40% of AI agent projects will be cancelled. Here’s the real reason.
It’s not the models. It’s the human judgment.
Gartner has a number that’s been showing up in board decks all year.
More than 40% of agentic AI projects will be cancelled by the end of 2027. Gartner made that call back in June 2025. A Forbes piece resurfaced it a few weeks ago, and most of the coverage is still treating it like breaking news.
Most people read that number and land on the same conclusion: the tech isn’t ready yet. Unfortunately, that’s the wrong conclusion.
Look at the reasons Gartner actually gave. Escalating costs. Unclear business value. Inadequate risk controls. Model capability isn’t on this list. Neither is hallucination. Neither is technical complexity.
None of the reasons on that list represent a technical issue where the machine fell short. In fact, all of these reasons point to a judgment call that a human leader had to make, that did not go as planned.
Before I go further, a quick, honest plug.
Those judgment calls, the ones that decide whether an AI project ships or gets cancelled, are exactly what I teach in a one-day workshop: the AI Strategy Playbook for Product and Engineering Leaders. You learn battle tested frameworks, get your hands dirty with real problems during the session, and leave with an arsenal of frameworks and mental models that you can apply in your job. The next cohort is August 8 and has only 20 seats.
Grab a seat today! Till Friday, July 31, take $350 off with code 350OFF.
Why we blame the tech?
“The models aren’t good enough yet” is a comfortable thing to believe.
If it’s the tech’s fault, you don’t have to change anything. You wait for the next release and try again.
If it’s your judgment, that’s harder. It means the problem was in how you decided what to build, not in what you built it with.
The hype makes this worse. Gartner even has a name for part of it: agent washing. Vendors take an old chatbot or a script, rebrand it as an “agent,” and sell it into the wave. The bad decision starts before a single line of code - at the moment someone picks a shiny thing and calls it a strategy.
What actually kills these projects?
Take Gartner’s three reasons and look at what each one really is.
Unclear business value. This is a problem-selection failure. The project got greenlit because AI was exciting, not because the problem was worth solving. Nobody asked whether this was a real problem or just a demo waiting to happen. So it ships, it moves nothing, and it gets cut.
Escalating costs. This is a design failure. Nobody decided where the agent stops. Nobody set what a good outcome should cost, or when the system should give up and hand off to a human. An agentic loop with no limit will happily burn your budget trying to be clever. Cost wasn’t part of the design. It was a surprise on the invoice.
Inadequate risk controls. This is an ownership failure. Nobody drew the line on where the agent decides for itself and where it defers. Nobody said who is accountable when it gets something wrong. So the first real mistake in production becomes the reason to shut the whole thing down.
The root cause - every one of them is a call that got made poorly, or never got made at all.
The thing underneath all of it
None of these projects died in the execution stage. They died in the decision stage.
The teams that got cancelled didn’t lack a better model. They lacked someone stopping to ask the boring questions before the work started. Is this the real problem? How big should this actually be? Where would a mistake break trust? Who owns the outcome?
When execution was expensive, you could get away with skipping those questions. The cost itself forced some discipline. You couldn’t afford to build the wrong thing, so you thought a little first.
That check is gone now. Execution is cheap. You can stand up an agent in a weekend. So nothing stops you from building the wrong thing fast - except your own judgment.
That’s the issue. When execution is cheap, the project doesn’t fail in the code. It fails in the call you made before the code was written.
Four questions that keep you out of the 40%
You don’t need a better model to beat this number. You need to ask four things before you greenlight anything agentic.
What real problem does this solve, and how do we know it’s real? Not “where could we use AI,” but “what’s actually broken, and would fixing it matter.”
What does a good outcome cost, and where does the agent stop? Decide the budget per result and the limit up front, not after the bill arrives.
Where would an error break trust, and what stays human? Some mistakes are cheap. Some end the relationship. Keep the second kind under human control.
Who owns it when it’s wrong? “The agent did it” is not an answer. Name the accountable person before you ship.
None of these are technical questions. All of them decide whether you land in the 60% that ships or the 40% that gets cut.
So should you be worried?
Everyone is reading this number as bad news for AI - I like to invert it.
40% cancelled means 60% ship. And the thing that separates those two groups won’t be budget, and it won’t be model access. Everyone has the same tools now. The difference will be judgment. Which problems they chose. How big they aimed. What they trusted the machine to do, and what they kept for themselves.
The cancellation wave isn’t a warning about how far AI still has to go. It’s the clearest proof yet that the human call is still the whole game.
That’s the skill worth building. The models will keep getting better on their own. Your judgment won’t, unless you work on it.
P.S. Those four questions are the spine of a one-day workshop I run, the AI Strategy Playbook for Product and Engineering Leaders. If you would rather practice them on a real problem and learn by doing it, the next cohort is August 8 and has only 20 seats.
Grab your seat today! Till Friday, July 31, take $350 off with code 350OFF.