Gartner Predicts 40% AI Project Cancellations: Here's the Fix

In June 2025, Gartner made a prediction that should have stopped every AI team in its tracks: over 40% of agentic AI projects will be cancelled by the end of 2027. The causes cited were escalating costs, unclear business value, and inadequate risk controls.

That prediction landed in boardroom slide decks, got a few worried nods, and was promptly filed away while teams continued deploying agents without the infrastructure to govern them.

A year later, the prediction is aging well — and not in a good way.

The diagnosis is precise

Gartner's analysts were specific about what's going wrong. According to Senior Director Analyst Anushree Verma, most agentic AI projects are "early stage experiments or proofs of concept that are mostly driven by hype and are often misapplied." Organizations are "blind to the real cost and complexity of deploying AI agents at scale."

More damning: Gartner estimates that only about 130 of the thousands of agentic AI vendors are real. The rest are engaged in "agent washing" — rebranding chatbots, RPA bots, and AI assistants as agentic AI without any meaningful autonomous capability. Companies buying these solutions are not just failing to get value; they're building organizational processes around tools that can't deliver on their promises.

But here's what I find most significant. Of the three reasons Gartner cites for cancellation — escalating costs, unclear business value, inadequate risk controls — two of them are governance problems in disguise.

Costs escalate when you can't measure what agents do

The "escalating costs" problem isn't primarily about compute or API spend, though those are real. It's about the hidden costs of ungoverned agent activity.

When an AI agent runs a workflow, it makes decisions at each step. Some of those decisions trigger API calls. Some trigger downstream agents. Some trigger retries when the output isn't what a human expected — because nobody defined what the output should be in the first place.

Without visibility into agent decision-making, you can't optimize. You can't identify which decisions lead to unnecessary compute. You can't spot agents that are retrying tasks in loops. You can't distinguish between productive agent activity and waste.

Consider a parallel from human organizations. If you had employees making thousands of decisions daily with no record of what they decided or why, and your costs were spiraling, the first thing you'd do is create visibility. You'd want to see the decisions. You'd want to understand the patterns. You'd want to set boundaries.

We do none of this for AI agents. We deploy them, pay the invoices, and wonder why costs keep climbing.

Business value is unclear when you can't audit outcomes

The "unclear business value" problem is even more directly a governance failure.

How do you measure the value of an AI agent? You look at its outputs and assess whether they improved outcomes. But if you can't trace which decisions led to which outcomes — if the agent's reasoning is a black box — then you can't attribute value. You can't distinguish between an agent that's genuinely helping and one that's producing plausible-looking outputs that humans rubber-stamp without verifying.

Gartner's own data tells the story from the other side. They predict that 15% of day-to-day work decisions will be made autonomously by agentic AI by 2028, up from 0% in 2024. And 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.

That's an enormous volume of autonomous decisions entering the enterprise. If even a fraction of those decisions are invisible and unauditable, the aggregate risk is staggering. And when the risk becomes visible — through an incident, a compliance failure, or simply a board member asking "what exactly are our AI agents doing?" — the response is predictable: cancel the project.

Not because the technology failed. Because the governance infrastructure was never built.

Inadequate risk controls: the root cause

The third factor — inadequate risk controls — is the honest one. It's not a symptom of something else. It's the root cause that makes the other two inevitable.

Risk controls for AI agents require three capabilities that most organizations lack:

Visibility. You need to see what every agent is doing, in real time and historically. Not just the outputs — the decisions. What did the agent consider? What did it choose? What did it reject?

Auditability. You need structured records of agent decisions that can be queried, reviewed, and presented to regulators, auditors, or leadership. "The AI decided" is not an audit trail.

Governability. You need the ability to define policies — "this agent requires human approval for actions above this threshold," "no agent may access production data outside business hours," "all financial decisions above $10,000 require a decision gate" — and enforce them at runtime, not after the fact.

These three capabilities — visible, auditable, governable — are what I've been calling Trust Infrastructure. They're the infrastructure layer that sits beneath your agents and makes them manageable, the same way Kubernetes sits beneath your containers and makes them orchestratable.

The real fix

Here's what Gartner's prescription usually boils down to: be more strategic about where you deploy agents. Start with high-value, low-risk use cases. Build internal expertise. That's reasonable advice, but it's incomplete. It tells you to be careful. It doesn't give you the infrastructure to be careful.

Being strategic about agent deployment requires knowing what your agents are doing. That requires visibility. Measuring business value requires tracing agent decisions to outcomes. That requires auditability. Managing risk requires defining and enforcing boundaries. That requires governability.

Without this infrastructure, "be more strategic" is just "be more lucky."

The 40% cancellation rate isn't a prediction about technology failing. It's a prediction about organizations deploying powerful tools without the infrastructure to manage them. The technology works. The agents are capable. What's missing is the layer between the agent and the organization — the layer that makes AI agent activity visible, auditable, and governable.

That layer is what needs to be built. The organizations that build it — or adopt it — will be the 60% that survive. The rest will keep deploying agents in the dark, keep losing visibility into what those agents decide, and keep cancelling projects when the consequences become impossible to ignore.

The Gartner prediction is a warning. But warnings without infrastructure are just anxiety. It's time to build the fix.

---

Ryohoshi is building deAria — open-source Trust Infrastructure for AI agents. This article expands on the data discussed in Agent Dark Matter: The Invisible Crisis in Your AI Stack.