The falling cost of finding out.

It has never been cheaper to find out whether an idea works. A working demonstration of almost anything — a screening tool, a product concept, a new way of reading a market — can be put together in days now, sometimes in an afternoon. Not long ago the same experiment meant a vendor contract, a budget line, and a quarter of somebody’s attention. This is a real change in what a small organization can attempt, and it deserves more credit than it usually gets.

It also explains a lot of disappointment. Bain surveyed 951 companies in June and found that among those that actually measured their AI savings, nearly 40% came in under 10%; most had been aiming for 11 to 20. Ninety percent are raising their budgets anyway. The year before, MIT put the share of corporate AI pilots with no measurable P&L effect at 95%. Numbers like these get read as a verdict on the technology. They are closer to a verdict on how organizations use demonstrations.

A demo runs on the documents its builder picked, tended by the person who made it, with nothing riding on the outcome. Real work arrives differently: longer, less consistent, formatted by whoever sent it, read by people on deadline who never saw the demo and will drop anything that fails them twice. Between those two situations sits everything that actually takes time — the exceptions, the second month, the maintenance, the person who answers for the thing when it is wrong. None of that shows up in a demonstration. Nearly all of it is the work.

The gap is visible wherever someone measures honestly. In the same Bain survey, only 7% of companies run fully autonomous AI agents in production; the standard arrangement is a human approving what the system proposes. Most business cases assumed the economics of automation and got the economics of supervision. MIT’s less-quoted finding pointed at how the gap does get crossed: deployments customized to an organization’s own work, usually with an outside partner, succeeded about twice as often as generic internal efforts. Customizing is not polish. It is the crossing.

For people who build things — and most serious investors are also builders, of ventures, of firms, of capabilities — the falling cost of demonstration is an opportunity dressed up as a problem. You can now afford to be wrong about ten ideas in the time it once took to be wrong about one. A thesis can meet real documents before anyone commits capital to it. A product idea can meet users in weeks. The discipline this asks for is modest but unforgiving: know what question each experiment is supposed to answer. A demo without a question is theater with a budget.

Which changes what a dead pilot means. One that ends in a clear no — the documents are too inconsistent, the failure mode cannot be tolerated, the upkeep would eat the saving — settled something real at a low price. The expensive failures are the demonstrations that impressed everyone and answered nothing. The measure of an experiment was never the applause. It is whether, when it ended, you knew what to build.