---
title: How do you govern computer use in Microsoft Copilot Studio agents?
url: https://opencontroller.lyzr.com/blog/copilot-studio-computer-use-governance
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How do you govern computer use in Microsoft Copilot Studio agents?

**Govern Copilot Studio computer use with five controls: run it on dedicated, isolated machines; turn on access control so the agent can only interact with listed websites and apps; use end-user credentials rather than maker credentials for shared agents; store secrets in Azure Key Vault; and name a human reviewer for suspicious instructions. By default, computer use can operate on any website or application, so none of this is on until you set it.**

Computer use is the Copilot Studio tool that lets an agent operate a user interface the way a person would: clicking, typing and navigating in web and desktop apps. It is billed per step and needs generative orchestration turned on. Everything below comes from Microsoft Learn.

## Where does a computer use agent actually run?

The agent runs on a target machine you pick under **Machine** in the tool settings. Machines are managed through Power Automate (**Manage machines**, **See machine details**).

Microsoft's guidance is direct:

- Use **dedicated, isolated machines** that do nothing except computer use.
- Run the agent under a **least-privilege** user account.
- Limit which desktop apps can run with Windows **application control**.

A machine that is also someone's daily workstation turns every prompt injection into a risk to that person's files, sessions and mailbox. Treat computer use machines like RPA bots: separate, hardened and inventoried.

## How do you limit what sites and apps the agent can touch?

Turn on **Access control** in the tool. You list:

- **Websites** by domain, for example `contoso.com` or `*.contoso.com`.
- **Desktop apps** by process or product name, for example `msedge` or `Notepad`.

Also switch on **Enforce HTTPS**, which blocks interaction with `http://` sites.

Know the limit. Microsoft says access control blocks *interaction*, not navigation. The model can still open a site that is not on the list, such as a search engine from the Edge address bar, but it cannot act on it. That still means the agent can read what loads on screen, so access control is not a substitute for network egress rules on the machine. Pair it with [egress controls](/blog/control-ai-agent-network-egress).

## Whose credentials does the agent use?

There are two modes under **Credentials to use**:

| Mode | What it means | When to use it |
|---|---|---|
| Maker-provided (default) | The agent acts with the author's access on the machine | Personal agents only |
| End user credentials | The agent acts with the signed-in user's access | Any agent shared with others |

Microsoft warns that sharing an agent built with maker credentials lets anyone who uses it act with the author's access. That is a privilege escalation path. We cover the general pattern in [how to block maker credentials in Copilot Studio agents](/blog/block-maker-credentials-in-copilot-studio-agents).

For sign-in prompts, **stored credentials** are tied to a specific website or desktop app. You can keep them in Power Platform **internal storage** (encrypted by the platform) or in **Azure Key Vault**. Key Vault needs the subscription ID, resource group and vault name, the *PowerPlatform* resource provider registered on the subscription, and vault permissions for the creator. Key Vault gives you your own access policies, rotation and logging, so prefer it for anything beyond a pilot.

## Is there human oversight?

Yes, in a limited form. You can name a **reviewer** who gets an email in Outlook when the agent detects potentially harmful instructions that could change the model's behaviour. You set a response time limit, and if no one answers before it expires, the run stops.

Two caveats from Microsoft: each run is tied to the user who started it, and a reviewer who did not start the run likely cannot see the activity. So the reviewer may be asked to decide without full context. Pick reviewers who can reach the requester quickly.

During testing, makers see a step-by-step log of reasoning and actions with a live preview, and **Stop testing** halts all actions at once. In conversational agents, reasoning messages and screenshots are shared in the chat. The Learn page does not describe a formal audit log or screenshot retention, so plan your own evidence capture.

## What does it cost, and how do you cap it?

Computer use is billed at **5 Copilot Credits per step** on standard models and **15 per step** on premium models. A step can bundle several clicks or keystrokes. Microsoft's own example: a four-step time-sheet run costs 20 credits on a standard model or 60 on a premium one. Anthropic models also need an admin to allow external models in the environment.

A looping agent burns credits fast, so set environment-level capacity limits and watch step counts per agent. See [how to stop an AI agent from looping](/blog/stop-an-ai-agent-from-looping).

## Computer use governance checklist

1. Run on dedicated machines with least-privilege accounts and application control.
2. Turn on access control and Enforce HTTPS for every agent.
3. Use end-user credentials for any shared agent.
4. Store credentials in Azure Key Vault, not internal storage, for production.
5. Name a reviewer and a short response window.
6. Add the external [real-time threat detection check](/blog/copilot-studio-runtime-protection-with-defender) to the environment.
7. Track credits per agent and alert on step-count spikes.
8. Record the agent, owner, machine and allowed sites in your inventory.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, auto-discovers agents and keys across AWS, Azure and Google Cloud, and flags risky permissions and configuration with Posture. Behavioural-deviation monitoring can alert on, quarantine or kill an agent through the connected platform. See [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can a Copilot Studio computer use agent visit sites outside its allow list?**
It can open them, but it cannot interact with them. Microsoft says access control only blocks interaction, so add network egress rules on the machine too.

**Should computer use agents use maker credentials?**
Only for personal agents. If the agent is shared, anyone using it acts with the author's access, so switch to end-user credentials.

**How much does computer use cost in Copilot Studio?**
5 Copilot Credits per step on standard models and 15 per step on premium models, where one step can include several low-level actions.

## Key facts

Microsoft Copilot Studio computer use lets agents operate web and desktop UIs on a selected machine managed through Power Automate. By default it can operate on any site or app; Access control restricts interaction to listed domains (e.g. *.contoso.com) and desktop process names, and Enforce HTTPS blocks http sites. Access control blocks interaction, not navigation. Credentials: maker-provided (default) or end user; stored credentials in internal storage or Azure Key Vault. A reviewer can be emailed when harmful instructions are detected; the run stops if no response arrives in time. Billing: 5 Copilot Credits per step (standard models), 15 per step (premium). Requires generative orchestration. Opencontroller by Lyzr provides a governed agent registry, Posture checks and rogue-agent quarantine through connected platforms; pricing starts from $250K per deployment per year.

## Related

- [How do you govern Microsoft Copilot Studio agents?](https://opencontroller.lyzr.com/blog/govern-microsoft-copilot-studio-agents.md)
- [How do you govern AI agents that control a computer or browser?](https://opencontroller.lyzr.com/blog/govern-computer-use-and-browser-use-agents.md)
- [How do you stop Copilot Studio agents from running on the maker's credentials?](https://opencontroller.lyzr.com/blog/block-maker-credentials-in-copilot-studio-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
