The prediction, and what it actually says
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. The stated causes reward a close reading: escalating costs, unclear business value, inadequate risk controls. Not one of them is a limitation of the models.
Anushree Verma, the Gartner analyst behind the forecast, is equally direct about the supply side. Most agentic propositions carry no measurable return, and much of the market is practicing what Gartner calls agent washing: relabelling chatbots, assistants and existing automation as agents. Gartner counts roughly 130 genuine agentic vendors among the thousands claiming the label.
Pilots pass, production fails
The failure point is consistent. A pilot runs against clean sample data, with a friendly tester and no consequence for a wrong answer. Production has none of those. It has real permissions, real integrations, real money moving, and a colleague who has to explain the mistake to a customer.
Gartner describes this as a gap between demonstrated capability and deployed capability. In our own projects the gap almost always opens in the same three places: data access nobody scoped, failure states nobody designed, and accountability nobody assigned.
What the survivors do differently
The projects that reach production share a short list of habits, and none of them are technical.
- Autonomy is graduated, not granted. The agent starts by drafting for a person, then acts on the low stakes cases, then widens only where its record justifies it.
- Approval gates match the stakes. Sending an internal summary needs no gate. Issuing a refund does. That threshold is a business decision written down before launch, not a setting discovered afterwards.
- Return is checked per phase. A named finance owner signs off at each stage. A project that cannot show its payback at phase two rarely finds it at phase four.
- Every agent has an owner. Not a team, a person, answerable for its quality, its cost and its exceptions.
Choosing the first one
The cheapest way to stay out of the 40% is to pick a first workflow that is boring on purpose. Look for work that is high volume, low variance, already documented, and owned by somebody who feels the pain. Intake classification, quote preparation, order triage and support drafting all qualify. Anything where the correct answer is genuinely contested belongs to your third project, not your first.
Payback data supports the boring choice. Survey figures compiled across 2026 put the median payback for agent deployments at roughly five months, with sales development agents nearer three and finance workflows nearer nine. Those are ordinary software economics, which is exactly the point. An agent that needs an extraordinary story to justify itself is usually the one that gets cancelled.
The uncomfortable summary
When a project fails on cost, clarity or control, the fix was never a better model. It was a smaller scope, a named owner and an honest number. That is unglamorous work, and it is most of what separates the 60% from the 40%.
At Clodron this is why an automation engagement starts with workflow mapping and an evaluation set rather than a demo. The demo was always the easy part.