An AI agent is just another identity. Treat it like one

The question we get asked most when the word "agent" comes up isn't how much it costs or how long it takes. It's this: how do I keep it from showing someone what they shouldn't see?

It's the right question, and the answer is less exotic than it seems. An agent isn't a new creature that requires a governance framework invented from scratch. It is just another identity in your organization, and it is governed with the tools you already use to govern the others.

1 · It reads with the permissions of whoever is asking

When an agent relies on SharePoint, it queries with the permissions of the person talking to it. If that person can't open the payroll folder, the agent won't show it to them either: nobody gets a document through an agent that they couldn't open on their own.

This has an uncomfortable consequence worth saying out loud: if your SharePoint permissions are set up wrong, the agent won't fix them. It will make them obvious. Many agent projects end up starting with a permissions cleanup, and that's how it should be: that mess already existed, it's just that nobody was querying it at the speed an agent asks questions.

2 · It has an owner

Every agent that goes into production needs a named owner: someone who answers for what it does, reviews its sources when the business changes and decides when it gets retired. Without an owner, an agent outlives the person who created it, keeps answering with outdated information and nobody feels authorized to turn it off.

The control question is simple and worth asking before you start: if this agent gives a wrong answer tomorrow, who gets notified? If there is no answer, it isn't ready for production yet.

3 · It leaves a trail

An agent without traceability is an employee without an activity log: while everything goes well, it doesn't matter; the day something goes wrong, there is no way to reconstruct what happened. With Agent 365 you know who created each agent, what it accesses and what it has done, just as you know for any other account.

That turns a discussion about trust ("do we believe the artificial intelligence?") into a discussion about auditing, which is much easier to have.

The full lifecycle, which almost nobody defines

• Onboarding: why it exists, which sources it can read and who answers for it.
• Permissions: limited to what it needs, not to what would be convenient.
• Review: when the process that gives it a purpose changes, its sources are reviewed.
• Offboarding: an agent that is no longer used gets retired, just as an account does.

And adoption, which decides whether it was worth it

The same applies to Copilot: it is adopted by role, not by license. You identify the specific tasks of each profile (what a purchasing analyst does by hand, what a service coordinator repeats every week) and support the team until the tool becomes part of their work. Handing out licenses and waiting is not an adoption strategy; it's well-intentioned spending.

If you're thinking about putting an agent to work on your company's documents, let's talk about permissions, ownership and traceability first. It's half an hour that saves months.