Ask a team what permissions their AI agent needs and you will usually get a generous answer. Nobody wants the agent to fail in production on a missing permission, so access gets granted broadly, with the intention of tightening it later. Later rarely arrives.
Checking a policy against a checklist only gets you so far. A rule that flags wildcard access catches the obvious cases, but a policy can look reasonable on paper and still contain dozens of permissions the agent has never exercised. The question that matters in a review is not just what the agent is allowed to do, but what it has actually done.
The evidence for that already exists. AWS CloudTrail, the Azure Activity Log and Google Cloud Audit Logs record the calls that agent identities make. Comparing 90 days of that history with the permissions granted produces a short, defensible list: these are allowed, and have not been used.
Why 90 days, and why the wait matters. A shorter window mistakes a quarterly job for an unused permission. A longer one delays every review. Equally important, a brand-new agent has no history to judge, so evaluating it on day one would produce nothing but false alarms. A sound approach waits until there is a full window of history and says so, rather than reporting an empty result as if it meant everything was fine.
How to act on the list. Unused does not automatically mean remove. Some permissions exist for rare but important events like a failover or a monthly close. The value of the list is that it turns a vague conversation into a specific one: an owner can look at each unused permission and either remove it or write down why it stays.
CloudSentry AI runs this comparison continuously, reading real grants and real activity from the cloud rather than a spreadsheet, and flags granted-but-unused permissions as findings with the agent and owner attached. Coverage today is complete for AWS and covers Azure app-registration agents and Google Cloud service-account agents; the remaining native resource types are on the roadmap, and we say so plainly.
Least privilege stops being an aspiration when it has evidence behind it. For autonomous agents, that evidence is sitting in your logs.