Computer-Use Agents Hit GA: the Complete Copilot Studio Breakdown

Microsoft shipped computer-using agents in Copilot Studio to general availability on May 13, 2026 — rolling out across all commercial Power Platform geographies and establishing Microsoft as the first major hyperscaler to reach contractual GA with computer use. The GA build adds Anthropic Claude Sonnet 4.5 and OpenAI CUA as production models, Azure Key Vault credential storage, Microsoft Purview audit logging, and human-in-the-loop routing via Outlook — a governance triad that was absent from the September 2025 preview.

The strategic significance is real. As of May 22, 2026, Microsoft is at full GA, Anthropic is in a paid-plan beta (since December 2025), and Google's Gemini Computer Use remains in public preview. That makes Copilot Studio the only computer-use platform enterprise buyers can today deploy against a production SLA with audit-compliant logging, RBAC credential isolation, and geo-broad availability. For Power Platform shops running Power Automate, Dataverse, and Microsoft 365, the path to production is now clear.

This deep dive covers everything the GA announcement left buried in docs: the exact model lineup (GA vs Experimental), the first published dollar conversion of Microsoft's credit-based pricing across realistic workload sizes, the governance triad in plain language, the app-surface limitations Microsoft buries three pages deep, a three-platform comparison matrix, the Graebel design-partner blueprint, and a use-case routing guide for enterprise automation teams. For the developer-API alternative outside of Power Platform, see our Anthropic Computer Use API guide.

Microsoft announced the general availability of computer-using agents in Copilot Studio on May 13, 2026, via the Microsoft Tech Community blog post authored by Mustapha Lazrek of the Copilot Studio team. The announcement confirmed rollout to all commercial Power Platform geographies — expanding from the prior US and select EMEA preview footprint. Sovereign clouds (GCC, GCC High, DoD) are excluded; the accurate phrasing is “all commercial Power Platform geos excluding sovereign clouds.”

The feature had originally launched in preview in September 2025 with vision and reasoning for Windows desktop applications. The GA build adds four material capabilities that were absent from the September 2025 preview: model choice across OpenAI and Anthropic vendors, Azure Key Vault credential storage, Microsoft Purview and Dataverse audit propagation (with session replay, which was added in January 2026 preview and carries forward to GA), and Windows 365 Cloud PC pool support for ephemeral execution surfaces.

Critically, computer use in Copilot Studio is only available for agents that have generative orchestration enabled. Agents using the classic orchestration mode cannot use the CUA tool — a requirement that limits GA adoption to makers who have already migrated to the newer orchestration pattern. Enterprises running large fleets of classic-orchestration agents will need a migration step before they can take advantage of computer use.

Mustapha Lazrek wrote in the GA announcement: “Computer use in Microsoft Copilot Studio is now generally available, and we're expanding availability to all commercial geographies in Microsoft Power Platform.” The post also published a direct feedback alias — computeruse-feedback@microsoft.com — which signals an active roadmap with planned governance and scale investment. For teams evaluating governance readiness before deployment, our agent governance framework covers the RBAC, audit, and compliance checklist for production agentic deployments.

One of the most important facts buried in Microsoft's GA documentation is that not all models in Copilot Studio computer use are generally available. The Microsoft Learn computer-use docs (last updated ms.date: 2026-05-21) list four models in the selection table, but only two carry a GA designation: OpenAI Computer-Using Agent (CUA) and Claude Sonnet 4.5. Claude Sonnet 4.6 and Claude Opus 4.6 are listed as “Experimental” — which means they do not carry production support commitments.

This matters for enterprise procurement. A workflow architectured on Opus 4.6 is not on a production SLA — it is on an experimental track that Microsoft can change or remove without the same customer-notice obligations that apply to GA features. Teams building long-lived automation should anchor on OpenAI CUA or Claude Sonnet 4.5 for the production path, and pilot Sonnet 4.6 or Opus 4.6 in a sandbox environment while waiting for those models to graduate to GA.

The benchmark context is worth noting. Independent benchmark aggregator Vellum reports Claude Sonnet 4.6 at 72.5% and Claude Opus 4.6 at 72.7% on OSWorld-Verified as of February 2026, with Claude Opus 4.7 reaching 78.0% on the same benchmark per a later measurement. For context, Anthropic's original Claude 3.5 Sonnet computer-use launch in October 2024 scored 14.9% (screenshot-only) on OSWorld — the category has advanced rapidly in 18 months. Benchmark numbers vary by run and methodology; see Vellum's benchmark explainer for the full methodology context.

Note on model naming: the GA computer-use stack inside Copilot Studio lists Claude Sonnet 4.5 (GA), Claude Sonnet 4.6 (Experimental), and Claude Opus 4.6 (Experimental). No Sonnet release sits between 4.6 and the separately shipped Opus 4.7 as of May 21, 2026, and Opus 4.7 is not present in the Copilot Studio computer-use model table.

Microsoft bills computer use at 5 Copilot Credits per step on the standard tier (OpenAI CUA, Claude Sonnet 4.5) and 15 Copilot Credits per step on the premium tier (Opus 4.6). Every piece of coverage published on the GA announcement stops at “5 credits per step” and does not convert to dollars. Here is the conversion math no one else has run.

The Microsoft 365 Copilot Studio pricing page lists the prepaid pack at $200/month for 25,000 credits — $0.008 per credit. Azure pay-as-you-go meters at $0.01 per credit, independently confirmed by CloudZero's 2026 Copilot Studio pricing analysis. Microsoft's own documentation provides a 4-step time-sheet example that consumes 20 standard credits (~$0.16 prepaid, ~$0.20 PAYG) or 60 premium credits (~$0.48 prepaid, ~$0.60 PAYG). That is the baseline.

The numbers that matter are not the 4-step example — they are what happens when you scale to realistic enterprise workloads. A 25-step legacy SAP GUI flow costs $1.00 prepaid on standard per run; at 1,000 runs per day, that is $1,000/day. A 50-step end-to-end relocation case costs $2.00 prepaid on standard. Run that at 100 runs per day and you are spending $200/day, $6,000/month on standard — $600/day, $18,000/month on premium. The premium tier multiplier is 3x, but the performance uplift (72.5% to 72.7% on OSWorld-Verified) is marginal at the GA model tier. The cost-benefit case for premium is only defensible if Opus 4.6's reasoning quality materially improves task completion rates on your specific workflows.

For teams evaluating whether Copilot Studio computer use replaces existing RPA, see the Agent vs Zapier TCO calculator — the same framework applies to comparing computer use against Selenium-based or Power Automate Desktop automations.

The three GA security additions — Azure Key Vault credential storage, Microsoft Purview audit propagation, and Windows 365 Cloud PC pool support — form the minimum viable governance stack for regulated enterprise deployments. None of these were available in the September 2025 preview. Each addresses a different risk vector in production computer use.

Azure Key Vault integration.GA ships with the ability to store credentials in customer-controlled Azure Key Vaults rather than in Power Platform's internal credential storage. This is significant for enterprises with strict credential hygiene requirements — it means the agent's passwords and API keys never leave the customer's own Azure tenant boundary. The alternative (internal Power Platform credential storage) remains available and works for lower-risk automation, but Key Vault is the correct choice for any agent that operates on privileged accounts or sensitive systems.

Microsoft Purview and Dataverse audit logging. Run logs from computer-use sessions are propagated to Microsoft Purview (the compliance and data-governance platform) and to Dataverse for admin review. Session replay — the ability to play back a recorded agent session — was added in the January 2026 preview and carries forward to GA. This creates a tamper-evident audit trail that regulated industries (financial services, healthcare, legal) need to satisfy internal compliance and external regulators. For the governance framework that wraps this audit capability, see our enterprise agent governance framework guide.

Windows 365 Cloud PC pool support. GA adds support for Windows 365 for Agents — a pool of ephemeral Cloud PCs that serve as the execution surface for computer-use sessions. This means the agent runs inside a managed, isolated Cloud PC environment rather than on an endpoint belonging to a specific user. Cloud PC pools are the recommended architecture for background (autonomous) agents that do not need a live user session attached. They also provide the execution-surface isolation that enterprise security teams require for agents operating on privileged systems.

Together, these three capabilities make Copilot Studio computer use the most governance-ready computer-use platform currently available at GA. Neither Anthropic's computer-use beta nor Google's Gemini Computer Use preview has an equivalent audit-to-Purview integration or a managed Cloud PC execution surface.

Microsoft's human-in-the-loop (HITL) implementation routes low-confidence steps to a designated reviewer via Outlook. The workflow pauses, sends a structured email or Adaptive Card to the reviewer, and waits for a configured response timeout. If no reply arrives before the timeout, the run stops. This is a conservative default — the agent fails safe rather than proceeding autonomously past a step where it lacks confidence.

The HITL timeout is configurable, which is important for workflows where the reviewer is in a different time zone or has a slow response SLA. Teams should model HITL timeout intervals against their business process SLAs — a 1-hour timeout on a time-sensitive financial reconciliation workflow is a different risk profile than a 24-hour timeout on a relocation case that spans days.

The allow-list implementation is more nuanced than it appears in most coverage — and the nuance is a genuine security consideration for enterprise teams. Microsoft's documentation (confirmed in the computer-use access control section) states that the allow list “prevents actions on non-allow-listed sites” but does NOT stop the agent from navigating to those sites. A non-listed site can be opened in the browser; the agent is simply blocked from taking actions there. This matters because some websites execute scripts, set cookies, or trigger authentication flows on page load — meaning the navigation itself can have side effects even without the agent taking an explicit action.

Microsoft's own recommendation is to combine allow lists with browser policy enforcement via Microsoft Intune — specifically, configuring Edge browser policies to block navigation to non-listed domains at the network level. The allow list operates at the agent action layer; Intune browser policy operates at the network layer. Using both creates defense in depth. Relying on the allow list alone is a documented gap. For the broader governance and resilience checklist for agent deployments, see our agentic workflow resilience audit.

Microsoft's computer-use implementation supports password input on websites and the following Windows application frameworks: WinForms, WPF, UWP, WinUI, and Win32. These cover most modern Windows enterprise applications built on native Windows technologies — but they leave out a meaningful portion of the enterprise app estate.

The unsupported surfaces are: Electron-based apps (VS Code, Slack, Teams desktop, Discord, and many SaaS apps built on the Electron framework), Java-based applications (common in banking, healthcare, and legacy enterprise software), Unity-based apps and games, Citrix published apps, and other virtualized or containerized execution environments. This limitation is documented in the Microsoft Learn computer-use internal storage section but is not prominently highlighted in the GA announcement.

The practical implications depend heavily on your app estate. If your legacy automation targets are browser-based SaaS tools or native Windows apps in the supported framework list, you have a clear path. If your estate includes Citrix-published applications or Java thick-clients — both common in financial services, healthcare, and federal — Copilot Studio computer use is a blocker for those specific workflows today. The correct response is not to dismiss the platform but to inventory your app surfaces and prioritize the supported subset for the initial production deployment while tracking the roadmap for Electron and Java support.

Additionally: computer use works best for autonomous (background) agents per Microsoft's own documentation. Conversational use is supported, but each user needs valid machine credentials — which creates a credential-management overhead for high-user-count deployments. For the failure patterns that affect production agent deployments, see our guide on why 88% of AI agents never reach production.

As of May 22, 2026, three major platforms offer computer use: Microsoft Copilot Studio at full GA, Anthropic Claude Computer Use in a paid-plan beta (available to Claude API customers since December 2025), and Google Gemini Computer Use in public preview (launched October 7, 2025). The matrix below scores all three on the same axes. For the developer-API perspective on Anthropic's offering, see our Anthropic Computer Use API guide; for the function-calling architecture comparison across all three vendors, see our three-way comparison of OpenAI, Anthropic, and Google function-calling.

Microsoft named Graebel — a global talent-mobility company with approximately 1,500 employees — as the GA design partner in the May 13, 2026 announcement blog post. The Graebel case is worth examining carefully both as a technical blueprint and as a signal of the use cases Microsoft is prioritizing at GA.

Graebel built a Service Order Agent that operates their proprietary Global Connect platform via its native UI — without a vendor API. The architecture is a textbook “no-API legacy app” pattern: Global Connect has a UI but lacks a modern REST or GraphQL API for automation. Before computer use, automating Global Connect would have required brittle Selenium-based selectors or manual data entry. With Copilot Studio computer use, the agent navigates the actual UI using vision and reasoning — adapting when layouts shift or workflows branch.

The Graebel implementation adds a second layer: Azure Content Understanding is used to extract structured data from unstructured incoming emails (relocation service requests, policy documents, vendor communications). The extracted data then feeds the computer-use agent, which drives the Global Connect UI to create and update service orders. The agent is designed to scale across 30+ relocation service categories.

Matt Brownlee, Chief Revenue Officer at Graebel, said in the GA blog (a vendor-supplied design-partner testimonial): “By adopting Microsoft Copilot Studio and AI agents, we've moved beyond traditional automation to a more intelligent, scalable operating model. This initiative strengthens our ability to serve clients faster and more accurately while positioning Graebel for long-term growth.”

The architectural insight for teams evaluating computer use: the most defensible use cases at GA are exactly the Graebel pattern — legacy systems with UIs but no APIs, where vision-based navigation is not just convenient but structurally necessary. Systems that have modern APIs should use those APIs instead of computer use; the additional vision-inference cost and step-level billing make computer use economically correct only when the API path is genuinely unavailable. For the broader landscape of agent automation patterns, see our guide on workflow automation across AgentKit, Make, and Zapier.

Not every automation candidate is a good fit for computer use. The step-level billing model, the generative-orchestration requirement, and the app-surface limitations all shape which workflows belong in the initial production deployment. The routing logic below is drawn from the GA capabilities and limitations documented in Microsoft Learn, not from vendor positioning.

Deploy now — high-fit use cases. Legacy web portals with no API (vendor portals, government portals, ERP web interfaces). Native Windows desktop apps built on WinForms, WPF, WinUI, or Win32 with no automation API. Browser-based SaaS tools where no webhook or API integration is available. Background (autonomous) agents on Cloud PC pools — the architecture Microsoft recommends for production deployments. These use cases have direct support in the GA feature set, clear credential storage options (Key Vault), and no structural blockers.

Pilot carefully — medium-fit use cases. Salesforce Classic UI automation (browser-based, but Salesforce has APIs that should be evaluated first). Conversational computer use with individual user sessions (credential-per-user overhead is manageable at low user counts but scales poorly). High-step-count workflows at premium model tier (Opus 4.6 Experimental) — pilot in sandbox before committing to production volume given the 3x cost multiplier and Experimental support status.

Do not deploy yet — low-fit use cases. Electron-based apps (VS Code, Slack, Teams desktop) — not supported for password input. Java thick-clients — not supported. Citrix-published apps — not supported. Any workflow where an existing API exists and works reliably — computer use is structurally more expensive than a direct API call; reserve it for the no-API case. High-frequency, low-step workflows where per-step billing exceeds the cost of the existing manual process — run the math first.

For teams building the governance layer around any of these deployments, the agent governance framework covers the RBAC, audit, and compliance checklist, and the 80% enterprise apps embed AI agents checklist covers the organizational readiness criteria for the broader Power Platform agentic story. Our AI transformation services practice helps enterprise teams build the deployment and governance architecture for production computer-use rollouts.