Gartner and Agentic AI: Agents as digital delegates
Gartner is continuing its warning on agentic AI.
In June 2025, the analyst firm predicted that more than 40 percent of agentic AI projects could be cancelled by the end of 2027. The reasons were rising costs, unclear business value, weak risk controls, and a lot of agent washing.
In May 2026, the warning became more specific. Gartner is now describing less of a hype problem and more of an operating problem: many companies treat agents either like ordinary software features that are tightly restricted, or like systems that are given very broad trust. Both are a problem.
A copilot with poor data gives poor answers. That is annoying, expensive, and risky in some cases. But usually a human still sits between output and effect. Someone reads, checks, rejects, or corrects.
An agent works in a different category. It can write, forward, enter, prepare, change, or trigger a process step. A wrong assessment can therefore become not just a bad suggestion, but a wrong action in the system.
That is where the risk class changes. If the data context is missing, the agent executes wrong assumptions. If processes are unclear, it runs the wrong process "cleanly". If permissions are too broad, it can see or do more than intended. And if monitoring and responsibility are missing, the error may only become visible after damage has already occurred, possibly after several executions.
Companies that do not want to become part of Gartner's forecast should not treat agents like additional software tools. They are digital delegates. Such authority is not granted broadly. It needs a defined purpose, access limits, action scope, approval thresholds, and revocation. And before assigning it, the company has to check whether the system actually has the data, context, and process logic required to execute the task reliably.
That is the difference between AI chat and an agent: the chat supports work. The agent receives execution rights.