Four in ten AI agent projects will be cancelled. Here's how to avoid being one.

What happened

In June 2025, Gartner predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027. It named three causes: costs that escalate, business value that stays unclear, and risk controls that turn out to be inadequate.

It also flagged what it calls agent washing. Some vendors rebrand chatbots, RPA tools and assistants as "agents" without adding real agentic capability. Gartner estimated that only about 130 of the thousands of vendors claiming to sell agentic AI are the real thing.

Read the list of causes again. Cost, value, risk. Not one of them says the model was too weak.

So what for your business

Most agent projects start with a demo. The demo works, someone says "let's roll it out", and nobody has written down what success looks like, what a task is allowed to cost, or what the agent must never do without a person signing off. Six months later the invoice is bigger than the savings and the risk team has questions nobody can answer.

The projects that survive tend to look boring at the start. They pick one workflow, measure how it runs today, and compare the agent against that number. They decide in advance where a person has to approve the result.

Gartner's own advice points the same way. Use agents where there is a clear return. Rebuild the workflow rather than bolting an agent onto an old one. Match the tool to the job: agents for decisions, plain automation for routine steps, assistants for lookups.

What to do this month

1. Pick one workflow. Choose something frequent, painful and measurable. Write down today's cost, time and error rate before you touch anything.
2. Set a cost ceiling per task. If a single run can cost more than the work is worth, you will find out on the invoice. Find out in the design instead.
3. Write the "never alone" list. Payments, deletions, messages to customers, anything legal. These get a human approval step.
4. Ask vendors what the agent actually decides. If the answer is "it summarizes", you are buying an assistant. That's fine, but price it like one.
5. Review at 30 days. Compare against the baseline. Keep it, change it or stop it. Stopping a pilot early is a good outcome, not a failure.