There is a particular kind of meeting happening in enterprises right now. An agent pilot has worked. It demonstrated well. Someone asks what it would take to run it for real, against production systems, for every transaction rather than the twenty in the demo. The room goes quiet.
Gartner has put a number on how those meetings end: it expects more than 40% of agentic AI projects to be cancelled by the end of 2027. The reasons it gives are worth reading carefully, because none of them is about model quality: escalating costs, unclear business value, and inadequate risk controls.
That is an unusually actionable forecast. Every one of those three is a scoping decision, which means it is made before the build, by people who can make it differently.
The market is smaller than the marketing
Two figures from the same Gartner work are worth holding together.
The first is that in a January 2025 poll of 3,412 webinar attendees, 19% of organisations reported significant investment in agentic AI. The second is Gartner's estimate that of the thousands of vendors describing themselves as agentic, only around 130 genuinely are. Gartner calls the gap agent washing — existing products relabelled.
For a buyer this has a practical consequence. The question to ask a vendor is not whether their product uses agents. It is what the agent is permitted to do without a human, what happens when it is wrong, and what it costs per thousand actions at production volume. Those three questions separate the 130 from the rest faster than any feature list.
Gartner is not forecasting a collapse, incidentally. The same body of work expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by 2028. Both things are true: most current projects will be cancelled, and the category will still matter. That combination describes a technology in the awkward middle, where the determining factor is scoping discipline rather than ambition.
Where the cost actually escalates
Pilot economics mislead in a specific, repeatable way.
| Cost | In the pilot | In production |
|---|---|---|
| Inference | Dozens of runs, absorbed in a trial budget | Every transaction, every day, with retries |
| Retries and loops | Rare, and a human notices | Common, and an agent that loops costs money while failing |
| Integration | Mocked, or one read-only API | Write access, error handling, idempotency, rollback |
| Evaluation | Someone eyeballs the output | A standing test set, maintained as the business changes |
| Human review | The builder, informally | A named queue, staffed, with a service level |
The last row is where most business cases quietly break. If the agent handles 70% of cases and a person handles the rest, you have not removed a role — you have changed what that role does and added a supervision duty on top. That can still be an excellent outcome. It is a very different outcome from the one in the original slide, and finding out at month nine is expensive. Trade coverage of the same Gartner work made the point squarely: the forecast makes humans more indispensable, not less, because the supervision layer is what the economics actually rest on.
What the ones that survive have in common
The surviving projects we see are not the most technically ambitious. They share a shape.
- A bounded decision, not a bounded department. "Classify and route incoming supplier invoices that fail three-way match" survives. "An AP agent" does not. The first has a measurable error rate; the second has a remit.
- A reversible action. Agents should start where a wrong answer is caught cheaply — drafting, classifying, preparing, flagging — rather than where it posts to a ledger or sends to a customer.
- A baseline measured before the build. If nobody recorded how long the task took and how often humans got it wrong, there is no way to prove the agent helped, and the project dies at the first budget review regardless of how well it works.
- An owner in the business, not in IT. Agent behaviour drifts as processes change. Someone whose job depends on the outcome has to own the evaluation set.
- A documented stop condition. What the agent does when it is uncertain, and who it hands to. Projects without this answer tend to discover it during an incident.
A sequence that tends to work
Start with the decision, not the technology. Write down one decision your team makes many times a day, the information it needs, and how often it currently goes wrong. If you cannot complete that sentence, the problem is not ready for an agent.
Then instrument it. Spend a fortnight capturing the baseline — volume, handling time, error rate, rework. This is unglamorous and it is the thing that makes the business case defensible later.
Then build the narrow version, with the human in the loop by default and the agent's authority expanded only as the evaluation set says it has earned it. Govern the authority explicitly: what it may do alone, what needs approval, what it may never do. Write that down before the first line of code, because afterwards it becomes a negotiation.
Key takeaways
- Gartner's 40% cancellation forecast is driven by cost, value and controls — all scoping decisions, all made up front.
- Only around 130 of thousands of self-described agentic vendors are judged genuinely agentic; ask what the agent may do unsupervised, and what it costs at volume.
- Production cost is dominated by retries, integration and human review, none of which are visible in a pilot.
- Scope one repeated decision with a reversible action, measure the baseline first, and expand authority only against evidence.
- Expect a supervision role to remain. Budget for it rather than discovering it.