions. [3][4]
The report also drew substantial criticism. Critics argued that the 95 percent figure was asserted with little visible supporting data and that the report disclosed few details about its sample's demographics or how data was collected; some charts were published with unlabeled axes. Wharton professor Kevin Werbach and others questioned the framing, noting that the explicit "5 percent" success figure in the report referred specifically to custom enterprise AI tools reaching production, a narrower category than "all AI pilots," so the inverse "95 percent failure" headline arguably conflated different measures. [4]
A second line of criticism concerned a potential conflict of interest. Because NANDA's own mission is to promote agent-based, decentralized AI infrastructure (work that builds on protocols such as Anthropic's Model Context Protocol and Google's Agent2Agent), some critics suggested the project had an incentive to portray current enterprise AI approaches as failing. The report was also issued as preliminary findings rather than as peer-reviewed academic research, and several writers called on the authors to release the full underlying data. Defenders of the report responded that the core message, that the problem lay in integration and workflow rather than in model capability, was sound regardless of the precise percentage, and that the report had been widely misread as a verdict on AI itself rather than on how companies were deploying it. [3][4]
The "GenAI Divide" report became a reference point in the 2025 discussion of enterprise AI adoption and ROI skepticism. Its lasting contribution was less the exact 95 percent figure, which remained contested, than its framing of a divide between a small set of organizations capturing real value and a large majority stuck in pilots, and its argument that the decisive factors were integration, workflow fit, build-versus-buy strategy, and budget allocation rather than raw model performance. For executives, it crystallized a growing concern that high adoption of generative-AI tools was not translating into measurable profit-and-loss impact. The episode also illustrated how a single, vividly framed statistic from a non-peer-reviewed report could move markets and shape public narrative well ahead of careful scrutiny of its methodology. [1][3][4]