Artificial Intelligence
AI governance: rules and evidence for AI in your company.
Together we set which agents run, which data Copilot reaches and what staff may enter into which model. Purview keeps prompts and responses for every review.
For management and boards whose teams use Copilot, agents and other AI tools and who need rules and evidence for them.
- Microsoft Copilot
- Microsoft Agent 365
- Entra Agent ID
- Microsoft Purview
- Data Security Posture Management
- Copilot Studio
- Microsoft Foundry

A question from the board
The board wants to know which AI in the company processes customer data. The head of IT opens the agent list in the Microsoft 365 admin center: every agent with its owner, data sources and next review date, and two agents without owners are blocked. The data protection adviser puts the Purview report next to it, showing which prompts contained personal data and which rule applied.
From agent inventory to legal frame
Most companies start with pilots that work: Copilot in the management team, an agent in customer service, a team that translates with a model. With every further use the question grows of who approved what and how that can be shown. We set up controls and evidence and agree them with your legal or data protection team, which decides on the legal assessment. How an agent is built is described under Agentic AI.
- Agent inventory and lifecycleThe registry in the Microsoft 365 admin center lists agents from Copilot Studio, from SharePoint and from partners; agents from Microsoft Foundry and other platforms can be managed as their own identity with Entra Agent ID. For each agent we record owner, data sources, actions and a review date, hand agents over when their owner leaves and block those no longer needed.Result: a list of all agents in which each has an owner and a next review date.
- Data in Copilot and agentsCopilot and agents answer with the permissions of the person asking and respect sensitivity labels; content without the copy right is not summarised. A data loss prevention rule keeps files labelled as personal data out of the answers, and Data Security Posture Management in Purview shows which AI apps and agents touch confidential data, ChatGPT Enterprise included.Result: your data protection team sees where AI works with confidential data, and excluded content stays out of the answers.
- Logging, retention and disclosureThe audit log records who used Copilot or an agent and when; eDiscovery finds the prompts and responses themselves. A separate retention policy sets how long Copilot interactions are kept, apart from Teams chats. For Copilot in Teams meetings, both require a transcript.Result: an access request or an internal investigation includes the interactions with AI.
- Rules for staffAn AI policy with approved tools, a table of which sensitivity level may go into which model, and rules on when AI content is labelled, for example when customers write to an agent. When a rule stops a prompt in Copilot, the notice leads straight to your policy. Rollout & Adoption plans the training for it.Result: staff know what they may enter where and find the answer in the notice.
- Legal frameAccording to the FDPIC, the Swiss data protection act applies directly to AI, with duties on transparency, on human review of automated individual decisions and on an impact assessment where the risk is high. Anyone offering AI in the EU or having its output used there can fall under the EU AI Act. We map every duty to a setting and a piece of evidence; your legal or data protection team makes the legal assessment.Result: a table from requirement to setting that your data protection team countersigns.
- Risk per use caseFor every use we record which data it uses, which actions it triggers, where a person approves and what is logged. Whether an agent triggers a payment, sends an email to a customer or pre-sorts a job application is set on the sheet: on its own, with approval by a person or not at all. We check attack paths against MITRE ATLAS, the open knowledge base of attacks on AI systems, such as poisoned agent tools or hidden instructions in documents, and record the countermeasure for each case.Result: one risk sheet per use case, approved jointly by the business team and data protection.
What you decide
A member of management who stands up for AI governance, and your data protection or legal leads, who decide on the assessment. A first list of the AI tools and agents your teams work with today. Your existing sensitivity labels; where they are missing, the work starts with Compliance & Data Protection. A licence check, because the registry's risk signals require Agent 365 and some data loss prevention rules require higher plans.
Evidence you can present
An agent list with owner, data sources, actions and review date for every agent. Data loss prevention rules and retention policies for Copilot and agents, active and tested with your examples. A published AI policy with the list of approved tools. One risk sheet per use case and a table that maps every duty from the Swiss data protection act and, where it applies, the EU AI Act to a setting. Support & Operations can take over the regular review of the agent list.
Sound familiar? Tell us where you stand today.
ContactProduct and legal updates.
As of: 11 October 2026
- As of 21 September 2026
The agent registry shows agents without owners.
In the Microsoft 365 admin center, the registry lists all agents available to your organisation: from Microsoft, from partners, published by your organisation and shared by individual people. A tile counts the agents without owners, for example after the person who created them has left; these can be blocked or deleted. Risk signals from Purview, Entra and Defender require a Microsoft 365 E7 or Agent 365 licence.
Source: Microsoft Learn: Agent Registry in Microsoft 365 admin center - As of 5 October 2026
Data loss prevention rules apply in Copilot.
A DLP policy for the Microsoft Copilot and Copilot Chat location keeps files and emails with selected sensitivity labels out of the answers and blocks web search when a prompt contains sensitive information. Blocking prompts with sensitive information entirely and excluding emails from external senders are available in preview. Files that someone uploads directly into the prompt are not checked by the rule.
Source: Microsoft Learn: Microsoft Purview DLP for Microsoft Copilot - As of 3 August 2026
EU AI Act: high-risk duties later, the rest applies.
The AI Act has applied since 2 August 2026; prohibitions and AI literacy since 2 February 2025, the obligations for general-purpose AI models since 2 August 2025, the transparency rules from August 2026. With the AI Omnibus, in force since 27 July 2026, the obligations for high-risk systems in areas such as employment and education apply from 2 December 2027, and for systems in products from 2 August 2028.
Source: European Commission: AI Act - As of 14 September 2026
MITRE ATLAS describes attacks on AI agents.
The open knowledge base MITRE ATLAS organises attacks on AI systems by tactics and techniques, based on real incidents and research. Release 2026.09 covers 16 tactics, 120 techniques, 88 sub-techniques, 40 mitigations and 73 case studies; since July 2026 it has added, among others, poisoned agent tools and agent-to-agent communication.
Source: MITRE ATLAS
Governance for every platform.
The same questions in three environments: who owns what, who can reach what and how you prove it.
- Governance
Microsoft 365
Ownership, lifecycle and sharing for sites and teams, Purview labels and the sharing review before Copilot.
Microsoft 365 - Governance
Power Platform
Environments, connector data policies, pipelines and an inventory of every app, flow and agent.
Power Platform - Governance
AI/SI
Agent inventory, data in Copilot, prompt logs and the legal frame for AI in Switzerland.
Where AI is built and run.
The pages on which agents are built, models chosen and staff trained.
Which tasks agents take on and how we build them.
Enterprise AIModels, platforms and data boundaries for your own AI systems.
Copilot Studio agentsWhat Copilot Studio can do and how agents are built there.
Microsoft CopilotWhere Copilot fits into the working day.
Security OperationsEvaluate security alerts and check the security baseline regularly.
Rollout & AdoptionTraining on the AI policy and contacts in each department.
Entra IDIdentities for people and agents in one place.
Agentic AI Readiness CheckCheck how clearly task, data access and ownership are defined for your next agent.
Models as insider risksSatya Nadella's principles and how we turn them into identity, logging and an emergency stop.
Where an agent already runs with clear rules.
BMC develops and runs its Copilot agent for product and supplier knowledge itself. We developed the security and governance model together.
Related to your task
Labels, data loss prevention and retention under the Swiss data protection act for the whole tenant.
Microsoft 365 GovernanceReview sharing and site ownership before Copilot answers from them.
Power Platform governanceEnvironments and data policies for everyone who builds apps, flows and agents.
Zero Trust & Identity SecuritySign-in, conditional access and admin rights for people and agents.
Support & OperationsHave the agent list and the reports on AI use reviewed on an ongoing basis.
Questions from management and data protection.
Does the Swiss data protection act apply to AI?
Yes. The FDPIC states that the current data protection act is worded in a technology-neutral way and applies directly to AI-supported data processing. From this it derives: disclose purpose, functioning and data sources, tell people when they are corresponding with a machine, have automated individual decisions reviewed by a person on request, and carry out a data protection impact assessment where the risk is high. We build the settings for this; whether they suffice is for your data protection team to assess.
Does the EU AI Act affect us?
It can, if you offer an AI system in the EU or its output is used there; whether that is the case is for your legal team to clarify. As of 11 October 2026, prohibitions and AI literacy have applied since 2 February 2025, the obligations for general-purpose AI models since 2 August 2025 and the remaining rules since 2 August 2026. Regulation (EU) 2026/1744 moves the high-risk obligations to 2 December 2027, for example for recruitment, and to 2 August 2028 for systems in products.
What is Switzerland planning?
Switzerland signed the Council of Europe convention on AI on 27 March 2025. The Federal Council intends to ratify it and has asked the Federal Department of Justice and Police to draft a consultation bill by the end of 2026, covering transparency, data protection, non-discrimination and oversight in particular; laws are to be adapted sector by sector where possible. When we checked on 11 October 2026, the consultation had not yet opened. Until new rules apply, the data protection act is the yardstick.
Do we need Microsoft Agent 365 for this?
Not to start with. Entra Agent ID is available to all Microsoft Entra customers, and the inventory in the Power Platform admin center shows Copilot Studio agents with their owners. Agent 365, generally available since 1 May 2026, is needed for Entra's security features for agents and for the risk signals in the registry. Which feature requires which licence we record before the start.
May staff use ChatGPT, Claude or Gemini?
Your AI policy decides, and we implement it technically. Approved tools with a company contract we connect to Purview, for example a ChatGPT Enterprise workspace. On managed devices, a data loss prevention rule detects when confidential data is pasted into AI websites and warns or blocks. Where a vendor processes the data we check per tool and contract.
Who decides whether an agent may act on its own?
The business team responsible for the agent, together with your data protection team, on the risk sheet for the use case. We contribute the technical side: which action needs which permission, where an approval belongs in the flow and what the log records. When an agent gets a new data source or action, the sheet comes back on the table.
Which attacks on AI do you take into account?
We work with MITRE ATLAS, an open knowledge base that describes attacks on AI systems from real incidents and research. For agents, hidden instructions in emails and documents, poisoned tools and data sources, and overly broad permissions matter most. For each use case we record which techniques are relevant and which countermeasure applies, from limited permissions to approval by a person.
Agree on a useful next step
Tell us about your current workflow, the systems involved and the result you need. In our first conversation, we identify what else we need to know to plan a useful next step.

