Why agentic AI projects get canceled

What the 40% number is actually measuring

It is a forecast, not a body count. Gartner published it on 25 June 2025, and the projects it covers are mostly early-stage. Gartner's supporting poll put 19% of organisations at significant investment in agentic AI, with 31% undecided or still waiting. So the prediction is largely about abandonment before production, not about mature deployments falling over. That distinction matters if someone is using the number against your programme. A pilot that stops because it was never scoped to prove anything is a different event from an agent pulled out of production after it did something nobody sanctioned.

Agent washing inflates the count

Gartner also estimates that of the thousands of vendors claiming agentic AI, only around 130 offer genuine agentic capability. The rest is rebranded chatbots, robotic process automation and assistants wearing a new label. That has a direct effect on the forecast: a share of the projects predicted to be cancelled were never agentic to begin with, so they were going to disappoint whatever the team did. Before you accept the number as a verdict on your own work, check which category your system is in. An assistant that drafts text and an agent that books, refunds or escalates carry different risk and deserve different scrutiny.

The more recent forecast is the one worth reading

The May 2026 Gartner release is fresher and lands closer to the operational reality. Its prediction is that 40% of enterprises will demote or decommission autonomous agents by 2027, and its diagnosis is blunt: enterprises treat agent governance as binary, either locked down or fully trusted, and Gartner names that as the root cause of failure. Locked down means the agent does nothing useful and the business case dies quietly. Fully trusted means the agent eventually does something expensive and gets pulled. Gartner's recommended fix is proportional governance, classifying agents across distinct autonomy levels with a trust boundary and a governance requirement for each. That is a design instruction, not a warning.

The three patterns behind a cancellation

Underneath both forecasts, the same failures repeat. Each one is visible early if anyone is looking for it.

• Cost without measured value. The programme spends on model calls and integration work while its business outcome is described in adjectives rather than a number anyone tracks.
• Scope nobody bounded. The agent accumulates tools and permissions because each addition looked reasonable on its own, and nobody can now say what it may not do.
• No accountable owner after the fact. When an action goes wrong, the investigation has to reconstruct what happened from application logs that were never designed to answer the question.
• Governance applied as a switch. The agent is either blocked from anything interesting or trusted with everything, because there was no way to grant authority in proportion to the risk of a specific action.

A short test for your own programme

Run these five questions past the team that owns the agent. Anything you cannot answer in a sentence is where the cancellation risk sits.

• Name the business outcome and the number that moves. If nobody owns that number, the programme has no defence at budget time.
• List the actions the agent may complete alone, the actions that need a person, and the actions it should never have. A list that has never been written is a list nobody agreed.
• Ask who approved the last tool the agent gained, and when. If the answer is a deployment, not a decision, scope is drifting.
• Pick one action from last week and reconstruct it end to end: who it acted for, what it proposed, what was allowed, what it changed. Time how long that takes.
• Ask what happens when a check cannot reach a verdict. If the honest answer is that the action proceeds, the failure mode is already in production.

How Difinity changes the odds on the second forecast

The binary-governance failure Gartner names is a structural problem, so it needs a structural answer rather than a policy document asking teams to remember to scope things. In Difinity, an agent's authority is the intersection of three things: what its version binds, what the caller is entitled to, and what the use case permits. It is never a union, so an agent cannot accumulate more reach than a deliberate decision gave it. An agent version can bind at most 32 tools, and under each bound tool the author writes, in their own words, when the agent may use it. Anything that changes something elsewhere is judged before it runs whatever that written policy says, and where the judge cannot settle it, an effectful tool escalates to a person rather than defaulting to permit. Every run is written to an append-only run trail as it happens, so the fourth question in the test above takes minutes rather than a week.

Where these forecasts could be wrong

Both numbers are predictions from one research firm, not measurements, and both are about the end of 2027. Gartner has an obvious commercial interest in enterprises buying advice about agent governance, which does not make the analysis wrong but is worth holding in mind. The agent-washing effect cuts both ways: it inflates the projects counted as failures and it also hides the genuine deployments inside a noisy denominator. Model capability is moving fast enough that the 2025 judgement about maturity may read differently by 2027. What is unlikely to change is the mechanism: a programme with no measured outcome and no bounded authority stays fragile whatever the underlying models do.