Copilot works with the access your people already have. That makes the review before a wider rollout mostly about what is reachable, who can build agents, and whether you can see what happens afterwards. This checklist covers seven surfaces. Use it as a list of things to review and confirm, then assign each finding to an owner.
What "read-only review" means
A read-only review only reads. It looks at configuration, inventory and logs through Microsoft Graph, Power Platform APIs, Azure RBAC and Purview, and it does not change a setting. You grant read access, you get a list of findings, and your own administrators decide what to fix and when. That also makes the review safe to repeat.
Before the review: what to collect
Gather a few things first, so the review starts from facts and not from guesses. None of it needs a change to the tenant.
- Tenant list. Every Microsoft tenant in scope, with a named owner for each. A review of one tenant says nothing about the one next to it.
- Licence counts. How many Microsoft 365 Copilot licences and how many Copilot Studio licences you hold, and who holds them. This is the baseline for the question of who should have access.
- Published agents and their owners. The list of agents that are live today, each with a person or team beside it. Agents with no owner are the first thing to chase.
- Sensitivity label taxonomy. The labels you have defined and what each is meant to protect. Without this, you cannot tell whether a gap in labelling is a gap or a choice.
- Purview audit retention setting. How long audit records are kept today. The logging surface depends on it, and it is quick to look up.
- Conditional access policies. Any existing policy that mentions Copilot or Power Platform. These show what you already decided, so the review can check it still holds.
Put these in one shared page. The reviewer, the tenant administrators and the data owners should all be looking at the same list.
The seven surfaces
1. Microsoft 365 Copilot and Purview
- Review sensitivity labels: which content is labelled, and which sensitive sites have no label.
- Review oversharing in SharePoint, such as sites open to everyone in the organisation or links open to anyone.
- Confirm Purview auditing is enabled and that Copilot interactions appear in the audit log.
- Review data loss prevention policies and confirm whether they apply to Copilot interactions.
- List who holds Copilot licences and compare it with who should.
2. Entra ID
- Review conditional access policies that apply to Copilot, Copilot Studio and other agent sign-ins.
- List app registrations and enterprise applications that agents use, and who owns each one.
- Review app consent settings: who can consent, and which consents are tenant-wide.
- List service principals with broad directory or mail permissions.
- Confirm multi-factor authentication covers the people who can build or publish agents.
3. Copilot Studio
- List every agent, its owner, and whether it is published.
- Review each agent's authentication setting and confirm which ones are open to people outside the tenant.
- Review the connectors and actions each agent can use, especially those that write data or send messages.
- Confirm who can create and publish agents, and whether that group is as small as it should be.
4. Azure AI
- List Azure AI and Foundry resources, models and deployments across subscriptions.
- Confirm content filters are set on each deployment.
- Review network isolation: which resources accept traffic from the public internet.
- Review how resources authenticate, and list any that still depend on shared keys.
- Review who holds roles that can create or change deployments.
5. Azure platform
- Review RBAC assignments on subscriptions and resource groups that hold AI resources.
- Confirm diagnostic logging is turned on for those resources.
- Review which identities have owner or contributor rights, and whether each is still needed.
- List where secrets for AI workloads are stored.
6. Shadow AI
- List AI applications seen on managed devices.
- Review DNS or proxy logs for traffic to AI services you have not approved.
- List AI tools connected to Microsoft 365 through OAuth consent.
- Decide, for each one, whether to approve it, replace it or block it.
7. Logging
- Confirm retention periods for audit and diagnostic logs meet what you need for investigations.
- Review coverage: which of the surfaces above actually send logs to a place you can search.
- Confirm logs reach your SIEM, and that someone owns the alert queue.
- Test one scenario end to end, such as finding who asked Copilot for a given file last week.
- Confirm that admin actions, such as changing a policy or publishing an agent, are logged as well as user activity.
After the review: turning findings into an owner list
A finding without an owner does not get fixed. Build the list in a simple order.
- Group by surface. Record each finding with the surface, the resource and what you observed, then gather them under the seven surfaces. Teams tend to own a surface, so this matches how the work gets done.
- Assign an owner per surface. Name a team, not "IT". The person who can change the setting owns the fix. If a surface has two owners, split it until each finding has one.
- Set severity by data exposure. Rank a finding by the data it puts within reach, not by how easy the fix is. An open link to a sensitive site outranks a tidy-up on an unused app registration.
- Sort into 30, 60 and 90 day buckets. Put the findings that expose the most data in the first bucket, the structural ones in the second, and the housekeeping in the third. A date makes the owner list a plan.
- Re-run the read-only checks. After each bucket, run the same checks again and compare. A fix that was never rechecked is a claim, and the review is cheap to repeat because it changes nothing.
Share the list with the people who own the data as well as the people who own the tenant, since oversharing is usually a business decision. Keep the first pass small. Fix the findings in surfaces 1 to 3 before the rollout expands, since they decide what Copilot and agents can reach. The rest can follow on a schedule.
Where Aynigma fits
Aynigma's Microsoft AI posture review runs 73 read-only checks across the same seven surfaces: Azure AI (17), Entra ID (16), Microsoft 365 Copilot and Purview (14), Copilot Studio (9), Azure platform (7), shadow AI (6) and logging (4). It reads through Microsoft Graph, Power Platform APIs, Azure RBAC and Purview and changes nothing in your tenant. Findings can feed the readiness packs in compliance monitoring, which are a readiness aid and not a certification. More on the review is at Microsoft AI posture.