Microsoft brings computer use to Copilot Studio agents

If you have ever had to automate tasks in a clunky enterprise system without modern APIs, you know how painful traditional robotic process automation can be. Move one button two pixels to the left, and the entire workflow instantly shatters.

Microsoft wants to change that. The company just introduced a computer use capability inside Copilot Studio, rolling it out as an early access research preview for businesses.

Basically, Microsoft is teaching its enterprise AI agents to look at screens, click buttons, and type into fields just like a human clerk would.

Giving AI hands and eyes on the desktop

This announcement builds on updates from last month, when Microsoft added deep reasoning capabilities, Model Context Protocol support, and general availability for agent flows.

With computer use, agents treat web and desktop applications as accessible tools. If a human can interact with a software interface, the agent can too, even when there is no API connecting the underlying systems.

It works across standard desktop programs and major web browsers like Chrome, Firefox, and Edge. Instead of relying on rigid, pre-programmed UI coordinates, the agent uses built-in reasoning to recognize screen elements and adapt in real time if a layout shifts.

Makers can build these workflows using natural language prompts without writing code. Microsoft added a split-screen testing environment that shows real-time video of the planned automation right next to the agent's chain-of-thought reasoning.

Every action is logged with captured screenshots and reasoning steps so admins can audit what the agent actually did. The whole thing runs on Microsoft-hosted cloud infrastructure, with enterprise data staying inside compliance boundaries and excluded from frontier model training.

Microsoft plans to share more technical details at its Build conference in May.

Why it matters

We think this is Microsoft making a deliberate, aggressive move directly into traditional RPA territory.

For enterprise users, the immediate upside is huge for legacy workflows. Tasks like invoice data entry, market research scraping, and cross-system database updates usually require expensive custom integrations or fragile RPA scripts. Letting an agent visually parse the screen bypasses the need for an API entirely.

That said, early access previews always come with caveats. While autonomous screen navigation sounds great on paper, visual agents can still misclick or misinterpret dynamic menus. In high-stakes areas like financial ledgers, nobody is going to let this run entirely unattended just yet.

Where Microsoft is being clever is distribution and enterprise trust. Anthropic kicked off the computer use wave late last year, but mostly aimed it at developers willing to spin up their own virtual machines. Microsoft is wrapping the capability directly into Copilot Studio, hosting the compute, and baking in compliance guarantees right out of the box. That puts immense pressure on legacy automation vendors like UiPath to prove why customers still need dedicated, brittle RPA platforms.