54% of Organisations Either Don’t Limit AI-Agent Access or Just Reuse Human Permissions, Gartner Says

Gartner says 54% of organisations in its survey either had no defined approach to limiting AI-agent access or relied on predefined human access. The figure comes from Gartner’s announcement on CISO priorities through 2026, and it points to a governance question that’s more specific than whether companies are adopting AI: what authority does an agent get once it’s allowed to act?

The number is a combined one. Gartner doesn’t publicly separate organisations with no defined approach from those using a person’s existing access as the basis for an agent’s permissions. That distinction matters. One group has no model at all. The other may have made a deliberate choice to delegate broadly. The Impact of AI reports the result comes from a Q2 2026 survey of 297 cybersecurity leaders, and notes that the public announcement doesn’t break down the two groups.

It’s a governance number, not an incident count. The finding doesn’t show that 54% of organisations suffered a breach, data loss, or unauthorized agent action. It records whether respondents reported a defined way to constrain an agent’s access. It also says nothing about how well the approaches used by the other respondents work. Access governance covers several questions: which identity an agent uses, what permissions it gets, which systems and data it can reach, and how its authority is monitored or revoked. The statistic addresses only whether an approach exists.

Reusing human access isn’t the same as agent-specific control. An organisation might give an agent the authority of the person who started or owns a task. The public material doesn’t say how Gartner evaluates that model, which permissions are involved, or whether they’re narrowed per action. It also doesn’t define what counted as an AI agent, whether internally built systems, vendor tools, developer assistants, or customer-service automation.

What isn’t published: the questionnaire, respondent geography, sampling method, or margin of error. Gartner’s announcement doesn’t name a preferred identity architecture, permissions model, or vendor, and the public material doesn’t establish that limiting agent access is one of Gartner’s five named CISO actions. Any claim that a particular product closes the gap would go beyond what’s been shown.

Bottom line: The finding gives security leaders a concrete question: is an agent’s authority defined separately from the user’s, or just borrowed from them? As agents get more ability to retrieve information and act, that difference becomes a governance decision, not a technical detail. The survey shows that many respondents haven’t publicly described an agent-specific access model. Whether that’s an absence or a deliberate choice, the data doesn’t say.