How do you discover shadow AI usage with Microsoft Entra Global Secure Access?

What “discover” means in practice for shadow AI

Discovery is not the same as blocking. With Microsoft Entra Global Secure Access, the useful signal is whether traffic is reaching AI services that are not on the approved list, then whether that activity is consistent with legitimate business use or a hidden tool being adopted outside policy. The point is to build an inventory from observed behavior, not from self-reporting.

For that reason, the first question is always whether you have enough visibility to distinguish sanctioned AI access from unmanaged usage. If the control plane can tell you which users, devices, and locations are reaching specific services, you can start separating routine access from shadow AI patterns such as repeated visits to consumer AI sites, unusual browser-based use, or traffic that appears only after hours.

That visibility matters because shadow AI often hides in ordinary web traffic. The practical discovery task is to identify the services being reached, compare them against policy, and then look for adjacent signals such as repeated blocked attempts, new destinations, or sudden growth in usage from a team that was not expected to adopt AI tools yet.

How to interpret the traffic signals you see

Discovery becomes reliable when you treat the network view as a starting point for investigation, not as the full answer. A single request to an AI service may be incidental, but repeated sessions, sustained uploads, or access from multiple endpoints in the same group can indicate an unmanaged workflow taking shape.

It also helps to compare observed usage with your approved access policy and user population. If a service is not sanctioned, but traffic is still flowing to it, that is a governance signal even before you know whether sensitive data is involved. If access is being blocked, the blocked events themselves are useful because they show both the attempted tool choice and where enforcement is preventing drift from policy.

Microsoft Entra Global Secure Access is most useful when it gives you central visibility across users and services, then lets you pivot from a destination to the source context. That makes it easier to answer the operational questions that matter: who is connecting, from where, on what device posture, and to which AI service.

What to do after you identify unmanaged AI use

Once a shadow AI service is identified, the next step is to decide whether the activity is a policy exception, a one-off test, or an unmanaged dependency that needs containment. If the service is business-relevant, bring it into the approved set and document the conditions under which it may be used. If it is not approved, restrict or block it and make sure the policy path is clear enough that users know which tools are acceptable instead.

This is where Microsoft Entra Global Secure Access should be paired with a clean approval process. Discovery is only useful if you can convert what you learn into an action: add the service to an approved catalog, tighten the policy, or investigate the user population that adopted the tool without review. The discovery result should always lead to a governance decision, not just a report.

When the evidence suggests risky behavior, treat it as an access and data handling issue as well as a usage issue. Unmanaged AI services can create exposure through prompt content, uploaded files, browser sessions, or third-party integration paths, so the response should consider whether users must be directed to safer approved tools, whether exceptions need time limits, and whether the access pattern shows broader adoption than policy expected.

Risk and Threat Considerations

Shadow AI is risky because users can move sensitive work into services that were never reviewed for data handling, retention, or integration behavior. Discovery helps you see the service before it becomes embedded in a workflow, but it also reveals where enforcement gaps, policy ambiguity, or user workarounds are already allowing unmanaged adoption.

Failure mechanism: Users route business work to unsanctioned AI services through ordinary web access, then upload content or credentials into a tool that the organisation cannot govern, monitor, or reliably audit.

Impact: Sensitive data can leave approved boundaries, approved controls can be bypassed, and the organisation may not learn about the exposure until usage patterns, blocked attempts, or downstream incidents reveal it.

Practitioner Guidance

What to verify: Confirm that your approved AI service list is current before you rely on discovery data. If the policy baseline is stale, you will misclassify legitimate usage as shadow AI or miss newly adopted tools that should have been reviewed.

Decision rule: If a service is consistently used but not approved, treat it as a governance gap first and an enforcement problem second. If it is a one-off access attempt, use the signal to validate whether the user simply needs a sanctioned alternative or whether a broader adoption pattern is emerging.

What good looks like: You can distinguish sanctioned AI use from unmanaged use, explain the difference by user, location, and service, and move quickly from discovery to either approval, restriction, or investigation. Discovery is working when it produces a defensible inventory of real usage rather than a vague list of suspicious destinations.

Practitioner takeaway: The goal is not to catch every AI request, but to expose where AI use has moved outside policy early enough that you can govern it before it becomes normalised.

Related resources from NHI Mgmt Group

• Guide: Shadow AI and AI Agent Discovery Guide
• How do you discover AI agents in use with Microsoft Global Secure Access?
• How do you secure AI agent identities with Microsoft Entra Conditional Access?
• How do you apply Conditional Access policies to AI agents in Microsoft Entra?
• How do you secure Model Context Protocol traffic with Microsoft Global Secure Access?