Microsoft Copilot Studio

Copilot Studio governance

Every agent your employees build is an AI system, with obligations and risks attached.

You roll out Copilot Studio to 3,000 employees, and any of them can build an agent. Modulos reads those agents, helps each maker register theirs and assesses it under the EU AI Act, ISO/IEC 42001 and the NIST AI RMF.

Request a demo Registration steps

Copilot Studio agent

Benefits helpdesk

L. Meier · HR production

AI system: Registered, owner People operations
EU AI Act: Transparency duty (Article 50)
Estimated loss: €20k a year

By Modulos·Reviewed 23 September 2026

Agent sprawl in Copilot Studio

A licence rollout makes every employee a possible maker. Makers in any department build agents that read SharePoint and send email, and they often share them with colleagues without a security review.

Six months later, the CISO asks
How many agents exist, who owns each one and which data can each one reach?

Legally, each agent is an AI system. The EU AI Act attaches transparency duties to agents that talk to people (Article 50), and the full high-risk duties when the use falls under Annex III. Wrong answers, leaked data and biased decisions also cost money.

Scout, the Modulos assistant, reads each environment that you connect and lists its agents. You can answer the CISO, and the makers can register their agents.

Read the connector documentation

An example Scout answer that lists the eight Copilot Studio agents in an HR environment, with the maker, knowledge and actions of each agent.

Which agents are in the HR production environment?

Scout • 23.09.2026 10:14

HR production has 8 agents. Six of them can send messages or change records in other systems, and one screens job candidates.

Agent	Maker	Knowledge	Actions
Benefits helpdesk	L. Meier	HR policy site	Outlook: send email
Candidate screening	S. Rossi	CV library	Outlook: send email
Onboarding buddy	J. Novak	Onboarding wiki	Planner: create task
Leave request helper	M. Dubois	Leave policy	Dataverse: update
Payroll questions	A. Schmid	Payroll FAQ	None
Interview scheduler	K. Brandt	None	Outlook: create event
Training finder	T. Weber	Course catalogue	None
Exit interviews	R. Costa	Exit survey library	Teams: post message

Session ID:5c9701be…fb1ddb35

Register an agent as an AI system

Registration gives each agent an accountable owner and a written purpose. Scout reads the agent and drafts the text, but it cannot save the form. The maker checks the draft, answers the scoping questions and saves, and Modulos then computes the EU AI Act classification from those answers.

1
Connect an environment
An administrator connects a Power Platform environment with a read-only app.
Benefits helpdeskConnected
Source
HR production environment in Copilot Studio
2
Start the registration
The maker opens a new AI system in Modulos, and Scout reads the agent.
Benefits helpdeskDraft
Source
HR production environment in Copilot Studio
Read by Scout
Instructions, topics, actions and knowledge
3
Review the draft
Scout drafts the text. The maker corrects it, answers the scoping questions and saves.
Benefits helpdeskRegistered
Source
HR production environment in Copilot Studio
Read by Scout
Instructions, topics, actions and knowledge
DescriptionDrafted by Scout
Answers benefits and leave questions from the HR policy library.
Intended purposeDrafted by Scout
Employee self-service for HR policy questions.
Owner
People operations
4
Assess and collect evidence
Modulos computes the classification, and the risk and evidence follow.
Benefits helpdeskAssessed
Source
HR production environment in Copilot Studio
Read by Scout
Instructions, topics, actions and knowledge
DescriptionDrafted by Scout
Answers benefits and leave questions from the HR policy library.
Intended purposeDrafted by Scout
Employee self-service for HR policy questions.
Owner
People operations
EU AI Act
Transparency duty (Article 50)
Estimated loss
€20k a year
Evidence
Scout answers, linked to their conversations

Example agents at different risk levels

A maker can build each of these agents in Copilot Studio in an afternoon. The EU AI Act class and the expected loss measure different things, so Modulos records both. The first two agents have the same AI Act duty, but the expected loss of the order agent is more than seven times higher.

Estimated annual lossExample figures for a company with about 2,000 employees. The bar is the 80% range and the dot is the median.

Transparency dutyBenefits helpdesk

€20k a year€6k to €60k

Transparency dutyOrder support agent

€150k a year€45k to €420k

High-riskCandidate screening agent

€650k a year€150k to €2.4M

Transparency duty

Benefits helpdesk

Answers employee questions about benefits and leave from the HR policy library.

Knowledge: HR policy site in SharePoint
Action: send an email through Outlook

EU AI Act: Not high-risk. Staff must know that they are talking to an AI system (Article 50).
Main loss scenario: Wrong answers on leave or benefits that end in a disputeAbout 8 disputes a year × €2,500 each

Transparency duty

Order support agent

Chats with customers on your website, looks up orders and issues refunds up to €200.

Knowledge: public help centre
Actions: look up an order, create a refund in Dynamics 365

EU AI Act: Not high-risk. Customers must know that they are talking to an AI system (Article 50).
Main loss scenario: Refunds outside policy, and order data shown to the wrong customer after a prompt injectionAbout 30 incidents a year × €5,000 each

High-risk

Candidate screening agent

Reads CVs from a SharePoint library, scores them against the job description and emails a shortlist to the recruiter.

Knowledge: CV library and job descriptions
Action: send the shortlist through Outlook

EU AI Act: High-risk under Annex III, point 4(a), because it filters job applications. Human oversight, log retention and the other duties in Article 26 apply.
Main loss scenario: A discrimination claim or a regulator finding about biased screeningAbout one case every three years × €2 million

The classifications are the Modulos reading of the EU AI Act for these example uses. Your role, your use case and the final text of each instrument decide which obligations apply.

Requirements by frameworkEU AI Act ISO/IEC 42001 NIST AI RMF GDPR

Read-only access by environment

The connection uses an app registration in Microsoft Entra ID. Your administrator adds that app to each environment as a Dataverse application user, with a security role that can only read agents and their components. You choose which environments Modulos can see.

Modulos sends only read requests to the Dataverse Web API. It never changes an agent and does not read connection credentials.
The Microsoft sign-in during setup only lists your environments. Modulos does not keep that sign-in token.
The source is off for every organisation until an administrator turns it on.
Scout handles agent instructions as untrusted content, apart from its own instructions.

Trusted by Leading Organizations

Questions about Copilot Studio governance

What is Copilot Studio governance?

Copilot Studio governance is the work of knowing which agents your organisation builds in Microsoft Copilot Studio, who is accountable for each one, which laws and standards apply, and what evidence supports each decision. Microsoft provides platform controls such as environments and data policies. A governance programme adds the inventory, the risk assessment and the record that an auditor or regulator asks for.

How does Modulos help with agent sprawl in Copilot Studio?

Microsoft’s admin tools decide who can create and share agents in each environment. Modulos adds the governance record. Scout lists the agents in the environments you connect, each maker registers their agent as an AI system with an owner and a purpose, and Modulos assesses the requirements that apply and the risk in money. You then know which agents exist, who is accountable for each one and what each one can cost when it fails.

Is a Copilot Studio agent an AI system under the EU AI Act?

Yes, in almost every case. An agent that uses a generative model to answer questions or take actions meets the definition of an AI system in Article 3(1). The obligations that attach to it depend on your role and on the use. An organisation that builds an agent and puts it into service under its own name can be its provider. A high-risk classification follows from the use case, for example recruitment.

What does Modulos read from a Copilot Studio agent?

Modulos reads the agent’s name, owner and state, and the dates when it was created, changed and published. It also reads the instructions and the model hint, the names and descriptions of the topics, each action with its connection reference and operation, and the type and URL of each knowledge source. Modulos does not read connection credentials.

Which permissions does Modulos need?

Your administrator registers an app in Microsoft Entra ID and adds it as an application user in the Power Platform environment, with a security role that can only read agents. Modulos uses the Dataverse Web API with that app and sends only read requests. Each connection covers one environment, so you decide which environments Modulos can see.

Does Modulos change our agents?

No. Modulos never writes to Copilot Studio. It reads the agent definition and its components. When Scout reads an agent, it handles the instructions and other agent text as untrusted content and keeps that text apart from its own instructions.

Does the integration replace the Power Platform admin center?

No. Keep using Microsoft’s admin tools for environments, data policies and sharing. Modulos adds the governance record for each agent: the owner, the purpose, the risk assessment, and the controls and evidence under each framework.

Is Modulos affiliated with Microsoft?

No. Modulos is an independent AI governance platform. It connects to Copilot Studio through Microsoft’s published APIs. Microsoft and Copilot Studio are trademarks of the Microsoft group of companies.

Sources and further reading

Microsoft and Copilot Studio are trademarks of the Microsoft group of companies. Modulos is not affiliated with Microsoft.

Govern the agents your employees build

Talk to an expert

Walk through your Copilot Studio environments with the team and see how an agent goes from registration to its first assessment.

Request a demo →

Keep exploring on your own

The documentation explains how to connect an environment, and the Trust Center has the material for your security review.

Read the connector documentation →Visit the Trust Center →Explore the EU AI Act requirements →