Most forecasts about agentic AI projects describe what they will do. This one describes what will be written off.

Gartner forecasts that more than 40% of agentic AI projects will be scrapped by the end of 2027. The reasons given are not technical: escalating costs, unclear business value, and inadequate risk controls.

That is a prediction about organisations, not about models. And it is the number worth putting beside every adoption statistic currently in circulation.

Agent washing

Gartner also flags a market problem it calls agent washing — existing products relabelled as agentic without the underlying capability changing. Chatbots become agents. Robotic process automation becomes agentic automation. The brochure is rewritten; the software is not.

Its assessment of the vendor field is blunt. Of the thousands of vendors claiming agentic AI, only around 130 are judged to be real. Gartner’s analysts have described most current projects as early-stage experiments or proofs of concept driven largely by hype, and have said plainly that today’s models lack the maturity and the agency to pursue complex business goals autonomously.

For a buyer, that converts a vague unease into a specific question to ask in the room: what does this product decide on its own, under what constraints, and what happens when it is wrong? A vendor with a real agent answers in operational terms. A vendor with a relabelled workflow answers with adjectives.

The forecast is not bearish

It would be easy to read a 40% cancellation rate as scepticism about the technology. Gartner’s other numbers say otherwise. The same body of analysis expects at least 15% of day-to-day work decisions to be made autonomously by 2028, up from effectively none in 2024.

Both things are forecast at once: a large share of current projects failing, and autonomous decision-making becoming ordinary within the same window. These are not contradictory. They describe a field where the destination is reasonably clear and the route is littered with projects started for the wrong reason, scoped without a business case, or deployed without controls anyone could name.

Cancellation at that rate is what a market looks like while it learns which problems are worth an agent — not evidence that none are.

What the cancelled projects will have in common

The three reasons Gartner gives have a shape, and each is diagnosable before a project starts.

Escalating costs point at consumption-based pricing meeting open-ended usage. An agent that retries, reflects and calls tools consumes tokens in a pattern nobody modelled, and the invoice arrives monthly whether or not the value did.

Unclear business value points at projects that began with a technology and went looking for a problem. The tell is a pilot that cannot name the metric it would move, or that names one nobody on the business side had been complaining about.

Inadequate risk controls point at the governance question nobody wants to own: when an agent acts, who is accountable for the decision?

The question that separates the two groups

That last one is the hardest and the most avoidable. A decision without an identifiable author is an audit problem in any regulated process, whether the author would have been a person or an algorithm. Organisations that settle it before deployment — recording who or what decided, on what basis, and what the alternative value would have been when someone overrides the system — are not the ones in the 40%.

It is worth noticing how little that has to do with model capability. The projects that survive will not be the ones that picked the best model. They will be the ones that could answer, on the day a regulator or a customer asked, why the system did what it did.

Source: Gartner press release, 25 June 2025.