Gartner put a number on it in June 2025: more than 40 percent of agentic AI projects will be canceled by the end of 2027, for "escalating costs, unclear business value or inadequate risk controls." Fifteen months later, on September 16, 2026, the same firm forecast worldwide AI spending at $2.7 trillion for 2026, up 49.5 percent on the year. Read those two together. The money is not slowing down. The projects are still dying.
Here is the claim, up front: most agentic AI project failure is not failure. The agent usually did the task. The project got canceled because it reached a gate it could not pass, and it could not pass because the pilot never collected the evidence the gate asks for.
Look at Gartner's three reasons again. Escalating cost. Unclear value. Inadequate risk controls. They read like three problems with three owners: finance, the sponsor, and risk. They are one problem. Each is a question the production gate asks. Each is a question a capability pilot cannot answer.
The pilot proves the wrong thing
A pilot is built to answer one question: can the agent do this? So it is scoped to show capability. Fifty tickets, a clean dataset, a demo for the steering committee. It works. Everyone agrees it works.
Then the production decision arrives and the questions change.
What does one action cost when the agent runs ten thousand of them a day, including the retries, the escalations, and the person who cleans up after the three percent it gets wrong? The pilot measured tokens. It did not measure cleanup.
Which outcomes can we attribute to the agent? If the number moved, how much was the agent and how much was the quarter? The pilot had no baseline and no record of which decisions the agent made alone.
What happens when it is wrong in a way nobody anticipated? Who stops it, how fast, and how do we know it stopped? The pilot never ran a stop drill, because nothing in the pilot was worth stopping.
Three questions. No evidence. The project is not killed because the agent failed. It is killed because nobody can prove it did not, at a price anyone can state.
Klarna measured one thing
Klarna is the clean public example because the CEO said it out loud. In February 2024 the company announced its AI assistant had handled two thirds of customer service chats in its first month, doing the work of 700 agents, and projected a $40 million profit improvement. By May 2025, Sebastian Siemiatkowski told Bloomberg the company was hiring humans again: "cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality."
Notice what that sentence admits. The system did what it was built to do. The evaluation measured cost. Quality was not on the scorecard, so quality arrived late, as a finding nobody had priced. That is Gartner's "unclear business value" in one company's words. Klarna did not cancel its AI; it dialed back a strategy and kept the assistant running. But the mechanism is the one that cancels projects: the metric that got the pilot approved was not the metric the business ends up living with.
Governance is the evidence
This is where the governance people and the delivery people talk past each other. Delivery hears "governance" and pictures a committee. What the production gate wants is a record, and the governance artifacts are the record.
The action log is the cost ledger. If every agent action is logged with its outcome, you can price an action at volume, retries and cleanup included, before you scale.
The approval trail is the attribution. If the pilot records which decisions the agent made alone, which it escalated, and what happened next, you can separate the agent's effect from the quarter's.
The stop drill is the risk control. If you pulled the agent mid-pilot, on purpose, and timed how long it took to halt and how long to confirm it halted, you have the one number risk will ask for.
None of this slows a pilot down. It changes what the pilot is for. A pilot that produces this record walks into the gate with answers. A pilot that produces a demo walks in with a story, and stories get canceled.
What to change on Monday
Rewrite the pilot's success criteria before the next sprint. "The agent completes the task" stays. Add three lines under it: a cost per action measured at the pilot's volume with cleanup included, an attribution method agreed with finance before the pilot starts, and a stop drill run at least once with the time written down. If a pilot cannot produce those three, it will not survive the gate, and learning that in week two is cheaper than learning it in month nine.
I spent 25 years making automated systems trustworthy in production, and the pattern held every time: the systems that shipped were the ones that arrived at review with the measurements already taken. Nobody argued with a log.
The 40 percent are not a verdict on agents. They are a verdict on pilots that proved capability and nothing else. Most agentic AI projects do not fail. They get canceled for lack of evidence, and evidence is something you can decide to collect.