Microsoft just flipped the script on enterprise automation. For years, getting AI to handle the messy, real-world parts of business—clicking through vendor portals, filling forms in legacy desktop apps, or extracting data from interfaces with no APIs—meant brittle RPA scripts that broke with every UI update or waiting for IT to build custom integrations. Now, with computer-use agents in Microsoft Copilot Studio reaching general availability on May 13, 2026, enterprises can build agents that see screens, reason like humans, and act directly on any website or Windows desktop app.[1]
This isn’t another chatbot upgrade. It’s the shift from agents that talk to agents that do—navigating UIs with vision-powered reasoning, adapting to layout changes on the fly, and operating without a single API call. Microsoft is the first major hyperscaler to deliver this at production scale across all commercial Power Platform geographies (excluding sovereign clouds like GCC).[2]
If your organization still relies on manual workarounds for long-tail processes or expensive RPA maintenance, this GA release changes the economics and the possibilities. Here’s what it means, how it works, and how to get started.
Why Computer-Use Agents Matter Now
Traditional automation hits a wall with “no-API” systems: custom internal tools, third-party SaaS portals, or older line-of-business applications. RPA tools record clicks but shatter when buttons move or pages redesign. API-based agents fail entirely on these surfaces.
Computer-using agents (powered by Computer-Using Agents or CUA models) solve this by combining computer vision with advanced reasoning. They interpret screenshots, understand context, click buttons, type into fields, navigate menus, and extract data—exactly as a human would. When an interface changes, the agent adapts instead of breaking.[3]
The GA announcement emphasizes enterprise readiness: global rollout to commercial geos, built-in credential management (including Azure Key Vault), human-in-the-loop checkpoints for risky steps, detailed run history with session replay logging to Microsoft Purview and Dataverse, DLP policies, environment isolation, and model choice between OpenAI’s CUA and Anthropic’s Claude Sonnet 4.5 (with experimental options like Sonnet 4.6 and Opus 4.6).[1]
This moves computer use from experimental preview (launched September 2025 in limited regions) to a governed, auditable production capability. It’s a direct response to the reality that many valuable workflows live in systems that will never expose clean APIs.
How Computer Use Works in Copilot Studio
Adding computer use is straightforward once generative orchestration is enabled on your agent (a requirement for the tool). In Copilot Studio:
- Go to your agent’s Tools page and select Add tool → New tool → Computer use.
- Configure the required fields: a clear Name, Description (so the agent knows when to invoke it), Model selection, and detailed Instructions (natural language steps, including target URLs or app names).
- Optionally set inputs for dynamic values, choose the target Machine (dedicated Windows environments recommended), manage Connections and Credentials (maker-provided or end-user), enable Human supervision for low-confidence actions, store credentials securely, and apply Access control allow-lists.[3]
The agent then executes on a configured Windows machine (local, Azure VM, or Windows 365 Cloud PC pool in preview enhancements), using a virtual mouse and keyboard. During runs, it logs reasoning steps and screenshots. For conversational agents, users can see activity in chat; autonomous agents run in the background.
Best practices are critical for success and safety:
- Use dedicated, least-privilege machines.
- Limit web access and installed apps via allow-lists and policies (combine with Microsoft Intune for browser controls).
- Write precise, step-by-step instructions with URLs, expected outcomes, and edge-case handling.
- Test thoroughly using the built-in test experience that shows live logs and previews.[3]
Note important limitations: Password input works on most Windows frameworks (WinForms, WPF, etc.) and websites but not Electron, Java, Unity, games, Citrix, or many virtualized environments. Access control prevents actions on non-allowed sites but doesn’t stop navigation attempts—layer with Intune policies for regulated environments.[2]
Standalone computer-use tools (preview) can also be added modularly to agent flows for reuse across workflows.
Real-World Impact: The Graebel Example and Beyond
Early adopters are already seeing results. Graebel, a global talent mobility leader, processes thousands of employee relocation requests annually. Their proprietary Global Connect platform lacked API support, forcing manual handling of unstructured emails with variable instructions and attachments.
Using Copilot Studio computer-use agents, they built the Graebel Service Order Agent. It interprets incoming emails, validates against business rules, operates Global Connect directly via the UI, and escalates exceptions through workflows. The agent now handles end-to-end processing across more than 30 service categories.[4]
Outcomes include reduced manual effort, faster turnaround, improved data consistency, and a repeatable model for intelligent automation elsewhere in the business. Matt Brownlee, Graebel’s Chief Revenue Officer, noted it moves the company “beyond traditional automation to a more intelligent, scalable operating model.”
This pattern applies broadly: invoice processing in legacy ERP screens, data entry into vendor portals, compliance checks across internal web apps, or report generation from desktop tools. Agents can combine with API actions and approvals in multi-step workflows (computer-use embedding now in preview), creating hybrid automations that are both structured and adaptive.[4]
Pricing, Models, and Cost Considerations
Computer use bills at 5 Copilot Credits per step on standard GA models (OpenAI CUA and Claude Sonnet 4.5). Premium/experimental models like Claude Opus 4.6 cost more (around 15 credits/step in some analyses).
Microsoft’s prepaid packs start at roughly $200 for 25,000 credits (~$0.008/credit), making a simple 4-step form fill about $0.16 prepaid. Real enterprise workloads scale differently—a 25-step legacy flow at 1,000 daily runs can reach thousands of dollars monthly on standard tiers.[2]
Factor in machine costs (dedicated environments), potential human review overhead, and the productivity gains from reduced manual work. Start with high-ROI, lower-volume processes to validate before scaling.
Model guidance: Anchor production on GA models. Pilot experimental ones for complex reasoning where benchmarks (e.g., higher OSWorld scores on newer Claude variants) show promise, but accept the support trade-offs.[2]
Governance, Security, and Getting Production-Ready
GA brings the controls enterprises demand: credential isolation (Azure Key Vault or internal storage), audit logging with session replay to Purview/Dataverse, human-in-the-loop for exceptions, DLP policies, and environment boundaries. Agents tie to specific users or makers, with clear accountability.[1]
Still, treat this like any powerful automation:
- Follow least-privilege principles.
- Implement allow-lists + Intune/ application control policies.
- Monitor run histories closely.
- Define clear escalation paths.
- Test for hallucinations or unintended actions in edge cases.
Sovereign clouds remain excluded for now. Generative orchestration must be enabled on agents.
For teams already in the Power Platform ecosystem, integration with Dynamics 365, Microsoft 365, and Power Automate is seamless. Standalone tools and workflow embedding expand flexibility.[4]
See our guide on building secure AI agents in Copilot Studio for deeper governance frameworks.
FAQ
What exactly are computer-use agents in Copilot Studio?
They are AI agents that use vision and reasoning to interact with graphical user interfaces on websites and Windows desktop applications, performing tasks like clicking, typing, and navigating without needing APIs. Powered by specialized CUA models, they adapt to UI changes in real time.[3]
When did this become generally available, and where?
Microsoft announced general availability on May 13, 2026, rolling out to all commercial Power Platform geographies (excluding sovereign clouds). It requires generative orchestration on the agent.[1]
How much does it cost to run these agents?
Standard GA models consume 5 Copilot Credits per action/step. Costs depend on workload complexity and volume—simple tasks are inexpensive, but high-volume or multi-step processes require careful modeling against your Copilot Studio credit allocation.[2]
Are there limitations I should know about before deploying?
Yes—unsupported app types include many Java, Electron, Citrix, and virtualized environments for password handling. Access controls have nuances (navigation vs. action). Always use dedicated machines and layered security policies. Test thoroughly, especially for regulated industries.[2]
The Bottom Line
Microsoft’s computer-use agents in Copilot Studio mark a genuine inflection point: production-grade UI automation that works where APIs don’t exist and adapts where scripts fail. Combined with workflows, governance tools, and the broader Power Platform, it gives enterprises a practical path to automate the “last mile” of digital work.
The question isn’t whether this technology will matter—it already does for early movers like Graebel. The question is how quickly your organization will experiment, govern, and scale it.
What process in your business still relies on manual UI work that an agent could now handle? Share your thoughts or a specific use case in the comments—we’d love to explore it.