Gartner says 40% of agentic AI projects will be cancelled. Here is which 40%.

What exactly did Gartner predict?

On 25 June 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing three drivers: escalating costs, unclear business value and inadequate risk controls (Gartner, June 2025). The prediction is grounded in a poll of over 3,400 organisations actively investing in the technology, so this is not a survey of sceptics. It is a forecast about the people who already bought in.

Read the word "cancelled" carefully. This is not a failure rate in the sense of projects that shipped and underperformed. It is a cancellation rate: projects abandoned mid-flight, before they ever reached the workflow they were meant to change. The money is spent; the return is zero. That is a worse outcome than not starting, and it is the modal outcome Gartner expects for two out of every five agentic initiatives.

We bring this number to the table for the same reason we bring MIT's. Not to frighten anyone off agentic AI (we build agentic systems), but because the three causes are avoidable, and naming them is the first thing a serious vendor should do.

Over 40% of agentic AI projects will be cancelled by end of 2027: escalating costs, unclear business value, inadequate risk controls. >Gartner press release (2025)

Every one of the three causes is a pre-contract decision. Cost overruns come from unbounded scope; unclear value comes from an unnamed workflow; absent risk controls come from skipping the evaluation harness. None of the three is a model problem.

The three named causes, in the order we see them

Unclear business value comes first, because it causes the other two. A project that cannot name the workflow it is changing cannot bound its cost or define its risk controls, because it has no fixed target to bound them against. Scope drifts, the demo expands, and the thing that was going to save time on invoice exceptions becomes a platform. We refuse to start an engagement we cannot name the workflow for, and this is why.

Escalating cost is second. Agentic systems fail expensively in a specific way: the happy path is cheap and the exception path is not. Every retry, every tool call, every fallback to a human costs tokens and latency, and the exceptions are where the volume hides. A project priced on the demo's cost per run discovers the real number in month three.

Inadequate risk controls is third, and the most quietly fatal. Without an evaluation harness, you cannot tell a regression from variance, so every model update is a gamble and every incident is a mystery. The project does not blow up. It loses the confidence of its sponsor and gets quietly defunded, which counts as a cancellation. Designing where the human stays in the loop is the same discipline seen from the other side.

Agent washing: how to test a vendor in ten minutes

Gartner estimates only about 130 of the thousands of self-described agentic AI vendors are genuine; the rest it calls "agent washing": existing assistants, RPA and chatbots rebranded as agents without substantial agentic capability (Gartner via MarTech, 2025). If roughly 130 of thousands are real, the base rate says the vendor in front of you probably is not.

Three questions separate the two in one meeting. Ask them to name the workflow before scoping: a washer will describe a capability, a builder will describe your queue. Ask what they would refuse to automate: a washer automates everything, a builder has a refusal list. Ask to see the evaluation harness they leave behind: a washer ships a demo, a builder ships a test suite. We wrote the longer version of this as ten questions to ask any AI vendor, including the ones we struggle to answer well ourselves.

Understanding what "agentic" actually means (orchestration, tool use, state, stopping conditions) is the cheapest defence against being sold a chatbot. That is what the difference between agentic AI and prompting comes down to.

Which projects are in the surviving 60%?

The survivors share the inverse of the three causes. They have one named workflow with a measured baseline, so value is not in doubt. They priced the exception path, not the demo, so cost does not surprise them. And they stood up risk controls (an evaluation harness, human checkpoints where errors are expensive, a defined escalation path) before go-live rather than after the first incident.

Notice that none of this is about picking the right model or the cleverest architecture. The survivors are not more technically sophisticated. They made three governance decisions early, and the decisions cost almost nothing at the start and everything if deferred.

This is why we insist on naming the workflow before the contract, and why our first paid step is a diagnostic rather than a build. An engagement scenario for EDI exception triage shows the shape: the exception queue is the automatable workflow precisely because it is bounded, measurable and already a source of pain: the opposite of an open-ended platform.

What does this mean for a GCC transformation owner?

Treat the 40% as a checklist, not a warning. Before you approve an agentic project, confirm three things in writing: the single workflow it changes, the fully-loaded cost per run including the exception path, and the risk controls that will exist on day one. If any of the three is missing, you are looking at a probable member of the cancelled 40%, and the time to find out is now, not in 2027.

And hold your vendor to the same test you would apply to any capital decision. The honest ones will welcome it, because the three questions above are the ones that disqualify their competitors.

Disclosure: Chokmah is a new practice with no completed client engagements. This page argues from Gartner's published research and from method. When we build an agentic system, the evaluation harness and the refusal list are part of the deliverable, not an upsell: that is what a governance and CoE retainer maintains after the sprint ends.

Keep your project out of the cancelled 40%

Name the workflow, price the exception path, stand up the risk controls before go-live. A workflow sprint ships all three, and the client owns the code.

Scope a Workflow Sprint · Book an Adoption Diagnostic first

Frequently asked questions

What is agentic AI?

Agentic AI describes systems that pursue a goal across multiple steps (planning, calling tools, reading and writing data, retrying on failure, and deciding when to stop), rather than answering a single prompt. The orchestration layer that coordinates those steps reliably is where agentic projects usually succeed or fail.

What is agent washing?

Agent washing is Gartner's term for rebranding existing products (AI assistants, robotic process automation, chatbots) as 'agents' without substantial agentic capability. Gartner estimates only about 130 of the thousands of self-described agentic AI vendors are genuine, so the base rate favours scepticism.

Why do agentic projects cost more than expected?

Because the exception path is expensive and the demo only shows the happy path. Every retry, tool call and human fallback consumes tokens and latency, and exceptions are where the volume hides. A project priced on the demo's cost per run discovers the real cost once real traffic includes the messy cases.

Is 40% a failure rate or a cancellation rate?

A cancellation rate. Gartner predicts over 40% of agentic AI projects will be abandoned before completion by end of 2027: money spent, no workflow changed. That is distinct from projects that ship and underperform, and arguably worse, because the spend produced nothing at all.

How do you build risk controls into an agent project?

Start with an evaluation harness that scores every run against versioned test cases, so you can tell a regression from variance. Add human-in-the-loop checkpoints where errors are expensive and hard to reverse, and a defined escalation path. Stand all of this up before go-live, not after the first incident.