Enterprise AI ROI: Why MIT Says 95% of Pilots Fall Short

The number that rattled the AI trade

A single statistic has become shorthand for enterprise AI disillusionment. A July 2025 report from MIT Media Lab's Project NANDA, The GenAI Divide: State of AI in Business 2025, found that roughly 95% of corporate generative AI pilots produced no measurable impact on profit and loss. This was despite an estimated $30–40 billion in enterprise spending. 12 Only about 5% of integrated pilots are pulling in meaningful value, which the authors describe as millions of dollars, while the rest stall. 2

The report's reach went well beyond its preliminary status. One retrospective argues the figure did what years of AI skepticism had not: it moved markets. 3 That reaction alone makes it worth examining what the study actually measured.

What the study looked at

The findings rest on 52 executive interviews, surveys of 153 leaders, and an analysis of 300 public AI deployments. 1 The central concept is the "GenAI Divide," a gap between widespread experimentation and rare transformation. More than 80% of organizations have piloted tools like ChatGPT, but few have turned those pilots into systems that change the bottom line. 1

The authors say the split applies on both sides of the market. It runs across buyers, from large enterprises to small businesses, and across builders, including startups, vendors, and consultancies. 2 According to coverage of the report, the steepest drop-off comes between pilot and production. 2

What the 5% do differently

The more useful part of the report is its account of what separates winners from everyone else. The sources mostly agree here. They differ only in which factor they emphasize.

• Adaptive systems over static tools. Success depends on systems that can learn and adapt, not off-the-shelf tools that stay frozen. 2
• Buying beats building. Purchased or partnered tools succeeded roughly twice as often as internal builds, at about 67% versus 33%. 3 Vendor partnerships come up repeatedly as a key to success. 1
• Back office over front office. Back-office automation produced clearer returns. Sales and marketing pilots absorbed an estimated 50–70% of budgets yet delivered less. 31

One analysis puts it bluntly: the failure was organizational, not technological. 3 That framing matters. It suggests the 95% figure says less about whether the models work and more about how companies choose, integrate, and measure them.

The counter-signal from workers

The pessimistic headline sits awkwardly beside survey data from people actually using AI tools.

A Productboard study, run with research firm UserEvidence, surveyed 379 product professionals at enterprises with 500 or more employees. Every team surveyed uses AI, and 96% use it consistently. Nearly half describe it as "deeply embedded" in their workflows. 4 The same study warns that governance is falling dangerously behind adoption. 4

Lenny's Newsletter's large-scale productivity survey is even more upbeat. Fifty-five percent of respondents said AI exceeded their expectations, and nearly 70% said it improved the quality of their work. More than half reported saving at least half a day per week on their most important tasks. 5 That survey also notes that engineers favor purpose-built tools such as Claude Code. Agentic platforms, though, have seen slow real-world adoption in 2025. 5

The MIT report itself points to the same pattern. It describes a "shadow AI economy" in which employees use consumer tools without employer approval. 1

Reading the divide

These findings are not really contradictory. They measure different things. The worker surveys capture individual productivity: time saved, better drafts, faster prototypes. The MIT study asks whether formal, budgeted enterprise programs show up on the P&L. Both can be true at once. Employees get real value from personal tools, while sanctioned corporate pilots, often aimed at flashy front-office uses, fail to scale.

That is my interpretation, but the evidence supports it. Adoption is nearly universal at the individual level. 45 Value is concentrated where companies pick narrow, back-office problems and partner with vendors rather than build from scratch. 13 Unapproved consumer tools are filling the gap. 1

Some caution is warranted. The MIT figure comes from a preliminary report with a modest interview and survey base. 13 The productivity surveys rely on self-reported gains from enthusiastic users. Neither side offers the final word.

The practical lesson for executives is not that AI fails. Their organizations may be measuring the wrong projects, funding the wrong functions, and overlooking where employees are already finding value. Closing the divide looks less like a model upgrade and more like a management problem.