Microsoft has moved computer-using agents in its Copilot Studio platform from preview to general availability (GA), the company announced through its Tech Community and Copilot blogs. <cite index="3-1">The feature became generally available on May 13, 2026, across all commercial Power Platform geographies, enabling enterprises to automate Windows workflows by visually interpreting and interacting with application interfaces.</cite>
What the feature does
<cite index="1-17,1-18">Computer use gives an agent the same tools a person has: a browser, a screen, a keyboard, and the ability to read what's on the page and take the next logical step. Instead of brittle selector-based automation, the computer use tool uses vision and reasoning to navigate live UIs—adapting when layouts shift, fields move, or workflows branch.</cite> The capability targets workflows that have historically resisted automation because the underlying systems expose no application programming interfaces (APIs).
<cite index="1-4,1-5">Until now, automating long-tail, UI-driven business processes meant either building and maintaining brittle robotic process automation (RPA) scripts or waiting on APIs that legacy systems were never going to expose. That gap has kept some of the most valuable workflows—the ones buried in vendor portals, internal web apps, and proprietary line-of-business systems—out of reach for modern automation.</cite>
Enterprise controls and architecture
The GA release adds production-grade governance features. <cite index="1-15,1-16">These include secure authentication with built-in credentials and Azure Key Vault when signing in to website or desktop applications, enterprise governance allowing lists for websites or desktop applications, and native Power Platform governance capabilities such as data loss prevention (DLP) policies and environment isolation. Customers can also choose between models from OpenAI and Anthropic.</cite>
<cite index="4-19">Microsoft Learn documentation lists OpenAI's Computer-Using Agent (CUA) and Anthropic's Claude Sonnet 4.5 as the GA-supported models, with Claude Sonnet 4.6 and Claude Opus 4.6 marked experimental.</cite> <cite index="3-10,3-11">The agent's screen, keyboard, and mouse streams are captured and acted upon through a session host that communicates with the Copilot Studio orchestration service, with Microsoft stating that screen data processing takes place within the customer's tenant boundary.</cite>
<cite index="5-9,5-10,5-11">Copilot Studio logs sessions to Microsoft Purview and Dataverse, with session replay carried forward from a January 2026 preview update. Human-in-the-loop checkpoints can be configured for steps that require approval or fall below a confidence threshold. DLP policies and environment isolation extend from the existing Power Platform governance stack.</cite>
Pricing, scope, and limits
<cite index="5-21,5-22">Pricing runs on Copilot Credits at five credits per step, which works out to roughly US$0.04 per step on the standard prepaid pack, making a four-step workflow cost around US$0.16.</cite> <cite index="4-28">Sovereign clouds—the Government Community Cloud (GCC), GCC High, and Department of Defense (DoD) environments—are excluded from the initial GA rollout.</cite>
Notable technical limitations remain. <cite index="4-6,4-7">Password input is supported on websites and Windows app frameworks (WinForms, WPF, UWP, WinUI, Win32) but not on Electron, Java, Unity, games, Citrix, or other virtualized environments—a production blocker for enterprises running Java thick-clients common in banking and healthcare, or Citrix-published apps.</cite> <cite index="4-8,4-9">Computer use is only available for agents that have generative orchestration enabled; agents using the classic orchestration mode cannot use the tool.</cite>
Customer deployment
Microsoft cited relocation services firm Graebel as an early customer. <cite index="1-25,1-26">The Graebel Service Order Agent operates the company's Global Connect platform directly through its UI—navigating screens, entering data, and completing transactions exactly as a trained human operator would, without APIs or platform redevelopment, while escalating exceptions and low-confidence cases through human-in-the-loop workflows.</cite>
Competitive context
<cite index="5-13,5-14">Microsoft is currently the only major platform with contractual general availability for computer-use agents; as of the announcement, Anthropic's Claude Computer Use remains in a paid-plan beta and Google's Gemini Computer Use is in public preview.</cite>