Microsoft Copilot Studio agents can now control your computer

The most painful enterprise workflows rarely involve cutting-edge innovation. More often, they look like an employee manually copying numbers from a supplier portal and pasting them into a clunky desktop ERP that hasn't had an API update in a decade.

Traditional RPA tried fixing this years ago, but the moment a website updated a button color or moved a layout element, the entire brittle script would crash. Microsoft is taking another crack at the problem, this time handing the steering wheel over to multimodal AI.

Microsoft announced that computer use is now available in public preview inside Copilot Studio for customers in U.S.-based environments.

Instead of writing complex automation scripts, users can simply describe what they want done in plain language. The agent then operates a computer to perform tasks across websites and desktop applications, relying on real-time reasoning to adapt when interfaces change.

Giving agents their own keyboard and mouse

Microsoft rolled out several enterprise-ready enhancements alongside the preview. The biggest friction with running automated agents has always been environment setup, so Microsoft integrated hosted browsers powered by Windows 365. For web tasks, users do not need to configure virtual machines at all.

If you need to automate custom desktop software or internal enterprise portals, you can still register your own machines. Existing setups used for Microsoft Power Automate desktop flows are also fully compatible, letting companies layer adaptive reasoning directly on top of legacy hardware.

To make adoption faster, Microsoft added starter templates for common scenarios like web market research, cross-site inventory tracking, and data entry into forms without APIs. Builders can test their workflows using a side-by-side view that shows both the live computer screen and the agent's step-by-step reasoning chain.

On the security side, agents can now pull from securely stored credentials to sign into services. Microsoft also built in an allow-list feature that restricts agents to specific apps and domains. The moment an agent tries to navigate outside that approved list, the task stops immediately.

Why it matters

We think this is one of Microsoft's smartest enterprise product packaging moves this year. While Anthropic grabbed headlines by showing that frontier models could operate desktop computers, it remained largely a developer experiment. Microsoft took that concept, wrapped it in Windows 365, integrated it with Power Automate, and bolted on enterprise credential governance.

For businesses, this opens up massive practical value. If you have teams wasting hours manually extracting data from supplier dashboards or keying data across air-gapped systems, testing this on low-risk workflows is a no-brainer.

That said, handing full keyboard and mouse control to an LLM is inherently risky. Even with domain allow-lists, letting an AI navigate real application interfaces means accidental clicks and unintended form submissions remain real possibilities. It looks great for read-heavy tasks and data aggregation right now, but we wouldn't let it touch mission-critical write operations without human approval in the loop.