Every AI pilot produces one kind of proof on its own. The invoices arrive. The token bill lands. The engineers' hours get logged. Six months in, you can say to the dollar what the pilot cost. Ask what it's worth and the room goes quiet. That gap is the whole story of agentic AI ROI, and it's how pilots become money pits: proof of cost is automatic, proof of value is something you have to build, and most pilots never build it.
McKinsey's 2026 State of AI survey, reported in September, put numbers on the gap. The share of large companies scaling AI agents in at least one function went from 27 percent to 40 percent in a year. The share of all respondents who can attribute any EBIT impact to AI stayed at 37 percent, flat on the year. About 6 percent qualify as high performers, meaning they attribute at least 5 percent of EBIT to AI and call the impact significant. You will see that last number passed around as "94 percent get nothing." It doesn't say that. It says most companies can't yet show a profit line they'd put their name to. Spending scaled. Proof didn't.
Why a pilot you can't measure can't be killed
Here is the mechanism, and it's uglier than "pilots fail." A pilot with no value metric doesn't fail. Failing would take evidence, and there isn't any, in either direction. So it gets extended. Someone asks for one more quarter to gather more data, and nobody can argue, because nobody can prove it isn't working any more than anyone can prove it is. The cost line keeps climbing, because cost documents itself. The value line stays blank, because nobody instrumented it. That is a money pit. Not a loud failure. A pilot that can't be judged, so it can't be stopped.
I saw this pattern for 25 years in production systems before anyone put the word AI in front of it. The projects that died in extended pilot were rarely the ones with the weakest technology. They were the ones where "success" was going to be defined after the fact.
Agentic AI makes the problem worse in two ways. The cost is harder to see, and the value is easier to fake.
Harder to see, because the budget line says tokens and licenses. It doesn't say reviewer hours. An agent produces an output in seconds. A person decides whether it's right, and that person is the real unit cost. Issue 011 was about making verification cheap enough to actually happen. A pilot that doesn't book the verification at all is reporting a cost figure that is wrong, and wrong in the flattering direction.
Easier to fake, because an agent generates activity by the thousand. "Handled 8,000 tickets" is a count. It isn't value until you know how many were handled correctly, how many came back, and what a person would have cost to do the same work. Activity numbers look like results. They are not results. They are the raw material results are made from, and most pilots stop at the raw material.
Agentic AI ROI is decided before the first dollar
The fix is not a better dashboard at month six. It's three decisions made before the pilot starts, and I would not fund one without them.
First, the value metric, in a unit the business already counts. Not accuracy on a test set. Hours of a named role's time, dollars on a named cost line, days off a named cycle. And a baseline, measured before the agent touches anything, because you cannot show a change against a number you never took.
Second, the full cost, people included. Tokens, licenses, integration hours, and reviewer time per output at the reviewer's rate. If the agent saves an analyst two hours and a senior engineer spends one hour checking the output, both go on the sheet. One of them is usually paid a lot more than the other.
Third, the decision date, and what happens on it. A date, written down, on which the pilot either converts to production with its own budget or stops. Not "we'll review it." Converts or stops. The job of a pilot is to produce a decision, and a pilot with no decision date has no reason to ever end.
That third one is the governance people push back on, because it sounds like a brake. It's the opposite, and Issue 005 made the general case. A kill date is what lets you fund five pilots instead of one, because the losers will actually die and the money comes back. Without it you fund one, it never ends, and it eats the budget for the other four. Governance as an enabler, in the one form every CFO already understands: money not wasted is money you get to spend.
The pilot that gets canceled, which Issue 018 took apart, is not the worst outcome. The worst outcome is the pilot that never gets the chance to be canceled, because nobody set it up to be judged. Proof of cost shows up on its own. Proof of value only shows up if you built it in.