Start free trial
AI Governance

Tackling the EU AI Act: What Microsoft Purview can (and can't) do for you

Tackling the EU AI Act: What Microsoft Purview can (and can't) do for you

For many organizations, the number of agents in Microsoft 365 is growing quickly. Colleagues build agents in Copilot Studio, publish them in Teams and SharePoint, and involve IT or security only after they pressed publish.

At the same time, the rules around agents are getting stricter. In July, the European Union changed the deadlines in the EU AI Act. The headlines say the EU AI Act is delayed. Sounds like good news, right? In our conversations with customers, the reaction is usually the same:

“The deadline moved, so we can wait!”

That reasoning is understandable, but it leads to the wrong conclusion. Some deadlines moved by more than a year. One did not move at all, and it is the one that applies to chatbots and agents.

Microsoft Purview can help with a large part of the problem, but not all of it. The gaps really matter and require your attention. Let’s take a look at what changed, what Purview covers, what you still need to organize yourself, and how Rencore can help.

Business case & scenarios for the EU AI Act

One of the challenges with AI governance is making a compelling business case for the stakeholders responsible for funding it. “The EU says so” isn’t going to unlock any budget. We have been in this situation before, and that’s why we created three scenarios you can use to help persuade your stakeholders.

Fines that scale with revenue

Violating the EU AI Act can cost up to €15 million or 3% of worldwide annual turnover, whichever is higher. For most organizations, that is a financial risk. Put your own number on a slide, and the business case becomes a lot harder to ignore.

Copilot adoption depends on trust

If leadership, risk, compliance, and security teams do not trust how agents use corporate data, Copilot adoption slows down. Governance protects both compliance and your Microsoft 365 Copilot investment.

Your customers will ask before the regulator does

This is not just a regulatory concern. Public procurement guidance is already moving in the same direction: the UK Government’s procurement policy note on AI use tells commercial teams to identify and manage AI-related risks, ask suppliers to disclose AI use in tenders, and apply proportionate due diligence where AI changes the risk profile. In other words, buyers are being told to ask the question before regulators do. Organizations that can explain where AI is used, how it is governed, and what evidence they can provide will move faster through procurement. Organizations that cannot will have a harder time closing deals.

These three scenarios show that AI governance is about more than avoiding a fine. Together, they give you a stronger case for securing funding for your project.

What changed, and what did not

In July 2026, the EU adopted a new regulation that changes parts of the EU AI Act timeline. That change is now law: Regulation (EU) 2026/1744 was published in the Official Journal of the European Union on 24 July 2026 and entered into force on the third day following publication, meaning 27 July 2026. The important detail is that not every deadline moved. The high-risk AI deadline moved back, but the transparency deadline for AI systems that interact with people did not.

The easiest way to see the difference is to separate the deadlines that moved from the ones that stayed in place. The table below highlights the dates that matter most for organizations deploying Microsoft 365 Copilot and agents.

DeadlineWhat has to be true
2 August 2026People must be told when they are interacting with an AI system. Deepfakes must be labeled.
2 December 2026AI-generated content must carry a marking software can detect. Only for systems that already existed before 2 August.
2 December 2027The full set of requirements for high-risk AI. Moved back from August 2026.
2 August 2028The same, for AI built into regulated products such as machinery and medical devices.

The delay reported in most headlines applies to the high-risk AI deadline in the third row. It does not apply to the transparency obligation in the first row, which has applied since 2 August.

Be aware: there is one extension inside the transparency rules, and it gets overstated constantly. Only the machine-readable marking of AI-generated content gets four extra months, and only for systems already live before 2 August. It does not delay the duty to tell people they are talking to AI, and it does not apply to anything published after 2 August. Two smaller changes are worth knowing:

  • The training requirement was softened. You used to have to ensure staff had a sufficient level of AI literacy. Now you have to take measures to support it.
  • The list of high-risk uses was not touched. Any classification work you already did still counts.

Your agents are in scope

The AI Act does not appear to use “agent” as a defined category. That has led some people to conclude that agents fall outside the AI Act. In reality, many agents can still fall within scope when they meet the definition of an AI system.

The Act defines an AI system as something operating with varying levels of autonomy, working out from its input how to produce an output. That can describe many agents, depending on how they operate. Autonomy is part of the definition, not an exception to it.

The European Commission published final guidance on the transparency rules on 20 July. That guidance makes the practical consequence clear: when an agent is designed to interact with people, it must identify itself as AI, and if there is a realistic chance it will interact with a person, the safer approach is to disclose that upfront. For example:

A Copilot Studio agent greeting that introduces itself by name but does not say it is AI

A generic greeting such as “Hi, how can I help?” may be insufficient if it does not make clear that the user is interacting with AI. One saying it is an AI assistant is a safer approach. We recommend reviewing every published agent, because you should not rely on a single admin setting to solve this across all published agents.

Provider or deployer?

This distinction determines which obligations apply to your organization, so it is important to understand it before assessing your agents. The following explains the difference between a provider and a deployer:

  • Provider: the party that develops an AI system or places it on the market under its own name. Provider obligations are the most extensive and can include technical documentation, conformity assessment, risk management, and post-market monitoring.
  • Deployer: the party that uses an AI system in its own organization. The obligations are different from provider obligations, but they still remain with the organization using the system.

As a deployer, you need to use the system as intended, put someone competent in charge of watching it, be careful what data it can reach, retain relevant logs for at least six months where those logs are under your control, and inform your employees before using a high-risk system on them. This part is manageable, but the exact obligations depend on the role, system, and use case.

Here is the part that catches organizations by surprise. If someone changes an agent substantially, your organization can become the provider. For example:

A colleague builds an agent in Copilot Studio and points it at HR data to help shortlist job applicants. There is no change request and no involvement from IT. That single decision may move your organization closer to provider-type obligations and may also place the system in a high-risk category, because recruitment and worker-management use cases are explicitly sensitive under the AI Act.

Be aware: your colleagues building agents are potentially making decisions with legal consequences, usually without knowing it. If there is no approval step before an agent is published, that is the first gap to close.

Managing AI with Microsoft Purview

The following topics are essential for managing AI under the EU AI Act:

  • Discovery and visibility
  • Controlling what the AI reads
  • Logging and retention
  • Oversight and monitoring

The following sections describe these topics with the corresponding Microsoft Purview services.

Discovery and visibility

Data Security Posture Management (DSPM) lists the AI apps and agents active in your organization, how many are high risk, and how many handled sensitive information. For example:

The AI observability page in Microsoft Purview DSPM showing cloud agents, AI apps and risky interactions

We recommend using the AI observability page when Agent 365 is in scope, because Microsoft identifies that DSPM experience as the place with additional information and support for active Agent 365 instances.

To see colleagues using ChatGPT or Gemini in a browser, you also need devices onboarded to Microsoft Purview and the Purview browser extension installed for Chrome or Firefox. Without these, your dashboard can look reassuringly quiet while colleagues use unmanaged AI apps outside your monitored estate.

Controlling what the AI reads

Sensitivity labels decide which files (documents and emails) Microsoft Copilot and agents can use. When a label encrypts a file, the employee needs the right to copy and extract content before anything is returned. For more detail, read our article on managing and monitoring your data with sensitivity labels.

Data Loss Prevention adds live controls. One control is designed for a specific attack pattern: someone emails your staff a message with hidden instructions aimed at your AI. Excluding external email from what Copilot reads shuts that down and takes minutes to configure. For example:

A DLP rule in Microsoft Purview that blocks Copilot from processing email received from people outside the organization

Be aware: these controls do not inspect a file a user uploads straight into a prompt. If a colleague downloads a confidential document and attaches it to a Copilot message, the rules protecting stored files do not apply. For more detail, read our article on managing and monitoring your data with Data Loss Prevention.

Logging and retention

For high-risk AI, deployers may need to retain relevant logs for at least six months where those logs are under their control. Copilot and supported AI agent interactions can be logged in Microsoft Purview Audit when the relevant workloads, licensing, and audit events are supported. Those records can include when the interaction happened, where it happened, and references to Microsoft 365 resources the AI used to generate the response.

Microsoft Purview Audit retention depends on licensing, workload, and record type. Audit Standard retains audit records for 180 days. Audit Premium can retain Exchange Online, SharePoint, OneDrive, and Microsoft Entra audit records for one year for licensed users, while other activities and some users may still be retained for 180 days.

If you rely on these logs for EU AI Act evidence, configure audit log retention policies before the relevant activity happens. Retention changes are not retroactive: logs already recorded keep their original expiry date.

Oversight and monitoring

The Act wants a competent human overseeing the system. That means someone needs to see what the AI is doing.

Communication Compliance can be configured to review selected prompts and responses and flag concerning ones. Microsoft provides pseudonymization controls, and organizations should verify the default and role-based configuration in their tenant. Mention that early to HR and your works council, because “we review AI conversations” sounds a lot worse than it is.

Insider Risk Management includes capabilities and policies designed to detect risky agent activity, depending on licensing and feature availability. These capabilities can help spot agents sharing files outside the organization, accessing risky sites, generating sensitive responses, or behaving differently from normal. For example:

The default Risky Agents policy for agents in Microsoft Purview Insider Risk Management and its triggering events

Where the prerequisites are met, Microsoft describes the Risky Agents policy as available by default and ready to generate alerts based on observed agent activity. It may already be producing findings in your tenant that nobody has reviewed. It is worth checking before you plan a broader deployment.

Where Microsoft Purview stops

Let’s be fair: Purview helps quite a lot. That said, it was built to protect data, not to manage agents as things your organization owns. That is where many organizations get caught off guard.

Microsoft Purview tells you an agent opened a confidential file. It will not tell you who built that agent, whether that person still works for you, whether anyone used it since March, or how many almost identical copies exist.

This matters because the Act’s obligations apply to each AI system you deploy. If you cannot list your agents, name an owner for each one, and show when each was last reviewed, it becomes harder to demonstrate compliance for them. However well configured your policies are, that is still a governance gap. An abandoned agent that still has access to your data is exactly the kind of thing an auditor may look for.

There are four more things Purview alone generally does not solve:

  • Deciding which agents are high-risk. Whether an agent is high-risk depends on what it is used for, not only what data it reads. Purview alone may not determine that your friendly internal chatbot moved into a high-risk use case when someone connected it to performance review data.
  • Spotting the moment you may become a provider. Your tooling may not alert you when a colleague’s configuration choice moves your organization closer to provider-type obligations.
  • Checking that agents introduce themselves. Purview policies generally do not replace a publishing review that checks whether a published agent tells people it is AI. That needs a review step before publishing, not a report afterwards.
  • The human obligations. Training colleagues, assessing impact on people’s rights, and registering high-risk systems with the EU are all organizational work.

The Agent 365 registry listing all agents in the organization, with agents at risk and agents without owners

This is why Agent 365 matters: it moves agent governance from scattered admin views toward a central inventory and lifecycle model, but organizations still need clear ownership, review processes, and governance controls around it.

For additional information, read our articles about managing and securing AI agents with Microsoft Agent 365 and Purview, Copilot and agent governance for Microsoft 365, and what IT leaders need to know about Microsoft AI agent governance.

Adoption & education

The main adoption challenge is that agent creation no longer happens only inside IT departments. Business users in marketing, HR, finance, and other teams can now build and publish agents with relatively little technical effort. That makes education essential, because the people creating agents are also creating regulatory exposure, often without realizing it.

We recommend the following:

  • Give makers a one-page standard covering disclosure wording, approved data sources, and who to contact before publishing.
  • Add an approval step so someone reviews the agent before it goes live.
  • Run short training sessions instead of only sending a policy document.

Microsoft does offer useful training and guidance. The most relevant materials are:

A note from Rencore

Every gap in the section above starts from the same missing piece: a list of your agents, with an owner, a purpose and a review date against each one. Purview tells you what an agent did with data. It does not hand you that list.

That is where Rencore fits, next to Microsoft Purview rather than in front of it. The AI & Agents module discovers every Copilot Studio agent, SharePoint agent, declarative agent and Azure AI Foundry deployment across the tenant, whether or not anyone registered it. Each agent gets a named owner, a lifecycle status and a risk score based on audience exposure and data sensitivity, and policy checks flag the patterns this article warns about: unauthenticated agents touching SharePoint, agents shared with hundreds of users, agents nobody has reviewed in months. High-risk agents can be routed for review or unpublished, with the trail to prove it.

Be clear about what this does not solve. It will not decide your provider or deployer role, write your disclosure wording, or train your makers. Those stay organizational work, exactly as this article argues. What it does is give you the inventory the EU AI Act assumes you already keep, and an answer to “who owns this agent, and is it still needed” that does not depend on someone remembering.

Conclusion

The deadline for the full set of high-risk AI requirements moved to December 2027. That gives your organization more time to design the right governance model for AI and agents.

But the transparency deadline did not move, and it may apply to agents your colleagues have already published. Telling people they are talking to AI should be handled in the agent experience and design, not treated as something Purview alone will solve. It is a design choice made in Copilot Studio by whoever built the agent, often months ago.

Microsoft Purview gives you strong data controls, useful logging, and monitoring. What it does not give you on its own is a complete governance inventory of your agents, including owners, purposes, and review dates. That is the kind of inventory you may need to demonstrate governance and accountability under the EU AI Act.

That distinction is the key takeaway. Purview helps you prove that data controls, logs, and monitoring are in place. Agent governance is different: it is about ownership, purpose, review, and operational accountability. The table below summarizes that boundary.

Purview handles thisYou need something else for this
What data an agent usedWho built it, who owns it, whether it is still needed
Whether sensitive data appeared in a promptWhether the agent’s purpose makes it high-risk
Logs and how long they are keptA list of your agents with owners and review dates
Blocking content the AI should not readChecking each agent tells people it is AI
Evidence your settings are configuredStaff training, impact assessments, registration

Ready for the next step? Read how Microsoft Purview Compliance Manager turns these controls into evidence you can share with an auditor.

Common questions on this

The EU AI Act was delayed. Can we wait before we act on our agents?
No. Regulation (EU) 2026/1744 entered into force on 27 July 2026 and moved the full set of high-risk AI requirements back to 2 December 2027, and to 2 August 2028 for AI built into regulated products. The transparency obligation did not move. Since 2 August 2026, people must be told when they are interacting with an AI system, and that applies to the chatbots and agents your colleagues have already published. Only the machine-readable marking of AI-generated content got four extra months, and only for systems that were live before 2 August.
Are Copilot Studio agents in scope of the EU AI Act?
Many are. The Act does not treat "agent" as its own category, but it defines an AI system as something that operates with varying levels of autonomy and works out from its input how to produce an output. That describes a lot of agents, and autonomy is part of the definition rather than an exception to it. The European Commission's transparency guidance makes the practical consequence clear: an agent designed to interact with people must identify itself as AI, so a greeting such as "Hi, how can I help?" is not enough on its own.
Can one of our own agents turn us from a deployer into a provider?
It can. A deployer uses an AI system in its own organization; a provider develops one or places it on the market under its own name, and carries the heaviest obligations, including technical documentation, conformity assessment and post-market monitoring. If someone changes an agent substantially, your organization can move toward provider obligations. A Copilot Studio agent pointed at HR data to shortlist job applicants is the classic case, because recruitment is also an explicitly sensitive use under the Act. If there is no approval step before an agent is published, that is the first gap to close.
Does Microsoft Purview cover the EU AI Act for our agents?
It covers the data side well. Data Security Posture Management shows which AI apps and agents are active, sensitivity labels and DLP control what Copilot can read, Purview Audit logs the interactions, and Communication Compliance and Insider Risk Management help with oversight. What Purview does not do is manage agents as things your organization owns. It will not tell you who built an agent, whether that person still works for you, whether the agent is still used, whether its purpose makes it high-risk, or whether it tells people it is AI.
How do we build the agent inventory the EU AI Act expects?
Start from a list you did not have to compile by hand, because agents are published by business users without a change request. Rencore discovers every Copilot Studio agent, SharePoint agent, declarative agent and Azure AI Foundry deployment in the tenant, whether or not anyone registered it, and gives each one a named owner, a lifecycle status and a risk score based on audience exposure and data sensitivity. That runs next to Purview rather than instead of it: Purview protects the data, the inventory answers who owns each agent, what it is for and when it was last reviewed.

Last updated 17 November 2026

Related articles

Rencore newsletter

Subscribe to our newsletter

Get the latest insights on governing Microsoft 365, Copilot and AI agents, delivered to your inbox.

Loading form