Gartner's headline number -- more than 40% of current agentic AI projects forecast to be cancelled by 2027 -- reads like a bearish call on the technology if you stop at the headline. The stated drivers say something narrower and more useful: unclear ROI and weak risk controls, not "the models don't work" or "the use case wasn't real."
Unclear ROI, on the evidence from this year's adoption surveys, usually traces back to measuring the wrong thing. A pilot gets funded on a demo; a production system has to justify its cost against a baseline the organization can already measure -- support tickets resolved, hours saved, error rate versus the human process it replaces. Plenty of agent pilots were never instrumented to produce that number, which makes a go/no-go decision at renewal time a political question instead of a data one, and political questions are exactly the ones that get resolved by cancellation when budgets tighten.
Weak risk controls is the other half, and it's the same gap this digest has covered from two other angles this week: the adoption-scaling gap (agents that work in a pilot but weren't built to handle the 20% of cases a pilot never exercises) and the documented AI-assisted attack activity (capability without governance is a liability, not just a shortfall). A project that can't answer "what's the worst thing this agent can do, and what stops it" doesn't survive a security review at scale, however well it performed in the demo.
Read together with the adoption numbers, the forecast isn't really about whether agentic AI works. Three-quarters of enterprises using agents in some form is not a technology that's failing. It's a forecast about which projects -- the ones scoped, measured, and governed like a six-week pilot instead of a system meant to run unattended -- don't survive contact with a second budget cycle.
A cancellation forecast driven by "unclear ROI and weak risk controls" is a governance and measurement problem, not a capability problem -- the fix is instrumenting an agent project against a real baseline and answering the blast-radius question before scaling it, not waiting for a renewal decision to force the issue.