In June 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. It also warned about “agent washing”, the relabelling of existing products as agents.
That is a forecast, not a measurement, but the causes it names are useful because none of them is about model quality. All three can be addressed before you write a line of agent code.
Cost: price the task, not the model
Demos hide the real cost. In production, an agent retries, calls tools and carries long context, and every one of those steps costs money. Before you build, estimate the cost per finished task at expected volume and compare it with what the task costs today. Then set hard limits on tool calls and tokens, with an alert before the budget runs out.
Value: choose a task you can measure
“Use agents in finance” isn’t a project. “Reduce the time from invoice due date to first reminder” is. Pick one process with an owner and a number that exists today. If nobody can state the current number, measuring it is your first deliverable, and it may well show that a plain workflow would do.
Risk: earn each permission
Treat an agent as a new hire. Give it its own account and start it on reading only. Move it to drafting once its reads are reliable, then to acting with approval, and let it act alone only within narrow limits. Log every tool call so that “why did it do that?” always has an answer. Test how it handles instructions hidden in documents and emails, since an agent can’t always tell data from commands.
The fourth cause: the data underneath
Gartner made a related prediction in February 2025: through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. An agent that drafts a purchase order from duplicate products and stale vendor prices just produces mistakes faster. This is why we start with the ERP. A clean system of record is what lets the rest stay simple.
A plan that survives contact with production
- Pick one process with an owner and a baseline number.
- Fix the records it depends on.
- Check whether a workflow or a simple model solves it first.
- Write an evaluation set from real examples, including the awkward ones.
- Launch read-only, then raise permissions on evidence.
- Review cost and quality on a schedule, and be willing to switch it off.
If you want a second opinion on a project in this position, tell us where it stands.
Sources: Gartner, 25 June 2025; Gartner, 26 February 2025.