An agent that can only read from a script and one that can place orders on its own shouldn’t be governed by the same rules

In a press release dated May 26, 2026, Gartner grouped enterprise approaches to managing AI agents into two extremes: lock them down or trust them completely. The real trouble lies elsewhere: applying the same baseline to roles with vastly different scopes of action. The ones slowed down or bypassed are precisely those that can actually get work done.

The basis for classification should be whether, after this role signs off, money will be paid out, data will leave the organization, or a customer will receive a message. Gartner’s four levels range from observation only, to recommendations only, to requiring approval before acting, to acting autonomously. As the level rises, the focus of controls shifts accordingly: the first two levels are controlled at the input stage; the third at the approval gate; and the fourth through runtime guardrails, circuit breakers, rollbacks, and a named person accountable. Turning the levels into approval steps is the easiest approach—and the one most likely to deteriorate into blindly clicking “Approve.”

The same material also includes a prediction: by 2027, 40% of enterprises will downgrade or shut down autonomous agents because of governance gaps discovered only after an incident. This article is from our company’s official WeChat account. Investing involves risks.