Gartner: AI agents need guardrails and named owners, not just governance policies

By The National CIO Review · October 20, 2026.

Gartner senior analyst Danielle Casey warned at a company event that businesses are deploying AI agents whose behavior they cannot always predict or control. She compared it to driving on unfamiliar roads without guardrails. According to The National CIO Review's report, her thesis is that governance policies are not enough: clear owners, multi-layered controls and continuous evaluation are needed.

According to the outlet's summary, Gartner stresses that agents (AI systems that act with a degree of autonomy to complete tasks) differ from traditional software in three ways: they operate with varying degrees of autonomy, they open up new security and privacy risks, and they can become increasingly expensive to scale.

The recommendations, as reported by the source:

• Separate governance (policies, decision rights, accountability, oversight) from technology management (choosing tools, integrating them, implementing them). And govern what is used today, not hypothetical scenarios. • Guardrails in four layers: technical, administrative, strategic (with leadership involvement) and external interaction. Casey highlighted the compounded risk in multi-agent workflows, where information passed from one agent to another can create unexpected privacy or security exposures. • For each agent: defined objectives, measurable success criteria, continuous evaluation and a "code of conduct" analogous to employees', with escalation procedures or a "kill switch" in case of failure. • To counter agent sprawl, trial periods after which agents that don't add value are retired. • A formal lifecycle (development, deployment, monitoring, retirement), with activity logs and documented evidence.

The proposed timeline is immediate for clarifying who decides, drafting the code of conduct and defining escalation; 90 days for implementing layered guardrails, trial periods and the lifecycle; and one year for strengthening oversight and tying it to business outcomes. The stated goal is "agentic velocity": innovating fast without losing control.

Our take. The most valuable part of the approach isn't the list, but the distinction between governance and guardrails. A written policy doesn't stop an agent that goes off script; a technical limit does. This ties in with something we have argued before: approving permissions with a click is fatigue, not control, and real security comes from design, with least privilege by default and limits the agent cannot bypass.

We also think the idea of trial periods is on point. It turns the agent into something like a new employee: it is measured, given a goal and, if it doesn't perform, retired. It is a pragmatic answer to "agent sprawl" and to cost, the other silent brake on this technology.

Caveats. This is a journalistic summary of a talk, not a report with data: there are no incident figures or evidence that these measures work. In general, consultancy frameworks tend to be sensible but generic, and their value depends on how much they translate into concrete controls. And the multi-agent risk, which Casey only touches on, is probably the least resolved part.

In the short term, the transition toward companies populated by agents will be messy: errors, data leaks, soaring costs. But that break-in period is precisely what builds the trust needed to delegate increasingly complex tasks. If the road to abundance runs through agents working for us, it's worth having them start out with guardrails.