Skip to content
Back to overview Abstract illustration of network traffic routed through a central control point to approved AI services
Global Secure Access AI Entra ID

Using Global Secure Access to control which AI tools are used in your organisation

8 min read

Key takeaways

  • A blanket AI ban is rarely followed in practice. AI gets used anyway, just where IT is not looking.
  • A defined set of approved tools, such as Microsoft 365 Copilot and Claude Enterprise, works better. Everything else is made visible or blocked.
  • Microsoft Entra Internet Access, part of Global Secure Access, implements this with web filtering, a category for generative AI, domain-level allow rules and Conditional Access.
  • Roll-out starts with observation; blocking comes later. The logs then serve as evidence for auditors.

One pattern comes up again and again in our conversations with managing directors and IT leads: AI chat tools are prohibited or unregulated, and AI is used every day regardless. Employees paste proposals, draft contracts or customer data into free services, often in a private browser profile or on their phone. This is rarely carelessness; the tools are useful and there is no approved alternative.

This article shows how to allow approved tools with Microsoft Entra, make the rest visible or block it, and document the result for an audit.

Why do AI bans usually fail in organisations?

A ban without enforcement and without an alternative only moves usage out of sight. The risks remain, but nobody can quantify them, and the board does not know which data is leaving the company.

A sustainable approach has three parts: approve a few AI tools under a business contract, make other usage visible, and block what is not acceptable. The technical foundation is Global Secure Access, Microsoft’s service for securing network access, integrated with Microsoft Entra ID.

What risks arise when employees use free AI chat tools?

Data leaving the company. Inputs sit with a provider the company has no contract with. Whether they are used for training, how long they are kept and where they are processed is governed by consumer terms, not by your requirements.

Liability. With personal data, there is usually no data processing agreement under Art. 28 GDPR. For organisations in scope of NIS2, Art. 20 of the NIS2 Directive requires management bodies to approve cybersecurity risk-management measures and oversee their implementation. Uncontrolled AI usage belongs in that risk management.

No audit trail. When an auditor, regulator or customer asks which AI services are in use, there is no reliable answer without technical records.

How does Global Secure Access control access to AI tools?

Microsoft Entra Internet Access routes internet traffic from managed devices through the Global Secure Access client to Microsoft’s Security Service Edge, where the filtering rules apply. Microsoft Entra Conditional Access decides which rules apply to whom. IT works with four building blocks:

  1. Web filtering policy (now Web Filtering in the portal, previously Web Content Filtering). Microsoft provides the web category “Artificial Intelligence” for sites using generative AI. You can add rules for individual domains (FQDNs) and, with TLS inspection, full URLs.
  2. Allow rules. A policy with the action “Allow” contains the domains of approved services, for example Microsoft 365 Copilot and Claude Enterprise, and gets a higher priority in the security profile than the block on the AI category. Use the domain lists published by Microsoft and Anthropic, not guessed entries.
  3. Security profile. It bundles the policies in the intended order.
  4. Conditional Access policy. It links the profile to users or groups through the session control “Use Global Secure Access security profile”, for example to a pilot group first.

Visibility comes from shadow AI discovery under Global Secure Access > Applications > Insights and Analytics. It matches traffic against the Microsoft Defender for Cloud Apps catalogue and shows which AI services are used, by whom, how often and with how much data, with a risk score per service.

In what order should you introduce the rules?

The key decision for the board is not to block on day one. Blocking without data hits business-critical workflows and costs employee acceptance. We recommend three phases:

Phase 1: Observe. Enable the Internet Access traffic forwarding profile for a pilot group, initially without block rules. Web filtering knows the actions “Allow” and “Block”; There is no documented log-only mode. Observation therefore relies on the traffic logs and shadow AI discovery.

Phase 2: Evaluate. After two to four weeks you can see which services are used, which departments have a genuine need, whether the approved tools cover it and which exceptions are justified.

Phase 3: Enforce. Then activate the block on the AI category with the allow rules in front of it, first for the pilot group, then for everyone. Communicate beforehand which tools are available and how to request approval. Newly assigned security profiles only take effect with a new access token, so expect a delay.

What prerequisites and licences do you need?

  • Licence. Microsoft Entra Internet Access is sold standalone or in Microsoft Entra Suite and is included in Microsoft 365 E7. Microsoft Entra ID P1 is a prerequisite. Check the current bundles before purchasing.
  • Client and devices. The Global Secure Access client runs on Windows, macOS, iOS and Android. On Windows, internet traffic is only acquired on Microsoft Entra joined or hybrid joined devices; registered-only devices currently get Private Access only. Deploy the client through Microsoft Intune.
  • Network settings. Disable DNS over HTTPS; the client does not acquire IPv6 traffic.

Where are the limitations?

  • Unmanaged devices. Private devices without the client are not covered. Conditional Access can restrict the approved business services to compliant devices, but it does not stop a private device from reaching arbitrary AI sites. The acceptable-use policy has to cover that.
  • Same domain, different account. A domain allow rule does not distinguish between a company and a private account with the same provider. For Microsoft services, Global Secure Access offers Universal Tenant Restrictions. For Claude, enforce sign-in through single sign-on and clarify with Anthropic how usage can be restricted to your organisation.
  • TLS inspection. Without it, filtering uses the server name, which is enough for categories and domains. URL rules, content inspection and prompt logging require TLS inspection, with your own certificate authority or a Microsoft-managed certificate (the latter marked as preview, as of October 2026). On mobile devices, certificate pinning can break individual apps.

What evidence do the logs provide?

The traffic logs under Global Secure Access > Monitor > Traffic logs show user, destination, applied policy and action for each connection. They demonstrate that a rule exists, who it applies to and that it is enforced; shadow AI discovery adds the overview of services in use. The portal retains logs only for a limited period, so export them through diagnostic settings to Azure Monitor Log Analytics or Microsoft Sentinel.

To capture prompt content, Microsoft offers Generative AI Insights, marked as preview (as of October 2026) and dependent on TLS inspection. Be cautious: this processes employees’ personal data. Agree purpose, scope and retention with your data protection officer and, where one exists, the works council first.

The next step

Global Secure Access controls which tools are used. To make sure the approved tools themselves do not expose data, you need a clean foundation in Microsoft Entra ID, SharePoint and Microsoft Purview. How we make AI adoption audit-ready is described on our AI Foundation page. If you first want to know where your tenant stands, our fixed-price Gap Assessment is the right starting point.

Book an introductory call

Frequently Asked Questions

Can Global Secure Access block ChatGPT and similar services while allowing Microsoft 365 Copilot?

Yes. Block the web category “Artificial Intelligence” and create an allow policy with the domains of the approved services, with a higher priority in the security profile. Conditional Access assigns it to users or groups. Test with a pilot group before rolling out the block.

Is there a test mode in which rules are only logged?

Web filtering policies themselves know the actions “Allow” and “Block”. For an observation phase, enable traffic forwarding without block rules and review the traffic logs and shadow AI discovery. Activate the blocks only after that.

Which licence do I need to control AI tools with Global Secure Access?

Microsoft Entra Internet Access, standalone or as part of Microsoft Entra Suite, on top of Microsoft Entra ID P1. Internet Access is also included in Microsoft 365 E7. Regardless of licence, private devices without the Global Secure Access client are not covered.