Secure Copilot and Microsoft 365 AI rollout
Copilot sees only what the user can see. So before the rollout, we get permissions, sensitivity labels, and data protection rules in order.
Before you switch it on
Where Copilot shows more than it should
If an employee can see more than they should, Copilot will find it for them, even if they would never have gone looking. We wrote about the permission risks of enterprise AI in this article.
Accumulated permissions
Someone moves to another department, but their old access stays. Regular access reviews prevent this.
Overly broad sharing
A folder shared with the whole organization, or a link anyone can open. Their contents can also show up in Copilot’s answers.
Unlabeled sensitive data
If you do not know what counts as sensitive data and where it is stored, you cannot monitor where it goes. A sensitivity label makes it traceable.
A simple test
A junior colleague asks the assistant how much the managing director earns. If the assistant does not clearly refuse, your permission model is broken.
Rollout
How we roll it out
We work in four steps, each building on the one before.
-
Sharing and permissions
We identify where a user or group can see more than their work requires, and narrow that access.
-
Labels and DLP rules
Sensitive documents get sensitivity labels, and data loss prevention rules define what may leave the organization.
-
Pilot group
A designated group uses Copilot first. Based on their experience, we adjust the settings before access is widened.
-
Phased expansion
The remaining users get access group by group, on a schedule agreed in advance.
Deliverables
What you get at the end
items handed over
- Overly broad sharing and permissions identified and fixed
- Sensitivity labels and data loss prevention rules
- Pilot group launch, then phased expansion
- User and administrator guide
Related services
What goes with Copilot
These services are worth doing before or alongside a Copilot rollout.
Microsoft 365 security hardening
Checking and fixing the security settings of your Microsoft 365 and Entra ID tenant.
Data loss prevention (DLP)
Stopping sensitive data from leaking by email, cloud storage, printer, or removable media.
GenAI data loss prevention
Keeping sensitive data out of AI tools, specifically for traffic going to AI services.
Security and AI usage training
Live training tailored to your company, for managers and staff, on using AI tools safely.
Frequently asked questions
The questions we hear most
How is this different from Microsoft 365 security hardening?
Hardening reviews the tenant settings: authentication, Conditional Access, sharing, email, and logging. A Copilot rollout sorts out which data your users can reach and which of it is sensitive. We covered common tenant mistakes in our article on Microsoft 365 security.
What can a malicious email or document make Copilot do?
Copilot also processes incoming email and the documents users open, and these can contain hidden instructions. This is indirect prompt injection. It cannot be eliminated entirely, so the defense focuses on limiting the impact: the model should reach only the data it needs, and risky actions should require human approval. In our AI security controls deployment service, we install vendor solutions that govern the use of AI services and filter out malicious instructions.
What about the other AI tools employees use?
Copilot does not replace them automatically. It is common for a department to use an AI tool on its own subscription and upload its own data while IT knows nothing about it. A ban on its own does not work, because employees find ways around it. So alongside a well-configured company tool that people can actually use, it also pays to set clear rules on what can be entered where. The Shadow AI assessment uncovers the tools actually in use.
