The forecast, and what it actually measures
In mid-2025 Gartner projected that over 40 percent of agentic AI projects would be canceled by the end of 2027, pointing to rising costs, unclear business value, and inadequate risk controls. It is worth being precise about what a number like that measures. It is not a verdict on whether agents work. It is a count of programs that could not justify their spend or defend their risk to the people holding the budget. Cancellation is a governance event before it is a technology one: someone in a review could not answer what is this returning and what could it break, so the project died.
Three failure patterns behind the number
The reported causes map cleanly onto what stalls these programs in practice. Cost with no value line: teams pay for tokens and infrastructure while nobody can point to a business metric that moved, so the spend looks like a hobby at renewal time. Scope without limits: an agent is given the ability to act on real systems before anyone defined what it may not touch, and the first bad action ends the trust. No accountability trail: when a review board asks who approved this action and why, the honest answer is nobody knows, which is an automatic no in any regulated setting. None of these are model problems. All three are decisions a delivery team either made up front or skipped.
The washing problem inflates the count
Gartner also flagged agent washing, vendors relabeling chatbots, assistants, and rules engines as agentic. That matters for reading the cancellation figure honestly. Some of the 40 percent were never agentic programs to begin with; they were ordinary automation dressed in new language, and they got canceled when the label did not survive contact with a real workload. Strip those out and the signal that remains is sharper: genuine agentic projects fail when value and control were treated as things to add later rather than conditions to start.
Where the survivors are different
The programs that make it to 2027 will not be the ones with the most capable agents. They will be the ones that scoped narrowly, measured a real number, and decided the boundary before the build. This lines up with the wider evidence on enterprise AI value. MIT's 2025 NANDA work found roughly 95 percent of generative AI pilots produced no measurable profit-and-loss impact, and the gap was consistently execution and integration, not model quality. Agentic AI raises the stakes of that same gap, because an agent that acts on production systems turns a weak value case into a weak value case with a blast radius.
The optimistic reading, and where it falls short
There is a fair counterargument: early cancellation is healthy pruning. Portfolios are supposed to kill weak bets, and a 40 percent cull could mean enterprises are getting more disciplined, not that agents are failing. That reading has some truth, and it is better than the panic version. It falls short on timing. Most of these cancellations arrive after real money and months of engineering are already spent, because the value and control questions were deferred until a late review instead of asked at the start. Pruning at the funding gate is discipline. Pruning after a six-month build is waste dressed up as discipline. The projects worth defending are the ones where someone did the hard scoping before the budget cleared, so the weak bets never got funded in the first place.
How to not be in the 40 percent
Treat every agent as a use case that has to earn production, not a capability to show off. Give it a baseline and a target number so value is arguable at budget time. Define the tools and actions it may reach, and the ones that always stop for a human, so the risk is bounded before the first run. Keep a trail detailed enough to answer a regulator or a customer after the fact. For teams in financial services, healthcare, and government, that last point is the difference between a pilot and something an auditor will let you keep. Operators who have taken AI to production at 300,000-organization scale sequence agents this way on purpose: value you can prove, inside controls you decided up front. That is the opposite of the profile Gartner is describing.
The takeaway for 2026 planning
If you are budgeting agentic work for next year, the forecast is useful less as a prediction and more as a checklist of ways to get canceled. Every project that dies for cost, value, or control died for a reason you could have caught in a one-week readiness pass. The cheapest agentic project is the one you decline before you fund it, and the second cheapest is the narrow, governed one that ships and gives you a number to defend. Run every candidate through that checklist before it clears funding, and the forecast stops being about you. The teams that will look prescient in 2027 are not the ones who bet biggest on agents. They are the ones who said no early, often, and on paper, and put their money behind the few that could answer the value and control questions on day one.