MIT report finds 95% of generative AI initiatives show no P&L impact

MIT NANDA's report The GenAI Divide: State of AI in Business 2025, published in August 2025, found that about 95% of enterprise generative AI initiatives show no measurable P&L impact. Fortune reported the findings on 18 August 2025, set against $30-40 billion of enterprise investment in generative AI.

What the MIT report found

According to MIT NANDA, only around 5% of custom or embedded AI tools reach production with measurable impact. The report found that deployments run with external partners succeeded about 67% of the time, against about 33% for internal builds. It also found that over 90% of employees use personal AI tools even where official pilots fail, a pattern the report calls shadow AI. The study wasn't peer reviewed and its sample was modest, so the figures show a direction and shouldn't be read as precise measurement. Forbes covered the findings on 21 August 2025.

Why pilots stall before production

Most pilots are designed as demonstrations. A typical one runs on a cleaned extract of data, sits in a browser tab beside the real systems and is operated by the keenest team in the building, with no baseline and nobody's job changing if it works. Each of those choices makes sense for a six-week test, and together they hide the conditions production depends on. When production arrives, so do the edge cases, the integration with your ERP, CRM and ticketing, and the questions from compliance and security that the pilot postponed.

What it means for engineering teams

Your employees already use AI, as the shadow AI figure from MIT NANDA shows, so demand for the tools is already there. The harder work sits in data access, permissions, integration and deciding whose workflow changes when the system runs. Three checks tell you quickly whether an initiative can reach production. It touches real data under real permissions from the start, someone's actual workflow changes and that person agreed in advance, and there's a named owner for month six after the pilot team moves on.

What it means when you hire engineers

The gap MIT NANDA reports between partnered and internal builds puts attention on who does the building. An engineer who has only built demos will tend to optimise for the demo, while an engineer who has shipped AI features into production will ask about permissions, edge cases and integration in the first week. When you hire, it's worth seeing how candidates work in a real codebase, both with AI tools and without them, so you know whether they can carry a feature past the pilot stage. Asking where their AI work broke in production, and what they changed, tells you more than a polished demo does.

Miyagami engineers go through two sixty-minute assessments in a real codebase, the first with AI tools off and the second with their own agents on, and both scorecards are shared with you. You can read more about how we vet.