Updated 19 October 2026: Microsoft has blocked new enablement of Restricted SharePoint Search (retiring 31 January 2027), so this post now points to Restricted Content Discovery and SharePoint Advanced Management instead, and the DSPM for AI general-availability date is corrected to its mid-2026 rollout.
Why DSPM for AI is the third leg of the Copilot governance stool
Microsoft Purview Data Security Posture Management for AI (DSPM for AI) reached general availability in mid-2026, rolling out between June and July according to Microsoft's Message Center notices. For Australian Microsoft 365 Copilot rollouts it closes a measurement gap that the prior toolset could not. Restricted Content Discovery (Microsoft Learn — Restricted Content Discovery, accessed 19 October 2026) keeps nominated sites out of Copilot responses while their access is fixed. SharePoint Advanced Management surfaces which sites are unsafe. Neither tells the governance team whether the constraints are working in practice. DSPM for AI does.
This walkthrough is a how-to companion to the guide on choosing which sites get Restricted Content Discovery and Restricted Access Control. It assumes the reader has already identified high-risk sites with SharePoint Advanced Management reports and is now operationalising the measurement layer for an Australian tenant subject to the Privacy Act 1988 (OAIC — Privacy Act 1988, accessed 19 October 2026).
Where this fits: the cadence below is what moves a tenant from L2 to L3 on the data protection pillar of the L1 to L5 Copilot Governance Maturity model. To benchmark your current level first, run the free Copilot Governance Maturity audit.
Prerequisites
The licence position is documented on Microsoft Learn in the current DSPM for AI article. Microsoft 365 E5 Compliance or the E5 Information Protection and Governance add-on unlocks the full feature set. Lower SKUs surface a subset of signals. Independent of the compliance SKU, SharePoint Advanced Management is bundled free with every Microsoft 365 Copilot licence, so the discovery feed is available to every tenant by default.
Activate the following before opening the Purview AI Hub:
- A published sensitivity-label policy with at least the four common Australian tiers: Public, Internal, Confidential, and Confidential — Personal Information.
- A Purview retention label policy applied to the SharePoint sites that currently have Restricted Content Discovery applied.
- Insider Risk Management enabled with at least the default policy template for Copilot risky use.
- A Conditional Access policy explicitly scoping the Microsoft 365 Copilot cloud app, as recommended by the ASD Blueprint for Secure Cloud (current revision).
The Purview AI Hub layout
The Purview AI Hub is the entry point. Microsoft Learn documents the layout in the DSPM for AI article. Four panels matter on day one:
- Overview. Sensitive interactions trend, unsafe prompt rate trend, top sites grounding Copilot answers.
- Recommendations. Microsoft-generated remediation suggestions based on the tenant signal mix.
- Activity Explorer. A queryable surface for individual Copilot interactions, filterable by user, site, and sensitivity-information type.
- Policies. The alert templates and active policies the tenant has enabled.
The Overview is where the weekly cadence lives. Activity Explorer is where individual investigations happen. Policies is where the alert configuration is governed.
Three day-one alerts every Australian rollout should enable
Microsoft Learn documents a catalogue of alert templates. Three of them are non-negotiable for an Australian governance team:
1. Unsafe prompts containing personal information
Configure the alert template that triggers when a Copilot prompt contains or returns content matching the sensitive-information types associated with the Privacy Act 1988. Microsoft Learn lists the Australian-relevant SIT templates: Australian Bank Account Number, Australian Business Number, Australian Company Number, Australian Driver Licence Number, Australian Medical Account Number, Australian Passport Number, Australian Tax File Number. Australian Privacy Principle 11 reasonableness considerations (OAIC — Australian Privacy Principles guidelines, accessed 19 October 2026) specifically reference the consequences of disclosure, so an alert on these types is the minimum baseline.
2. Unlabelled-content grounding on labelled sites
Configure the alert that triggers when a Copilot answer is grounded on a document that lacks a sensitivity label but resides in a site whose policy expects labelling. This is the early-warning signal that label coverage on a protected site is drifting. Microsoft Learn documents this template under the Data Security Posture Management for AI policy library.
3. Risky-user correlations
Enable the correlation between Insider Risk Management signals and Copilot use. This alert surfaces accounts whose broader behaviour pattern correlates with risk indicators and who are also actively prompting Copilot. The Office of the Australian Information Commissioner's guide to securing personal information (OAIC, current edition) treats user-behaviour signals as part of the reasonable-steps test under APP 11.
Three alerts is the floor. Microsoft Learn lists additional templates that mature governance programmes layer in over time, including over-permissioned access detection and AI-application risk scoring.
Weekly cadence: four metrics in fifteen minutes
The weekly Copilot governance review does not need to be a long meeting. Four metrics, fifteen minutes, recorded against the prior week:
- Sensitive interactions count. Trending down is the goal. A flat trend indicates labelling debt is unresolved. A trending-up signal correlates with a Copilot cohort expansion or a new pattern of use that warrants investigation.
- Unsafe prompt rate. A meaningful denominator only emerges after the cohort exceeds approximately two hundred and fifty users per Microsoft Learn guidance. Below that, focus on individual incidents in Activity Explorer rather than rate trends.
- Unlabelled-content grounding rate. The most actionable rate. A non-zero rate on a site listed under a labelling policy is a specific remediation ticket.
- Risky-user correlations. Investigate every correlation. A correlation between Insider Risk Management signals and Copilot use is the kind of finding the OAIC expects organisations to treat as a deliberate event rather than background noise.
Monthly cadence: re-review sites with Restricted Content Discovery
Once a month, the SharePoint and Teams Governance lead should re-review the sites with Restricted Content Discovery applied against the DSPM for AI signal mix. The questions to answer:
- Which sites generated the most sensitive interactions? Do they still need the restriction, or should their access be remediated?
- Which sites generated unlabelled-content grounding alerts? What is the remediation status?
- Are any sites ready to have Restricted Content Discovery removed because their SAM Data Access Governance posture has improved?
- Are any sites now showing risky-user correlations that justify applying Restricted Content Discovery or Restricted Access Control?
The monthly cadence aligns with the ASD Blueprint for Secure Cloud expectation that scope decisions are evidence-based and that the evidence is documented (ASD, current revision).
Where Activity Explorer matters
Activity Explorer is the investigation surface. When a weekly metric flags a concern, Activity Explorer translates the aggregate into individual events. Microsoft Learn documents the filterable fields: user, site, time, sensitivity-information type, prompt category. For Australian Privacy Act investigations the prompt category and user fields are most important. APP 11 expectations of reasonable steps include investigating events of concern; Activity Explorer is where that investigation happens. The Office of the Australian Information Commissioner's recent statements on AI-mediated processing reinforce that organisations are expected to be able to answer "what did the AI do with that data" questions in writing.
Common DSPM for AI configuration mistakes
Four mistakes recur:
- Enabling the AI Hub without publishing a sensitivity-label policy first. Without labels, the unsafe prompt rate and unlabelled-content grounding rate cannot be calculated meaningfully.
- Treating the Microsoft recommendations as optional. The recommendations panel surfaces tenant-specific guidance. Implementing a recommendation often closes a class of alerts entirely.
- Running the AI Hub against the full Copilot user base from day one. A staged enablement aligned to the Copilot pilot cohort yields cleaner baseline data and a faster signal-to-noise ratio.
- Not pairing DSPM for AI with the SharePoint and Teams Governance Audit or SAM. DSPM for AI measures the symptom. SAM identifies the cause. The audit translates both into a prioritised list of sites to remediate.
How DSPM for AI rolls back into site restrictions
The point of the measurement layer is to inform the scope layer. The walkthrough closes the loop:
- Weekly metric review yields a list of candidate sites to investigate.
- Monthly review of Restricted Content Discovery sites applies the SharePoint and Teams Governance Audit composite score to those candidates.
- SAM Data Access Governance reports provide the remediation evidence.
- Updates to Restricted Content Discovery and Restricted Access Control ratify the changes, and the restriction is removed once access is fixed.
That four-stage loop is what makes Microsoft Purview DSPM for AI more than a dashboard. For Australian organisations, it is the evidence trail the Privacy Act, the ASD Blueprint, and the ACSC Essential Eight Maturity Model (Australian Cyber Security Centre — Essential Eight Maturity Model, accessed 19 October 2026) all expect to see when AI applications are interacting with organisational content.
Practitioner takeaway
DSPM for AI is the measurement layer Australian Copilot rollouts have been missing. With it the governance team can answer the question the Office of the Australian Information Commissioner is most likely to ask: how do you know your AI is not exposing personal information? Without it, every Copilot rollout in Australia is operating on faith. With the right alerts, the right cadence, and the right loop back into Restricted Content Discovery and SharePoint Advanced Management, faith becomes evidence.
References
- Microsoft Learn — "Microsoft Purview Data Security Posture Management for AI" (general availability mid-2026)
- Microsoft Learn — "Restricted Content Discovery" (current revision)
- Microsoft Learn — "SharePoint Advanced Management — Data Access Governance reports" (current revision)
- Microsoft Learn — "Sensitive information types: Australia" (current revision)
- Australian Signals Directorate — "Blueprint for Secure Cloud" (current revision)
- Australian Cyber Security Centre — "Essential Eight Maturity Model" (current revision)
- Office of the Australian Information Commissioner — "Guide to securing personal information" (current edition)
- Office of the Australian Information Commissioner — "Australian Privacy Principles guidelines, APP 11" (current edition)
Sources referenced in this article: Microsoft Learn (Restricted Content Discovery), OAIC (Privacy Act 1988), OAIC (Australian Privacy Principles guidelines, APP 11), Australian Cyber Security Centre (Essential Eight Maturity Model), Microsoft Learn (Microsoft Purview DSPM for AI), Microsoft Learn (Sensitive information types: Australia), Microsoft Learn (SharePoint Advanced Management Data Access Governance), ASD Blueprint for Secure Cloud, OAIC (Guide to securing personal information).
