The counterweight to every agentic AI headline arrived this year. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. A March 2026 survey found 78% of enterprises running agent pilots and fewer than 15% in production; other counts put the pilots that never ship at 89% — while the survivors report 171% ROI.
Read the failure modes, not the percentage
Three causes come up repeatedly, and none of them is "the models aren't good enough": unclear value, data the agent can't trust, and no owner to hand the decision back to. Gartner separately expects organizations to abandon 60% of AI projects unsupported by AI-ready data — which is a polite way of saying the agent was pointed at a system of record nobody had kept current.
That failure needs an enterprise to happen. It needs a pilot with no P&L attached, a committee, and a data estate too large for anyone to fix in a quarter. An agency has none of those, which is the actual advantage here — not a smaller budget, a shorter distance between the agent and the person who can tell whether it helped.
What the survivors did
The 11% that shipped picked one loop that runs every day and measured it before and after. Not "AI transformation" — chasing overdue tasks, answering status questions, checking new asks against scope. Something with an obvious before, a visible cost, and one person who owns the outcome. Which loop to pick first matters more than which model you use.
The market is converging on the same conclusion. Salesforce shipped seven named agents in September, each scoped to one business function — an admission that "build your own agent from a blank canvas" is where pilots go to die.
The honest test
Before you buy or build: name the loop, name the week it should show up in, and name who gets told when the agent is wrong. If the answer to the third one is "nobody, it's autonomous", you've found the risk control the cancelled 40% were missing.
The ROI numbers are real — they're just collected from the projects that stayed small enough to finish.
← All posts