EU AI Act Compliance
An EU AI Act assessment for your Microsoft AI.
For any organisation the EU AI Act applies to, including those outside the EU whose AI is used in the EU. I list every Copilot, agent and model you use or build, set out what the Act asks of you for each one, and give you a dated action plan. Your legal counsel signs off the roles and risk levels.
- Microsoft 365 Copilot
- Copilot Studio
- Foundry
- Agent Framework
- Purview
- NowSince February 2025No prohibited practices, such as reading staff emotions from their voices or faces. Measures to support AI literacy for staff.
- NowSince August 2026People told when they are talking to AI. AI-generated content marked in machine-readable form, and deepfakes labelled.
- Next2 December 2026Two new bans: AI that makes sexual images of a real person without consent, or child sexual abuse material. Marking due for generative tools placed on the market before August 2026.
- Later2 December 2027High-risk rules for AI used in areas such as recruitment, credit scoring, life and health insurance pricing and public benefits.
- Later2 August 2028High-risk rules for AI built into regulated products.
Your role depends on what you build.
The Act gives duties to providers, who develop an AI system and put it into use under their own name, and to deployers, who use one in their work1. The role is set system by system, so it changes the moment you build an agent.
Microsoft is the provider. You are the deployer.Copilot Chat and Microsoft’s own agents too.
Microsoft tells users they are talking to AI2. You support AI literacy for the people who use it, avoid prohibited uses such as reading staff emotions from voices or faces, and label deepfakes and the AI text you publish on matters of public interest.
You built the agent, so treat yourself as its provider.Even when only your own staff use it.
An agent that talks to people must tell them it is AI, and the content it generates must carry a machine-readable mark where that is feasible. Copilot Studio keeps transcripts for 30 days by default, and none in Teams or developer environments3. A high-risk use needs at least six months of logs from 2 December 20274.
You built the system, so treat yourself as its provider.Microsoft supplies the platform and the SDK.
The same duties. Microsoft Agent Framework sends no telemetry until you add an exporter, leaves prompts out of its traces by default, and asks for approval only on the tools you mark5. Foundry’s evaluations, red teaming and tracing produce evidence for your technical documentation.
The model makers are the providers.GPT, Claude, Mistral and others.
Microsoft, OpenAI, Anthropic, Mistral AI and Google have signed the EU’s code of practice for general-purpose AI; xAI signed part of it, and Meta and DeepSeek have not6. Record which model each system uses and where it processes data: Claude is outside the EU Data Boundary in Copilot and Copilot Studio, and has no EU-only option on Foundry7.
Hiring, credit scoring and life and health insurance pricing are high-risk areas.So are promotion, dismissal, staff performance, exams and public benefits.
Use any of the above there and you are likely the provider of a high-risk system. From 2 December 20274, the provider carries risk management, data governance, technical documentation, logging, human oversight, a quality management system, conformity assessment and registration. The deployer keeps at least six months of logs, tells staff and the people affected and, for credit scoring, insurance pricing and public bodies, runs a fundamental rights impact assessment. Systems already in use are caught once their design changes significantly4. Microsoft’s guidance for Copilot says to avoid these uses2.
- Regulation (EU) 2024/1689, the AI Act, as amended: Articles 2 to 6, 25 to 27, 50, 111 and 113, and Annex III.
- Microsoft Learn, Microsoft 365 Copilot application card, checked November 2026.
- Microsoft Learn, Understand downloaded conversation transcripts from Power Apps and transcript controls, checked November 2026.
- Regulation (EU) 2026/1744, the Digital Omnibus on AI, amending Articles 111 and 113, in force since 27 July 2026.
- Microsoft Learn, Agent Framework observability and tool approval, checked November 2026.
- European Commission, The General-Purpose AI Code of Practice, signatories updated 7 October 2026.
- Microsoft Learn, Claude models from Anthropic in Microsoft Online Services and Claude hosting options in Foundry, checked November 2026.
What I work through with your teams.
Find what you have
- List every AI system: Microsoft 365 Copilot, agents built in Agent Builder, Copilot Studio and Foundry, AI inside your other software, and tools staff use without approval.
- Record each one’s purpose, owner, users, model and where its data is processed.
- Work out your role for each one: provider, deployer or both.
- Mark which ones the Act reaches: used by an EU office, offered in the EU, or with output used in the EU.
Fix what applies now
- Check for prohibited uses, such as reading staff emotions from their voices or faces.
- Check image, voice and video tools against the two new bans that apply from 2 December 2026.
- Make every agent that talks to people say it is AI, and confirm which of your models mark their output.
- Take measures to support AI literacy for staff and contractors, and keep a record of the training.
- Write an acceptable-use policy that keeps Copilot and agents out of high-risk decisions without an approved process.
Prepare for 2 December 2027
- Classify each system against the high-risk list. Where an Annex III use is not high-risk, record why before it goes live, and register it.
- For each high-risk system, keep logs for at least six months: a Purview retention policy, a longer Dataverse transcript job or Application Insights retention.
- Design human oversight into each high-risk system, and add approval to agent actions that move money or change records.
- Gather the documentation from Microsoft and the model makers, with written agreements where a system is high-risk.
- Where a fundamental rights impact assessment is needed, plan it alongside your data protection impact assessment.
- Appoint an EU authorised representative if you provide a high-risk system from outside the EU.
How I scope each assessment.
There is no fixed length. I agree the scope with you, from the Copilot, agents and models you run today and those you plan to roll out. Most organisations start from one of three places.
You use Microsoft 365 Copilot
Measures for AI literacy, the prohibited-use check, an acceptable-use policy, Purview audit and retention, and a decision on which AI models your staff may use.
Ask about thisYou build agents
All of the above, plus each agent’s role, its AI notice and content marking, approval on agent actions that move money or change records, and the model makers’ documentation.
Ask about thisYou use AI in decisions about people
All of the above, plus a classification note for each system and a plan for the high-risk work due by 2 December 2027: quality management, conformity assessment, six months of logs, human oversight, registration and, where needed, the impact assessment.
Ask about thisFour documents from the assessment.
- An AI register. Every system, with its owner, role and risk level.
- Role and classification notes. Written for your legal counsel to sign off.
- A gap register. What is missing against the duties that apply now and from December 2027.
- A dated action plan. Each action with an owner, a date and the Microsoft licence it needs.
This is a technical and governance assessment, not legal advice. Your legal counsel signs off the roles and risk levels. Then I can work through the plan with your teams, or hand it over.


