Prevent data leaks to shadow AI with Microsoft Purview.

People in your organisation are already pasting work into AI apps that IT never approved. This guide follows Microsoft’s four-step model for that problem. You find the AI apps in use and block the ones you won’t allow. Then you keep sensitive data out of the ones you do allow, and record what people send so you can investigate later.

It’s written for the security, Entra, Intune and Purview administrators who share the work. The model runs across four products: Defender for Cloud Apps finds the apps, Entra controls access, Intune controls installation and Purview protects the data.

Microsoft’s four steps on one page.

Microsoft’s blueprint sets out the model in four columns: discover, block access, secure data and govern data. Each activity carries a number for the product that does it: 1 Defender for Cloud Apps, 2 Entra, 3 Intune, 4 Purview.4

Prevent data leak to shadow AI blueprint: discover, block access, secure, govern
Microsoft’s blueprint for the model, shown whole. The numbers name the product behind each activity. Diagram · Microsoft Learn, Prevent data leak to shadow AI
  1. The blueprint’s govern column is headed “in Microsoft Edge”. That scope still holds: Communication Compliance, retention and eDiscovery for non-Microsoft AI apps work only in Edge, for the apps listed in Step 4.4, 14

Microsoft’s steps at a glance

StepWhat it doesOutcome
Step 1Discover AI appsUnknown AI app usage is discovered.
Step 2Block access to unsanctioned AI appsUnsanctioned AI apps are blocked.
Step 3Block sensitive data going to sanctioned AI appsSensitive data isn’t sent to sanctioned AI apps.
Step 4Govern data sent to AI appsPrompts sent to AI apps are retained for future investigations.

Four products, each licensed on its own terms.

Each feature has its own licence, and several features for non-Microsoft AI apps add pay-as-you-go charges on top. The table lists the plans each service description names. A feature marked pay-as-you-go needs that billing turned on as well as the plan.13, 15

Licences by feature 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25

FeatureLicenceNotes
Defender for Cloud Apps discovery, tagging and governanceDefender for Cloud Apps standalone, or in Enterprise Mobility + Security E5, Microsoft 365 E5, A5 or G5, Microsoft Defender Suite, Microsoft Purview Suite, or Microsoft 365 E5, F5 or G5 Information Protection and Governance.Cloud App Discovery, a subset of Defender for Cloud Apps, comes with Microsoft Entra ID P1, Enterprise Mobility + Security E3 and Microsoft 365 E3.
Blocking unsanctioned apps through Defender for EndpointDefender for Cloud Apps plus Defender for Endpoint, or Microsoft 365 E5.Devices onboarded to Defender for Endpoint, with network protection in block mode.
Entra Internet Access web content filteringMicrosoft Entra Internet Access standalone, or in Microsoft Entra Suite or Microsoft 365 E7.Users also need Microsoft Entra ID P1 or P2.
Intune app controlsMicrosoft Intune Plan 1, which most organisations get in a Microsoft 365 bundle such as E3, E5 or E7.
Insider Risk Management and Adaptive ProtectionMicrosoft 365 E5, A5 or G5, Microsoft Purview Suite, or Microsoft 365 E5, A5, F5 or G5 Insider Risk Management.
Sensitivity labelsManual labelling: Microsoft 365 E3 or E5 and other plans. Automatic labelling: E5-level plans.
Endpoint DLPMicrosoft 365 E5, A5 or G5, Microsoft Purview Suite, or Microsoft 365 E5, A5, F5 or G5 Information Protection and Governance.
DLP for cloud apps in Edge for BusinessMicrosoft 365 E5 or A5, Microsoft Purview Suite, Information Protection and Governance, or Office 365 E5 or A5.Pay-as-you-go when a managed device sends data to an unmanaged app, such as a consumer AI app.
Network Data SecurityWith Global Secure Access: Microsoft 365 E7, or Purview E5 or equivalent plus Entra Internet Access or equivalent. With a non-Microsoft SASE: Purview E5 plus pay-as-you-go.Non-Microsoft integrations are generally available or preview, depending on the integration, and some features are preview. Pay-as-you-go use is billed as Network Data Security (preview).
DSPM for AI (classic) for non-Microsoft AI appsPay-as-you-go.Microsoft 365 Copilot interactions are not billed this way.
Collection policiesNo licence of their own: you need the licence for the workload they cover.Network collection needs the tenant linked to an Azure subscription.
Audit of non-Microsoft AI appsPay-as-you-go. Enterprise subscriptions don’t include it.Records are kept for 180 days.
Communication ComplianceMicrosoft 365 E5, A5 or G5, Microsoft Purview Suite (including EDU, GOV and FLW), Microsoft Defender + Purview Suite FLW, Microsoft 365 E5, A5, F5 or G5 Insider Risk Management, or Office 365 E5, A5 or G5.Other AI apps are pay-as-you-go.
Retention policiesMicrosoft 365 E3 or E5, Business Premium, Office 365 E3 or E5, and other plans.Retention of AI interactions is metered pay-as-you-go.
eDiscoveryStandard search and hold: Microsoft 365 E3 or E5. Premium: E5-level plans.eDiscovery storage is metered pay-as-you-go.

Find the AI apps people already use.

Outcome
Unknown AI app usage is discovered.

This step watches and blocks nothing. By the end, you know which AI apps people use, how risky each one is, and which users send sensitive data to them.1

In today’s portal The second task uses DSPM for AI (classic). The current DSPM opens at Microsoft Purview portal > Solutions > DSPM, and the classic version sits beside it under its own name.3

  1. Review the generative AI apps in use

    Microsoft Defender portal › Cloud apps › Cloud discovery › Discovered apps

    1. In Browse by category, select Generative AI. The tab lists the apps your traffic reached, with their usage.1, 28
    2. Open each app your users reach and review its risk score and usage. Defender for Cloud Apps scores each app on more than 90 risk factors across general, security, compliance and legal categories.26
    3. Mark each app Sanctioned or Unsanctioned from the three-dot menu at the end of its row, following the AI app policy you agreed.1, 27
    Defender for Cloud Apps Tag app menu with Sanctioned and Unsanctioned outlined
    The three-dot menu at the end of an app’s row, where you tag it Sanctioned or Unsanctioned. Screenshot · Microsoft Learn, Govern discovered apps in Microsoft Defender for Cloud Apps

    Check The Discovered apps tab shows each AI app with the tag you set on it.28

    Watch out

    Marking an app unsanctioned doesn’t block it. It flags the app so you can monitor it. Blocking comes in Step 2, and only through a connected enforcement route.27

  2. Turn on the DSPM for AI discovery policies

    Microsoft Purview portal › Solutions › DSPM for AI (classic)

    1. In Overview > Get Started, complete Install Microsoft Purview browser extension and Onboard devices to Microsoft Purview. The policy that detects visits to AI sites needs both.14, 29
    2. Turn on the three one-click discovery policies under Extend your insights for data discovery. Each one runs in audit mode and blocks nothing.14, 29

    The three discovery policies 1, 29

    PolicyWhat it detectsBuilt on
    Detect when users visit AI sites.Visits to AI websites.Insider Risk Management.
    Detect sensitive info shared in AI prompts in Microsoft Edge.Sensitive information in prompts in Edge. It doesn’t capture the prompt text.A collection policy.
    Detect sensitive info pasted or uploaded to AI sites.Sensitive information pasted or uploaded to AI sites in Edge, Chrome or Firefox.Endpoint DLP.

    Why The catalog shows which apps are used. These policies show who uses them, and whether sensitive information goes with them.1

    Check Wait at least a day, then open Reports. User activity with AI apps starts to appear there.14

  3. Watch for risky AI use with Insider Risk Management

    Microsoft Purview portal › Insider Risk Management › Policies › Create policy

    1. On the Policy template page, select the Risky AI usage template, then complete the policy workflow.1, 30
    2. Know what it sees. Its browsing detection covers visits to generative AI websites, which is the part that reaches shadow AI. Its prompt and response detection covers only Microsoft 365 Copilot, Microsoft Copilot and agents.31
    3. Deploy the extension the template needs: the Microsoft Insider risk extension in Edge, or the Microsoft Purview extension in Chrome.31

    Why The template’s alerts raise a user’s risk score, which Adaptive Protection uses in Steps 2 and 4 to restrict risky users automatically.5, 31

Done when Every AI app your traffic reaches is tagged Sanctioned or Unsanctioned, the DSPM for AI (classic) reports show user activity, and Insider Risk Management is watching for risky AI use.

Next Once you know which apps to refuse, Step 2 blocks them.

Block the AI apps you won’t allow.

Outcome
Unsanctioned AI apps are blocked.

This step turns your Step 1 tags into blocks at three levels: for the whole organisation, for chosen users and groups, and on the device itself. It’s the first step your users will notice.32

  1. Unsanction the AI apps you won’t allow

    Microsoft Defender portal › Cloud apps › Cloud discovery › Discovered apps

    1. In Browse by category, select Generative AI.28
    2. Select the apps you won’t allow with the bulk action checkbox, then mark them Unsanctioned in one action.28, 32

    An unsanctioned tag blocks only when an enforcement route is connected. The first route is the Defender for Endpoint integration in the next task. The second is a supported secure web gateway integration: Zscaler NSS, iboss, Corrata, Menlo or Open Systems. The third is a block script for your own appliance, from Actions > Generate block script.27

    We recommend

    We recommend writing down the routes these blocks leave open from the start. Users can still reach blocked apps from unmanaged devices or personal networks, and the later steps close only some of those routes.32

  2. Enforce the block through Defender for Endpoint

    Microsoft Defender portal › Settings › Cloud Apps › Cloud Discovery › Microsoft Defender for Endpoint

    1. Check the prerequisites: devices onboarded to Defender for Endpoint, Microsoft Defender Antivirus with real-time and cloud-delivered protection, and network protection turned on in block mode.18
    2. Select Enforce app access.18
    3. Go to Settings > Endpoints > Advanced features and turn on Custom network indicators.18
    Defender for Endpoint integration with Enforce app access to block shadow AI
    Enforce app access, the setting that turns unsanctioned tags into blocks. Screenshot · Microsoft Learn, Govern discovered apps using Microsoft Defender for Endpoint
    Defender for Endpoint Custom network indicators setting turned on
    Custom network indicators, turned on under Advanced features. Screenshot · Microsoft Learn, Govern discovered apps using Microsoft Defender for Endpoint

    Check Confirm the block reaches your network controls: each unsanctioned app’s domains appear as custom URL indicators in Defender for Endpoint, usually within a few minutes. The Enforce app access setting can take up to 30 minutes to take effect, and a block through Defender for Endpoint up to three hours.18, 32

    Watch out

    Browsers other than Edge need the Microsoft Defender Browser Protection add-on before the block applies in them.27

  3. Restrict AI apps for specific users and groups with Entra Internet Access

    Microsoft Entra admin center › Global Secure Access › Secure › Web content filtering policies

    1. Turn on the Internet Access traffic forwarding profile and install the Global Secure Access client on devices. Turn off DNS over HTTPS and the browser’s built-in DNS client.6
    2. Select Create policy, then Add rule. Choose the Artificial Intelligence web category (sites that use generative AI), or enter an FQDN for one app. Set the policy action to Block.6, 33
    3. In Global Secure Access > Secure > Security profiles, select Create profile, then Link a policy > Existing policy, and add the policy.6
    4. In Entra ID > Conditional Access > Policies, create a policy for the users or groups to restrict. Set Target resources to All internet resources with Global Secure Access, and Session to Use Global Secure Access security profile with your profile.6

    Why Security profiles linked to Conditional Access let you allow some people limited access to an AI app while others are blocked.32, 34

    Check A new security profile can take 60 to 90 minutes to apply, and a Conditional Access change about an hour. Then open Global Secure Access > Monitor > Traffic logs to confirm traffic is blocked or allowed as you intended.6

  4. Block elevated-risk users with Adaptive Protection

    Microsoft Purview portal › Insider Risk Management › Adaptive protection › Dashboard

    1. Select Quick setup. It creates DLP policies in test mode and a Conditional Access policy called Block access for users with Insider Risk (Preview). That policy starts in report-only mode and covers Office 365 apps for users at the Elevated level. Setup can take up to 72 hours.5
    2. For your own policy, go to Entra ID > Conditional Access > Policies in the Entra admin center. Under Conditions > Insider risk, set Configure to Yes and select Elevated. Set Grant to Block access.5, 35
    3. Leave Enable policy on Report-only until you’ve confirmed who it affects, then set it to On.35
    Adaptive Protection dashboard: risk levels, DLP complete, Conditional Access not
    The Adaptive Protection dashboard, with Quick setup beside the Conditional Access row. Screenshot · Microsoft Learn, Help dynamically mitigate risks with Adaptive Protection
    Adaptive Protection Conditional Access page listing three insider risk policies
    Conditional Access policies that use the insider risk condition. The quick setup policy starts in report-only mode. Screenshot · Microsoft Learn, Help dynamically mitigate risks with Adaptive Protection

    Video · Microsoft Learn · 2 min 17 s

    How Adaptive Protection works. Insider Risk Management assigns each user a risk level. DLP and Conditional Access policies apply stricter controls at higher levels, and ease them again as the risk falls. From Help dynamically mitigate risks with Adaptive Protection, Microsoft Learn. Video by Microsoft, with Microsoft's captions in the player.

    We recommend

    We recommend setting the policy’s Target resources to All internet resources with Global Secure Access, since the quick setup policy reaches Office 365 apps only. Be clear what that does: with Block access, elevated-risk users lose all internet traffic through Global Secure Access, not only AI apps. This is our reading: Learn documents the insider risk condition and the internet target, but not the two together. Exclude the Intune endpoints first, or blocking internet access can lock users out of compliance remediation.5, 6, 36

    Watch out

    Turning on Adaptive Protection also creates a data lifecycle management policy (preview). It keeps SharePoint, OneDrive and Exchange Online content deleted by Elevated users for 120 days. If Adaptive Protection was already on, you must opt in to it. The Adaptive protection in Data Lifecycle Management setting turns it off.5

  5. Block installation of AI apps with Intune

    Microsoft Intune admin center

    • Windows. In Endpoint security > App Control for Business, select Create Policy. Apps a policy doesn’t explicitly allow are blocked from running.37
    • Windows, store only. A device restrictions profile with Apps from store only set to Store Only blocks app installs from the internet. It doesn’t stop installs from USB devices, network shares or other non-internet sources.38
    • Android. Corporate devices block Play Store apps by default. If you allow them, add the AI app as a managed Google Play app. Then assign it as Uninstall at Apps > Android > Android apps > the app > Properties > Assignments > Edit.39

    Why Blocking the website doesn’t stop someone installing the app’s desktop or mobile version.32

    Watch out

    An App Control for Business policy blocks every app it doesn’t allow, unless you set it to audit mode. Start in audit mode, which runs all apps and logs them.37

Done when Unsanctioned AI apps are blocked through a connected route, chosen users and groups are restricted, elevated-risk users are blocked, and managed devices can’t install the apps you refused.

Next Step 3 protects the data people send to the apps you allow.

Keep sensitive data out of the AI apps you allow.

Outcome
Sensitive data isn’t sent to sanctioned AI apps.

Approved AI apps still need guarding. This step stops sensitive information being pasted, uploaded or sent to them. Each control covers different routes, so the last task maps them.40

  1. Encrypt sensitive content with sensitivity labels

    Apply sensitivity labels that encrypt your most sensitive content. Other AI apps can’t decrypt content encrypted by Purview, so a labelled file pasted or uploaded into one can’t be processed.40, 41

    Watch out

    Labels protect the file, not the conversation. Purview doesn’t apply sensitivity labels or encryption to what people type into a non-Microsoft AI app.14

  2. Block paste and upload of sensitive information with Endpoint DLP

    Microsoft Purview portal › Data loss prevention › Policies

    1. Purview has a built-in Generative AI Websites group, which you can’t edit or delete. For your own list of AI sites, go to Data loss prevention > Settings > Endpoint settings > Browser and domain restrictions to sensitive data > Sensitive service domain groups. Select Create sensitive service domain group. A group holds up to 100 sites.42, 43
    2. To use the upload action, also configure the Sensitive service domains list in the same section. Upload to a restricted cloud service domain needs it.43
    3. Create a DLP policy with Devices as the location.40, 42
    4. Add a condition for content that contains sensitive information types or sensitivity labels.40
    5. Add the action Audit or restrict activities on devices. Under Service domain and browser activities, set Paste to supported browsers and Upload to a restricted cloud service domain to Audit, Block with override or Block.10, 42
    6. Select + Choose different restrictions for sensitive service domains, then Add group. Choose your AI group or Generative AI Websites, select Add, and set the action for that group.43, 44
    7. Run the policy in simulation mode first: nothing is enforced and events are audited.40, 45

    Check Review the DLP activity explorer for false positives and tune the conditions before you turn the policy on.40

    Watch out

    Copy to clipboard restricts copying from a protected file. It doesn’t look at which site the content is going to, so use paste and upload controls for AI sites.10

  3. Block sensitive prompts in Edge for Business

    Microsoft Purview portal › Data loss prevention › Policies › + Create policy

    1. Select Inline web traffic, then the custom template.46
    2. Select + Add cloud apps > Adaptive app scopes > All unmanaged AI apps > Add. These include ChatGPT (consumer), Microsoft Copilot (consumer), DeepSeek and Google Gemini.40, 46
    3. Under Choose where to enforce the policy, select Edge for Business.46
    4. Add the action Restrict browser and network activities. Set Text sent to or shared with cloud or AI apps and File uploaded to or shared with cloud or AI app to Block.46
    5. Set the policy mode to Run the policy in simulation mode. Choose the name carefully: you can’t rename a policy later.46

    Saving the first policy activates it in Edge automatically. Purview creates two security groups for included and excluded users, plus Edge and Intune configuration policies. With Block set, users in scope can’t open Firefox and other browsers, or Chrome without the Purview extension.9, 46

    Check The configuration shows in the Microsoft 365 admin center at Settings > Microsoft Edge, and in the Intune admin center at Devices > Configuration. If activation fails, Purview doesn’t enforce the policy until you fix the error.9, 22

    We recommend

    We recommend testing typed prompts separately from pasted text and file uploads. Edge treats typed text and file uploads as separate activities, and Endpoint DLP handles paste and upload as separate actions, so one passing test doesn’t prove the others.10, 22

  4. Extend detection to other browsers and apps with Network Data Security

    Microsoft Purview portal › Data loss prevention › Policies › + Create policy

    1. Connect a supported integration. With Entra Global Secure Access, devices must be Entra ID joined, with a TLS inspection policy and a content policy in Internet Access. For a non-Microsoft SASE, open the Security Store from the Data Loss Prevention solution and select Get solution.23, 47
    2. Create an Inline web traffic policy with Network and non-Microsoft secure browsers set to On.47
    3. Choose Audit only or Block for text sent, files uploaded, text received and files downloaded.23

    Check Allow up to 24 hours for policies to distribute and the first data to appear. After that, each request takes up to 30 minutes to show in the audit log and activity explorer.23

    Watch out

    Network Data Security sees only traffic routed through a supported integration, not all network traffic. Some of its features are in preview, and the activities and actions available depend on the integration.23, 47

  5. A note on Microsoft 365 Copilot

    Microsoft’s step also has Defender for Cloud Apps monitor Microsoft 365 Copilot use. It has dedicated detections for Copilot, and raises an alert on suspicious interactions. One example is someone reaching sensitive data through Copilot from a risky IP address. The alert carries the MITRE attack technique and IP address for your investigation.40, 48

  6. Check which routes each control covers

    No single control covers every route. Read across each row before you rely on a control for an AI app.

    Coverage by device, browser and network path 6, 10, 14, 18, 22, 23, 27, 32, 38, 39, 40, 42, 43, 47

    ControlDevicesBrowsers and appsNetwork pathDoesn’t cover
    Defender for Endpoint blocking (Step 2)Devices onboarded to Defender for Endpoint, with network protection in block mode.Edge. Other browsers need the Microsoft Defender Browser Protection add-on.Managed network connections.Unmanaged devices and personal networks.
    Entra Internet Access filtering (Step 2)Devices running the Global Secure Access client.Browsers with DNS over HTTPS and the built-in DNS client turned off.Traffic the client forwards to Internet Access.UDP (QUIC) traffic, unsupported in the current preview.
    Intune app controls (Step 2)Intune-managed Windows and Android devices.Installed apps, not websites.Not network based.Under Store Only, installs from USB devices or network shares.
    Sensitivity labels with encryptionAny device.Any AI app, which can’t decrypt the content.Any.Typed prompts, and files without an encrypting label.
    Endpoint DLPOnboarded Windows 10 and 11, and the three latest major macOS versions.Paste: Edge, Safari on macOS, and Chrome and Firefox with the Purview extension on Windows. Upload: Edge, or Chrome and Firefox with the Purview extension.Not network based: it acts on the device.Typed text. Pasted text is classified at the moment it’s pasted, whatever its source. Paste on macOS is preview.
    DLP in Edge for BusinessFor unmanaged AI apps, Intune-managed Windows 10 and 11 devices only, in every Edge profile including InPrivate.Edge for Business 144 or later, the two latest stable versions.Not network based: it acts in the browser.Unmanaged devices, B2B guest users, and other browsers.
    Network Data Security (some features preview)With Global Secure Access, Entra ID joined devices. Otherwise what the integration supports.Any browser, app, API or add-in whose traffic passes through the integration.Only traffic routed through a supported SASE or secure browser integration.Traffic outside the integration.

    We recommend

    We recommend marking each control you turn on against this table, and recording the routes that stay open for each sanctioned AI app.10, 22, 23

Done when Encrypting labels protect your most sensitive files. Endpoint DLP and Edge policies run in simulation or enforcement, traffic through a supported integration is inspected, and you know which routes remain open.

Next Step 4 records and governs what people send.

Record and govern what people send to AI apps.

Outcome
Prompts sent to AI apps are retained for future investigations.

The last step keeps a record you can act on. You audit AI interactions, capture prompts, detect inappropriate ones, retain them, search them when something goes wrong, and block users whose risk rises.49

In today’s portal The last task uses a DSPM for AI (classic) one-click policy. The current DSPM opens at Microsoft Purview portal > Solutions > DSPM.3

  1. Confirm auditing is on and add non-Microsoft AI apps

    Microsoft Purview portal › Audit

    1. In Exchange Online PowerShell, run Get-AdminAuditLogConfig | Format-List UnifiedAuditLogIngestionEnabled. True means auditing is on. Security & Compliance PowerShell always shows False, so don’t check there.12
    2. If it’s off, select Start recording user and admin activity on the Audit page. It can take up to 60 minutes to start.12
    3. Turn on pay-as-you-go audit for non-Microsoft AI apps. Your enterprise subscription doesn’t include these records, which are kept for 180 days as the AIAppInteraction record type.24
    4. Pay-as-you-go itself is turned on by a Global Administrator, in the Purview portal at Settings > Account details. Select an active Azure subscription and a resource group to bill against.51

    Watch out

    Audit records who used an AI app, when and where. To keep and search prompt content for non-Microsoft AI apps, you also need the collection policy in the next task.14, 24, 50

  2. Capture prompts with a collection policy

    Microsoft Purview portal › Data loss prevention › Classifiers › Collection policies › + Create policy

    1. Create the policy. The role you need is DLP Compliance Management.52
    2. Set the Content contains classifiers condition to All.53
    3. Select + Add data source and add the All unmanaged AI apps adaptive app scope, or unmanaged cloud apps in the generative AI category. Content capture works only for these AI sources.52, 53
    4. At the content capture step, change the default Don’t capture content to capture content.52, 53

    Or edit the collection policy that the Step 1 Edge one-click policy created. Collection policies from one-click policies can be edited like any other.53

    Why Communication Compliance, retention and eDiscovery see prompts and responses for non-Microsoft AI apps only through this setting. Without it, a policy records only the sensitive information it detects. The one-click Edge policy from Step 1 doesn’t capture content.14, 29, 53

    Check In Explorers > Activity explorer, AI interaction events show the prompts and responses. For unmanaged AI apps in Edge, they show text prompts only.29, 52

    Watch out

    Content capture doesn’t include the content of files people share with an AI app.53

  3. Detect inappropriate prompts with Communication Compliance

    Microsoft Purview portal › Communication Compliance › Policies

    1. Turn on pay-as-you-go billing, as in the first task. Communication Compliance needs it for other AI apps.25, 51
    2. Select Create policy, then a template or Custom policy. Name the policy, choose the users in scope and the reviewers. You can’t rename a policy later.54
    3. Or turn on the DSPM for AI (classic) one-click policy DSPM for AI - Unethical behavior in AI apps. It detects sensitive information in prompts and responses in Microsoft 365 Copilot, agents and other generative AI apps.29
    4. In the policy row, select More actions > Edit, then Next twice. In Choose locations to detect communications, select Other AI apps.25

    Creating policies needs the Communication Compliance or Communication Compliance Admins role group. Reviewers need Communication Compliance Analysts or Communication Compliance Investigators, and must be named as reviewers in the policy.8, 55

    Communication Compliance, retention and eDiscovery support other AI apps only in Edge, for ChatGPT, Microsoft Chat (consumer), Google Gemini and DeepSeek. That’s Microsoft’s current list.14

    Watch out

    Reviewing every interaction, with the review percentage at 100% and no conditions, can reach the storage limit.25

  4. Retain and delete prompts with a retention policy

    Microsoft Purview portal › Solutions › Data Lifecycle Management › Policies › Retention policies › New retention policy

    1. Keep the default of Full directory for admin units.50
    2. Select the Other AI Apps location.49, 50
    3. Set retention and deletion periods to your organisation’s requirements.49
    Retention policy timing: distributed within a day, applied within seven days
    How long a new retention policy takes to apply: up to seven days from submission. Diagram · Microsoft Learn, Create and configure retention policies

    Check Prompts are stored in a hidden folder in each user’s mailbox. The timer job that acts on them typically runs every 1 to 7 days, so a one-day deletion can take up to 16 days.56

    Watch out

    Other retention policies, Litigation Hold and eDiscovery holds suspend deletion.56

  5. Search AI interactions with eDiscovery

    Microsoft Purview portal › eDiscovery › Cases › Create case

    1. Create a case and add a search over the users’ mailboxes, where all prompts and responses from AI apps are stored.7, 14
    2. Add an Item class condition. Browser-based AI apps use IPM.SkypeTeams.Message.CloudAIApp.SaaS.<AppID>, and locally installed ones IPM.SkypeTeams.Message.ConnectedAIApp.Connector.<AppName>.7, 14
    3. Place a hold and export what you find.49, 57
    4. To review first, add the results to a review set, where you filter, tag and analyse them. Review sets are a premium eDiscovery feature, which needs Office 365 E5, Microsoft 365 E5 or a related E5 add-on.57

    Watch out

    Searching needs the eDiscovery Manager role group. Deleting needs the Search And Purge role, and removes up to 10 items per mailbox at a time.7

  6. Block users at every insider risk level from AI prompts in Edge

    Microsoft Purview portal › Solutions › DSPM for AI (classic)

    1. Find the one-click policy DSPM for AI - Block elevated risk users from submitting prompts to AI apps in Microsoft Edge. Despite its name, it blocks users at all three insider risk levels: Elevated, Moderate and Minor.29, 49
    2. Before you turn it on, check the conditions for each level at Insider Risk Management > Adaptive protection > Insider risk levels. The Minor and Moderate conditions decide how many people lose AI prompts in Edge.5, 29
    3. Then turn the policy on. It turns on Adaptive Protection if it isn’t already on.29
    Adaptive Protection insider risk levels page with Elevated, Moderate and Minor
    The conditions that put a user at each insider risk level. Screenshot · Microsoft Learn, Help dynamically mitigate risks with Adaptive Protection
    Adaptive Protection DLP policies page listing two active insider risk policies
    DLP policies that use the insider risk level condition, listed in Adaptive Protection. Screenshot · Microsoft Learn, Help dynamically mitigate risks with Adaptive Protection

    Watch out

    Microsoft’s Edge policies also block users in scope from browsers that aren’t protected.29

Done when Auditing is on, prompts to the supported AI apps are captured, checked, retained and searchable, and users at every insider risk level are blocked from prompting AI apps in Edge.

Next Read the limits before you sign the work off.

What these controls don’t cover.

  • Unmanaged devices and personal networks. Users can still reach blocked apps from them.32
  • Typed text. Only DLP in Edge for Business and Network Data Security inspect text people type. Endpoint DLP classifies text when it’s pasted, not as it’s typed.22, 23, 42
  • Apps outside the catalog. Defender for Cloud Apps can’t discover apps that aren’t in its catalog unless you suggest them or create a custom app. MCP servers and model providers have their own categories, separate from Generative AI.26
  • Changing app addresses. If a vendor changes or adds URLs, unsanctioned blocking through Defender for Endpoint can behave inconsistently. Some Microsoft services can’t be blocked at all.26, 27
  • Manual tags and policies. When a manual tag and a discovery policy disagree, the last operation applied wins.27
  • Edge for Business limits. Browser and network policies don’t apply to B2B guest users. They inspect up to 4 MB of text and files up to 3 MB. Some apps send content encoded or through changing endpoints. Tenants using multi-admin approval in Intune aren’t supported for unmanaged apps.22, 23
  • Overlapping policies. When the same user, context and action are in both, Endpoint DLP takes priority over DLP in Edge. Edge protections for managed apps might not apply to users also in scope of a Defender session policy or an Endpoint DLP policy.22
  • Adaptive Protection scope. Adaptive Protection’s DLP support covers Exchange, Microsoft Teams and devices. Turning it off can take up to six hours to stop assigning risk levels.5
  • Endpoint DLP changes. Policy updates generally take about an hour to reach devices.10
  • DSPM for AI (classic). Administrators restricted to an admin unit can’t create the one-click policies.29
  • Intune on iOS, iPadOS and macOS. This guide has no Learn steps for blocking AI app installs on these platforms. Check your Intune app restrictions for them before you rely on Step 2.

Where each step comes from.

  1. 1 Microsoft Learn Step 1: Discover AI apps
  2. 2 Microsoft Learn Learn about Data Security Posture Management for AI - (classic)
  3. 3 Microsoft Learn Learn about Data Security Posture Management
  4. 4 Microsoft Learn Prevent data leak to shadow AI
  5. 5 Microsoft Learn Help dynamically mitigate risks with Adaptive Protection
  6. 6 Microsoft Learn How to configure Global Secure Access web content filtering
  7. 7 Microsoft Learn Search for and delete AI application data in eDiscovery
  8. 8 Microsoft Learn Assign permissions in Communication Compliance
  9. 9 Microsoft Learn Automatic activation of your Microsoft Purview policy in Microsoft Edge
  10. 10 Microsoft Learn Learn about Endpoint data loss prevention
  11. 11 Microsoft Learn Cloud app discovery overview
  12. 12 Microsoft Learn Turn auditing on or off
  13. 13 Microsoft Learn Learn about Microsoft Purview billing models
  14. 14 Microsoft Learn Use Microsoft Purview to manage data security & compliance for other AI apps
  15. 15 Microsoft Learn Microsoft Purview service description
  16. 16 Microsoft Learn Microsoft Defender service description
  17. 17 Microsoft Learn Compare discovery capabilities for Defender for Cloud Apps and Cloud App Discovery?
  18. 18 Microsoft Learn Govern discovered apps using Microsoft Defender for Endpoint
  19. 19 Microsoft Learn What is Global Secure Access?
  20. 20 Microsoft Learn Microsoft Entra licensing
  21. 21 Microsoft Learn Microsoft Intune licensing
  22. 22 Microsoft Learn Learn about Data Loss Prevention for Cloud Apps in Edge for Business
  23. 23 Microsoft Learn Learn about Microsoft Purview Network Data Security
  24. 24 Microsoft Learn Audit logs for Copilot and AI applications
  25. 25 Microsoft Learn Configure a Communication Compliance policy to detect generative AI interactions
  26. 26 Microsoft Learn Find your cloud app and calculate risk scores
  27. 27 Microsoft Learn Govern discovered apps in Microsoft Defender for Cloud Apps
  28. 28 Microsoft Learn View discovered apps with the Cloud discovery dashboard
  29. 29 Microsoft Learn Considerations for DSPM for AI to manage data security and compliance protections for AI interactions - (classic)
  30. 30 Microsoft Learn Get started with Insider Risk Management
  31. 31 Microsoft Learn Learn about Insider Risk Management policy templates
  32. 32 Microsoft Learn Step 2: Block access to unsanctioned AI apps
  33. 33 Microsoft Learn Global Secure Access Web content filtering categories
  34. 34 Microsoft Learn Learn about Microsoft Entra Internet Access for all apps
  35. 35 Microsoft Learn Block access for users with insider risk
  36. 36 Microsoft Learn Universal Conditional Access through Global Secure Access
  37. 37 Microsoft Learn Manage approved apps for Windows devices with App Control for Business policy and Managed Installers for Microsoft Intune
  38. 38 Microsoft Learn Windows device settings to allow or restrict features using Intune
  39. 39 Microsoft Learn Manage AI on Android with Intune - A Guide for IT Admins
  40. 40 Microsoft Learn Step 3: Block sensitive data going to sanctioned AI apps
  41. 41 Microsoft Learn Learn about sensitivity labels
  42. 42 Microsoft Learn Help prevent leakage of sensitive content by restricting paste actions into browsers
  43. 43 Microsoft Learn Configure endpoint data loss prevention settings
  44. 44 Microsoft Learn Help prevent sharing of sensitive items with unauthorized cloud apps and services
  45. 45 Microsoft Learn Create and deploy data loss prevention policies
  46. 46 Microsoft Learn Help prevent sharing via Microsoft Edge for Business to unmanaged AI apps from managed devices
  47. 47 Microsoft Learn Use Network Data Security to help prevent sharing sensitive information with unmanaged AI
  48. 48 Microsoft Learn Protect your organization’s use of AI apps and data
  49. 49 Microsoft Learn Step 4: Govern data sent to AI apps
  50. 50 Microsoft Learn Create and configure retention policies
  51. 51 Microsoft Learn Enable Microsoft Purview pay-as-you-go features for new customers
  52. 52 Microsoft Learn Create and deploy collection policies
  53. 53 Microsoft Learn Collection Policies solution overview
  54. 54 Microsoft Learn Get started with Communication Compliance
  55. 55 Microsoft Learn Plan for Communication Compliance
  56. 56 Microsoft Learn Learn about retention for Copilot & AI apps
  57. 57 Microsoft Learn Learn about eDiscovery

Have your AI data controls checked.

LensIQ reviews your Defender, Entra, Intune and Purview settings against this model and tells you which routes to AI apps are still open.