Five Questions to Answer Before AI Agents Request Access

Agentic access is not a future problem. If your enterprise is piloting AI agents anywhere — coding assistants, support bots, workflow agents — those agents are already touching systems that your governance program is accountable for. Before the volume grows, your team should be able to answer five questions.

First: lineage. For every agent, who built it, who deployed it, and which human or team is accountable for what it does? An agent without a traceable owner is a service account problem wearing a newer costume.

Second: scope. What did the agent declare it would do, and does its actual access match? Scope drift in agents happens faster than in any human population, because agents request access programmatically.

Third: separation of duties. Can an agent accumulate combinations of access that would be flagged instantly if a human requested them? Most SoD engines have never been pointed at bots.

Fourth: expiry. Agents get decommissioned, forked, and replaced constantly. Does their access die with them, or does it linger the way orphaned service accounts always have?

Fifth: continuous evaluation. Agent context changes by the hour, not by the quarter. A certification campaign that reviews agent access annually is reviewing a snapshot of something that no longer exists. If your program can answer all five, you are ahead of the market. If not, that gap is exactly what a reasoning layer is for.