Companies are racing to deploy generative AI — but how often does a pilot actually move the bottom line?

This explores how often enterprise generative AI pilots turn into money you can measure, and why most of them don't. The headline figure in the collection is stark. MIT NANDA found that about 95% of enterprise GenAI pilots produce no measurable effect on profit and loss, while roughly 5% extract millions Why do most enterprise AI pilots fail to deliver returns?. The surprise is what explains the split. It isn't which model a company picked or the regulations it faced. It's how the company deployed the tool: whether it bought or built, whether line managers rather than a central lab drove adoption, and whether the tools could adapt over time. The same pattern shows up well beyond business pilots. An interdisciplinary review of generative AI's effect on inequality reaches the same conclusion: outcomes depend on access, integration, and incentives, not on the technology itself Does generative AI inevitably worsen or reduce inequality?.

Why would capable models so often fail to pay off? One answer is that the field has been measuring the wrong thing. An analysis of 960 real occupational workflows found that AI agents do well on contest-style benchmarks but break down on long, multi-step professional tasks Why do agent benchmarks not predict real economic value?. A company that picks a tool for its leaderboard scores may be buying skill at puzzles rather than at its own work. Related research finds that what predicts success on long tasks is persistence, meaning the willingness to keep testing, fixing, and retrying, rather than a strong first attempt. Most models quit early or waste their time budget What predicts success in ultra-long-horizon agent tasks?. Other work suggests that smaller, cheaper models trained on real execution can be the better economic choice, because cost and delay add up across a whole workflow Does model efficiency matter more than peak capability for real work?.

A second answer is that productivity gains can be real and still not show up in the books, or can be offset somewhere else. In a randomized experiment with 776 Procter & Gamble professionals, individuals using AI matched the output of two-person teams working without it, and their solutions drew more evenly on different professional backgrounds Can generative AI replace the benefits of having a human teammate?. That is a real gain, but it is a gain in output quality, not a revenue line. FactSet is a sharper warning. Its GenAI integration made analyst reports about 25% richer in sources and coverage, yet forecast accuracy fell, because analysts couldn't absorb the extra material. A machine model given the same inputs lost no accuracy Does AI-enriched analyst reports improve forecast accuracy?. A dashboard counting "more content produced" would have called that pilot a success.

A third answer concerns hidden overhead. Larsen-Ledet argues that today's AI tools make workers keep restating context the system ought to infer. This tedious explaining eats into whatever time the tool saves, and the tool only becomes a real workplace asset when that overhead goes away Can generative AI reduce overhead work that adds no value?.

The short answer is rarely, by the main measure in this collection: about 1 in 20. But the more useful lesson is that pilots seem to fail for organizational and measurement reasons more than technical ones. These include picking tools by benchmark, ignoring the cost of human attention, and treating richer output as better output. That points to where a company could change its odds. Note that the collection's direct evidence on the 95% figure comes mainly from a single study, so treat it as a strong signal rather than a settled base rate.