The Enterprise Copilot Rollout Checklist

The pilot went well. Everyone liked it. Then security asks who owns the agent and what it can reach, finance asks what the message bill is, and someone in the pilot cohort finds a salary band they were never meant to see. The rollout stops, and nobody can say for certain when it starts again.

This resource is the sequence that prevents that, written for the person who has to sign the rollout off rather than the person building it. Six phases and 52 checks, ordered the way failures actually surface, so the expensive discoveries happen in week one instead of month three.

Inside:

Licensing and entitlement. The licence mix, the capacity model, and the trap that catches nearly everyone: seats and consumption are two separate bills, and a per-user Copilot licence does not cover a single Copilot Studio message or a single Foundry token.

Data boundaries and permissions. The oversharing assessment, broad tenant-wide grants, sensitivity labels, restricted content and the Purview coverage legal will assume is already handled. Copilot respects existing permissions exactly, which is the risk and not the control.

Agent identity and registration. Entra Agent IDs, the Agent 365 registry, scoping permissions to the job rather than inheriting them from the maker, and locking the default Copilot Studio environment where shadow agents are born.

Prove it works. What a review board actually accepts as evidence, and why a golden set with a recorded pass rate is the thing that turns "we think it is fine" into a signature.

Rollout, audit and lifecycle. The pilot cohort, the acceptable use policy, the incident path, quarterly access reviews, and the offboarding trigger that stops a leaver turning into an orphaned agent with live permissions.

A printable sign-off sheet. One row per phase, the evidence required, and ruled lines for owner and date. This is the page people keep.