Computer use, honestly assessed

Computer use in Copilot Studio has been generally available since 13 May 2026. Agents operate web and desktop applications through the interface, by seeing and reasoning, even where there is no API. Rolled out to all commercial Power Platform regions, excluding the sovereign clouds.

That is the announcement. Now the part that is not in it.

The exclusion list

Not supported: Electron, Java, Unity, Citrix and virtualised environments. Password entry works on websites and in Windows application frameworks, meaning WinForms, WPF, UWP, WinUI and Win32, but not in Electron, Java, Unity, games, Citrix or other virtualised environments.

Read that list again, this time with your own application landscape in mind.

Java thick clients are not a fringe case in banking and healthcare. They are frequently the exact systems whose missing API motivated the interest in computer use in the first place. Citrix in many organisations is not one application but the way all applications are delivered. And Electron covers more tools than most people realise.

That inverts the logic: computer use is weakest precisely where you would need it most. Systems with a clean, modern interface usually have an API as well. The ones without an API are often the ones on the exclusion list.

What you do first

An inventory of application surfaces, before you plan computer use as your primary automation path. Not after. Three facts per target system: which framework does it run on? Is it delivered locally or virtualised? Does the procedure require a sign-in with password entry?

If any of those answers falls into the exclusion list, the topic ends there and you have saved yourself a pilot that runs for three weeks and then gets stuck on a dialog box.

What GA added

Compared with the preview from September 2025, four things are new that matter for operations: model selection, stored credentials via Key Vault, Purview auditing with session recording, and Cloud PC pools for scale. On top of that, standalone computer use tools are in preview, letting you embed UI automation modularly and reusably into agents and agent flows.

Purview auditing is the one I switch on first. An agent operating an interface on behalf of a user produces actions that will, in case of doubt, be attributed to somebody. Without session recording, the question "what exactly did that thing do" has no answer.

My position

Computer use is a tool for the emergency, not the default answer. Automation through the interface breaks on every redesign, it is slow, it costs per step, and it creates evidentiary questions an API does not.

The order in which I check options is unchanged: API first, then connector, then MCP, then computer use. Reverse that order because the demo was impressive and you build an automation that falls over at the target system's next update.

And when computer use is the right answer, then for a bounded procedure with a clear abort condition, not for a process that spans eight screens.

Sources

• Automate web and desktop apps with computer use, Microsoft Learn
• What's new in Copilot Studio, Microsoft Learn, on GA in May 2026 and the standalone computer use tools
• Copilot Studio Computer-Use Agents: GA Deep Dive, 22 May 2026, secondary source with the breakdown of unsupported frameworks