Microsoft 365 Copilot and Data Governance: Permissions Irish Employers Should Check

Before an organisation switches on Microsoft 365 Copilot, the most useful governance question is not "is the AI safe?" but "who can already see what?" Microsoft's own documentation says Copilot works within each user's existing access, which means any weakness in your permissions becomes visible the moment a person asks Copilot a question. This guide explains the Microsoft 365 Copilot data governance points that Irish employers should check, using Microsoft Learn pages and what they say, with the page dates. It suits IT, HR, finance and compliance teams in Dublin and elsewhere.

For the day-to-day use cases, see our guide to Microsoft Copilot for finance teams in Excel, Word and Outlook. This article is about the groundwork. Microsoft's documentation changes often, so check the live pages before you act, and this is general information, not legal or technical advice for your tenant.

The core principle: Copilot inherits permissions

Microsoft Learn's page "Data, Privacy, and Security for Microsoft 365 Copilot" (metadata date 9 July 2026, footer last updated 18 August 2026) says Copilot connects large language models to organisational data through Microsoft Graph, including documents, emails, calendar, chats, meetings and contacts. It says Copilot only surfaces data that the individual user has at least view permission to access, and that it is important to use Microsoft 365 permission models, such as SharePoint permissions, so the right people have the right access. That includes access granted to outside users through collaboration such as shared channels in Teams. The page says the semantic index honours the user's identity-based access boundary.

A Microsoft Learn training module on data security for Copilot (dated 24 April 2025) adds that Copilot does not search other tenants, for example one where the user is a guest.

The consequence is simple. Copilot does not create new access, but it makes existing access easy to use. A confidential HR spreadsheet that was shared too widely years ago, and that nobody ever found, can now be found by a natural-language question. That is why governance work starts with permissions.

Step one: find and fix oversharing

Microsoft's deployment blueprint, "Secure and govern Microsoft Copilot" (metadata date 6 May 2026, last updated 18 August 2026), is built on three pillars: remediate oversharing, set up guardrails and meet regulations. It recommends identifying high-risk sites and files, applying interim access restrictions if needed, fixing access issues and continuously enforcing guardrails. The page says SharePoint Advanced Management is included with the Microsoft Copilot licence.

The SharePoint Advanced Management preparation page (last updated 16 July 2026) lists tools you can use:

• Content management assessment: identifies overshared content, inactive or ownerless sites and Copilot readiness. The page recommends rerunning it every 30 days.
• Site lifecycle management: ownership, inactive-site and site attestation policies to reduce content sprawl, with archiving for inactive sites.
• Sharing settings: tighten SharePoint and OneDrive sharing and set link expiry. The page notes SharePoint defaults to the most permissive sharing setting.
• Data access governance reports: including site permissions baseline, the "Everyone except external users" report and sharing-link activity.
• Restricted access control: limit a site to members of a named group, so others cannot access the site or its contents.
• Restricted content discovery: keep a site accessible to current users but exclude it from Copilot and organisation-wide search, without changing permissions.

The last two are different tools. Restricted access control changes who can open a site. Restricted content discovery changes what Copilot and search surface, but, according to the page, leaves permissions as they are. Our suggestion: use restricted content discovery as a short-term guardrail for sensitive libraries, and treat it as a stopgap while you correct the underlying permissions.

Step two: labels, encryption and data loss controls

Microsoft's page "How data is protected and audited in Microsoft 365 and Microsoft Copilot" (metadata date 7 April 2026, last updated 18 August 2026) says Copilot works with Microsoft Purview sensitivity labels and encryption. For encrypted content, the user needs EXTRACT and VIEW usage rights for Copilot to interact with it. Responses can display sensitivity labels, reflecting the highest-priority label among the sources, and content Copilot creates from labelled sources inherits the highest-priority label where supported. The page also says information protection policies can restrict Copilot access based on sensitivity labels and data loss prevention conditions, and notes a limitation: Copilot agents cannot read files with user-defined permissions.

The privacy page adds that encryption can be applied through sensitivity labels or restricted permissions in Microsoft 365 apps, and that Copilot honours the usage rights granted to the user. In practice, an employer that has never rolled out labels should expect to do some of that work, at least for its most sensitive categories, such as HR case files, board papers, payroll and client financial data.

Step three: where data goes, and what is retained

The privacy page makes several points an Irish employer will want in its records:

• It states that prompts, responses and data accessed through Microsoft Graph are not used to train the foundation large language models used by Copilot.
• For EU users, it says EU traffic stays within the EU Data Boundary, while worldwide traffic may be sent to the EU and other countries or regions for processing. It also says models provided by Anthropic as a subprocessor are currently excluded from the EU Data Boundary, so check how that applies to your configuration.
• Copilot stores the user's prompt and response, including citations to grounding sources, as Copilot activity history. Administrators can manage this through Content search and Microsoft Purview and can set retention policies, and users can delete their own history through the My Account portal.
• When web search is used, Copilot generates a query that is sent to the Bing Search service, and administrators can control this through optional connected experiences.
• When agents are enabled, a search query based on the prompt, activity history and accessible data may be sent to an agent, and administrators control which agents are allowed.

The architecture page adds that Microsoft 365 can capture audit records for Copilot prompts, responses and referenced content, and that interaction data can be discovered, audited and retained with Purview. In our view, those audit and retention records may contain personal data, in which case GDPR is relevant; the Microsoft pages do not say this. Our guide to GDPR staff training requirements is a useful companion, and the Data Protection Commission's page on AI and large language models (18 July 2024) lists points for organisations using AI products, including considering a formal risk assessment, transparency and having a retention schedule.

Step four: tell people the rules

Technical controls do not replace a policy. A practical approach, labelled as our suggestion: write down what staff may and may not ask Copilot to do, who approves new agents, how to report a document that Copilot should not have surfaced, and who reviews the oversharing reports. Our article on what a generative AI acceptable use policy should cover sets out the topics. Brief staff that Copilot output needs checking and that the user remains responsible for what they do with it.

A short pre-launch checklist, again our suggestion:

• Run the content management assessment and review the "Everyone except external users" and sharing-link reports.
• Identify sensitive libraries (HR, finance, legal, board) and decide on permission fixes, labels or restricted content discovery.
• Decide on retention for Copilot interaction data and who can search it.
• Decide which agents and connected experiences are allowed.
• Record your data protection assessment and brief staff before go-live.

Learnsignal's offices are off Baggot Street in Dublin, and we offer in-person training from there. For current details, see our CPD pages.

Frequently asked questions

Can Copilot see files I am not allowed to open?

According to Microsoft Learn, Copilot only surfaces data the individual user has at least view permission to access. The risk is that existing permissions are wider than intended.

Is our data used to train the underlying models?

Microsoft's privacy page states that prompts, responses and data accessed through Microsoft Graph are not used to train the foundation large language models used by Copilot. Confirm the current wording in your contract and on the live page.

Does our data stay in the EU?

The page says EU traffic stays within the EU Data Boundary, with worldwide traffic possibly processed elsewhere, and that models provided by Anthropic as a subprocessor are currently excluded from the boundary. Check what applies to your tenant.

Can we exclude a sensitive site from Copilot?

The SharePoint Advanced Management page describes restricted content discovery, which excludes a site from Copilot and organisation-wide search without changing permissions, and restricted access control, which limits who can open the site at all.