Enterprise AI
Why Most AI Pilots Never Show Measurable ROI
A widely cited 2025 MIT study found most enterprise AI pilots show no measurable P&L impact — a sign of weak measurement, not that AI doesn't work.
The number everyone is citing
A widely reported 2025 MIT research effort found that the large majority of enterprise generative AI pilots were not delivering measurable profit-and-loss impact. The figure spread quickly because it matched what many operating leaders were already sensing about pilot programs that generated excitement but no budget line the CFO could point to.
A failed pilot and a failed measurement are not the same finding
Some pilots failed on technical or product-fit grounds. Many others were simply never set up to prove value in the first place: no baseline, no assigned owner, no agreed metric, and no plan to compare projected value against what actually happened.
Most of these projects had no baseline to fail against
A pilot cannot show ROI it was never asked to measure. If nobody documented the pre-AI cost, cycle time, or error rate before launch, there is no credible way afterward to say what changed or by how much.
Integration and operating cost are usually the missing line item
Model usage is rarely the largest cost once a pilot moves toward production. Integration work, data preparation, governance, monitoring and change management routinely exceed the visible subscription or usage cost, and a business case that omits them overstates the return before the project even launches.
Read carefully, this is an argument for measurement discipline, not against AI
A finding like this is not evidence that AI does not work. It is evidence that most organizations are not yet running AI investment the way they run other capital decisions: with a baseline, an owner, a total cost view, and a plan to check projected value against realized value.
What this means for the next pilot
Before starting, write down the metric that should move, who owns it, what the full cost will be at production scale, and how you will compare the forecast to the outcome. That discipline is what tends to separate a pilot that becomes a case study from one that becomes a statistic.