Computer Use (CUA) in Copilot Studio 🖥️

Computer use lets a Copilot Studio agent operate a Windows computer the way a person does: clicking buttons, choosing menus and typing into fields, driven by a vision plus reasoning AI model. If a human can use the app, the agent can too, even when there’s no API to connect to.

TL;DR: Computer use is a tool in Microsoft Copilot Studio that automates websites and desktop apps from natural-language instructions. It’s powered by Computer-Using Agents (CUA), an AI model that reads the screen visually and adapts when the UI changes. It’s currently in preview, needs an agent with generative orchestration turned on, and complements rather than replaces classic RPA.

⚠️ Preview. Computer use is a Copilot Studio preview feature (built on the standard harness). Preview features aren’t meant for production use and may change, so confirm current availability, limits and pricing before you design around it.

What it is

🔸 You describe the task in plain language, for example “open this PDF and use its contents to fill out the web form”, and the tool performs the necessary mouse clicks and keystrokes using a virtual mouse and keyboard.

🔸 It’s powered by Computer-Using Agents (CUA), an AI model that combines vision (reading the screen as pixels) with advanced reasoning to interact with graphical user interfaces.

🔸 Because it sees the screen, it adapts to interface changes. When buttons or layouts move, it keeps working instead of breaking like a hard-coded selector would.

🔸 No code required, and no API needed on the target app. It’s suited to legacy applications, desktop-only workflows and multi-app processes that lack modern integrations.

How it works: the perception, reasoning, action loop

Computer use runs an iterative three-step loop until the task is done or it needs your input.

flowchart LR
    A["Perception<br/>capture a screenshot"] --> B["Reasoning<br/>chain-of-thought over screen + history"]
    B --> C["Action<br/>click, type or scroll"]
    C -->|"task not done"| A
    C -->|"task done"| D(["Finished"])

🔸 Perception: it captures screenshots to visually understand the current state of the screen.

🔸 Reasoning: it evaluates the current state, previous actions and screenshots through a chain-of-thought approach to decide the next step.

🔸 Action: it clicks, types or scrolls, then loops back to perception with a fresh screenshot.

CUA vs classic RPA

Copilot Studio’s computer use (CUA) and Power Automate’s desktop flows (RPA) both automate UIs, but they work very differently.

Aspect	RPA	CUA
Type	Rule based	LLM driven
Interaction	UI tree / selectors	Vision (reads screen)
Authoring	Script	Natural language
Decisions	Predefined rules	Autonomous, visual
Flexibility	Limited	High
Error handling	Static	Self-correcting

Use RPA when only GA features are allowed, the UI is stable, the rules are clear, speed and high volume matter, and an RPA team already owns it.

Use CUA when UIs shift or vary widely across many apps, you need something fast without a full RPA build, the task depends on what’s visible on screen, and decisions are fuzzy enough that the agent has to reason and self-correct.

Local vs standalone computer use tool

🔸 The local, in-agent computer use tool is hosted by, and tied to, a single agent.

🔸 The standalone computer use tool (also preview) is a modular version: publish it once, then reuse it across multiple agents and agent flows. It centralizes governance (define allowed apps and websites), gives high-fidelity session replays and activity logs, and supports human-in-the-loop supervision for sensitive tasks.

Governance and safety

🔸 Computer use is powerful and comes with real security considerations. An ambiguous instruction or unexpected on-screen content could trigger an unintended action that touches personal, financial or enterprise systems.

🔸 Mitigate this by defining allowed applications and websites, reviewing session replays and activity logs, and adding human-in-the-loop approval for sensitive or complex steps.

Licensing

🔸 Computer use is a Copilot Studio capability, so it’s metered through Copilot Studio consumption (subscription or pay-as-you-go Copilot Credits) rather than a Power Automate RPA license. As a preview feature its exact metering can change, so confirm current pricing and capacity terms before you rely on it.

🔸 It requires an agent with generative orchestration enabled.

🛠️ FAQ

Is computer use the same as RPA? No. RPA (desktop flows) is rule-based and drives the UI tree; computer use is AI-driven (CUA) and reads the screen visually, so it adapts when the UI changes. They’re complementary.

Do I need an API on the target app? No, that’s the point. Computer use works purely through the visible UI, so it can automate legacy or desktop-only apps that have no API.

Is it production-ready? It’s currently preview and built on the standard harness. Preview features may change and aren’t meant for production use, so validate before you depend on it.

Where does it run? On a Windows computer you set up, driven by a virtual mouse and keyboard, orchestrated by your Copilot Studio agent.

How is it different from Azure AI Foundry’s computer-use model? Copilot Studio’s computer use is a low-code tool for makers; Azure AI Foundry exposes a computer-use-preview model for developers who build the screenshot-to-action loop in code. Same idea, different audience.

🔗 Related

🔸 Cloud Flows vs Desktop Flows, where classic RPA fits.

🔸 M365 Copilot vs Copilot Studio, which Copilot builds agents that take action.

🔸 All AI & Copilot tips · Knowledge base overview

Details reflect Microsoft Learn at time of writing. Computer use is a preview feature and evolves quickly, so always confirm current capabilities and licensing in the Microsoft Copilot Studio documentation.