Listen to this article
MIT's State of AI in Business 2025 report landed in late August like a bomb, because its headline number is brutal: 95 percent of enterprise GenAI pilots are failing to deliver any measurable P&L impact. Two and a half years into the generative AI era, with billions committed, nineteen out of twenty pilots produce nothing a CFO can find on a financial statement.
The number got the coverage. The diagnosis deserves it more. In Forbes, Jason Snyder pulled the thread the report leaves hanging: the failed 95 percent did not fail because the models were weak. They failed because the organizations deploying them avoided friction - the review steps, the verification, the accountability - and bought tools that demo well precisely because they skip all of it.
Two related but distinct measures from the same report: the share of pilots that never reach the P&L, and the share of custom enterprise tools that make it out of the pilot phase at all.
The shadow AI economy
The strangest finding in the report is that AI adoption is not actually failing - official AI programs are. MIT found that 90 percent of employees already use personal GenAI tools for work, while only about 40 percent of their companies have an enterprise GenAI subscription. The work is getting done on personal ChatGPT accounts, invisible to IT, with no logging, no review, and no accountability. MIT calls it the shadow AI economy.
Note the populations differ: the first bar measures individual employees, the second measures firms. Both point the same direction - usage is racing far ahead of anything the organization can see or govern.
That gap is not a temporary lag that procurement will close next quarter. It is the natural result of buying frictionless tools: if the sanctioned tool and the personal tool both skip review and verification, employees will pick the one without a login wall. The organization gets all of the risk and none of the visibility.
What the surviving 5 percent do differently
The report's pattern for the projects that actually work is consistent. External vendor partnerships succeed roughly twice as often as internal builds - not because outside code is better, but because a purchase forces the friction internal projects dodge: requirements, acceptance criteria, someone accountable for the outcome. Generic chatbots hit 83 percent adoption for trivial tasks like drafting emails, then stall the moment the workflow gets complex enough that a wrong answer costs money. The tools that survive are the ones embedded in real workflows, with verification built in and a human answerable for the result.
The failed 95 percent avoided friction. The surviving 5 percent institutionalized it.
This is the inversion worth sitting with. Friction has been treated as the enemy of AI ROI since the first pilot kicked off - every review step framed as drag, every approval as a blocker. MIT's data says the opposite: friction is what separates the pilots that pay from the pilots that die. What actually kills ROI is ungoverned frictionlessness - output nobody verified, actions nobody approved, and a shadow workforce of personal AI accounts nobody can audit.
The AuthorityGate take
The friction MIT found in the surviving 5 percent is not bureaucracy. It is specific, structural, and it sits in one place: between an AI's output and its effect. Review before the change lands. Verification before the answer ships. A named human accountable for the consequential calls. Those are gates, whether or not anyone calls them that.
That is why we built change validation as a pipeline rather than a policy document: every change, whether it comes from a person, a pipeline, or an AI agent, passes the same checks, and the ones that matter route to a human who signs before production. The 95 percent failed by engineering the friction out. The 5 percent got paid by engineering it in - on purpose, in the pipeline, where it cannot be skipped on a busy day.
The 95 percent number will get quoted for years, mostly as evidence that GenAI is overhyped. That is the wrong reading. Ninety percent of employees are already using these tools daily because they work. The failure is organizational: companies keep buying the frictionless demo and wondering why nothing survives contact with production. The report's real message fits in a sentence - if your AI program has no friction in it, you are not in the 5 percent.
Sources
Go deeper
Every agent action, validated before it takes effect
AuthorityGate's newsletter breaks down real AI incidents and the governance failures behind them. Our configurable 8-gate validation model is how organizations keep a named human accountable for what their AI actually does.