Gartner predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents they run today, because governance gaps surface only after a production incident (Gartner, May 2026). The problem its analysts name is not weak models. It is that companies treated autonomy as a switch, either locked down or fully trusted, and gave agents scope they had not yet earned. The capability was real. The trust was assumed.
Why It Matters
The instinct is to tie how much an agent runs on its own to how good the underlying model is. A stronger model feels like it deserves a longer leash, the way a proven hire would. It has not earned one. In Anthropic’s study of millions of real agent sessions, full auto-approval rose from about 20% for new users to over 40% for experienced ones, and the time agents worked unsupervised nearly doubled (Anthropic, Measuring Agent Autonomy). What changed was not the model. It was the operator’s track record with it, built by actively checking the work, not by waiting for time to pass. Autonomy grew with demonstrated reliability, not with a release date. The ones Gartner expects to backtrack skipped that step.
The Decision
The question worth asking this week is what actually earns an agent more room in your operation. If the answer is the vendor’s latest release, autonomy is running ahead of evidence. If it is a record you can point to, weeks of clean output that a human still checked, the leash is lengthening for a reason. Both feel like progress. Only one holds up after an incident.
What To Do This Week
- Take one agent that runs with little oversight and ask what earned that freedom. If the answer is that the model improved, you found your risk.
- For your highest-stakes agent, separate what it can do from what it is allowed to touch: drafting an email is not sending it, querying a database is not changing it. Then enforce that line in the system itself, not on paper, so the agent cannot act past it.
- Set the bar for a clean track record before you widen any agent’s scope, not during the incident review.
What Not To Do
Don’t read a model upgrade as a promotion for the agent running on it. Release notes describe capability, not your control over it. Don’t govern every agent the same way, a read-only summarizer and an agent that sends customer emails do not need the same leash. And don’t mistake fewer interruptions for more trust. In that same Anthropic data, experienced users let agents run longer yet interrupted more often, 9% against 5%.
Oversight did not shrink. It changed shape.
Signal Boost
Anthropic’s Measuring Agent Autonomy (2026) is a rare look at how people actually supervise AI agents at scale, not a benchmark. It shows autonomy widening with operator experience while oversight changes form. Read it before deciding any agent has earned a longer leash.