Almanac
Microsoft/power-automatePower Platform

Consultant KB for Microsoft Power Automate end to end: cloud flows, desktop flows and RPA, connectors and integration, AI and agent flows, process and task mining, approvals and human-in-the-loop, ALM and solutions, governance and security, licensing and capacity, monitoring and troubleshooting, plus cross-cutting decision guides. Implementation notes, configuration decisions and the gotchas that bite on real projects. Populated by the daily author agent from the Power Automate release plans, docs repo and product blog, plus the author's own consultant notes.

power-automate/02-desktop-flows-rpa/feature-computer-use.md
feature-computer-use.mdv1 · history
CurrentApplies to Desktop flows / Agent flowsUpdated last monthSource Microsoft Learn ↗

What it does

Computer use is a Copilot Studio tool that drives a Windows machine from natural-language instructions — a vision model looks at the screen and works out where to click and what to type. No captured selectors, no UI elements repository. It's the same idea as a desktop flow, aimed at the cases where building and maintaining selectors isn't worth it.

Key facts

  • It's configured in Copilot Studio, but it runs on a machine managed through Power Automate machine management, and the tool's configuration page links straight into the Power Automate machines list.
  • Only available for agents with generative orchestration turned on.
  • Model choice as of the May 2026 documentation: OpenAI's Computer-Using Agent (CUA) and Anthropic Claude Sonnet 4.5 are generally available on the standard tier; Claude Sonnet 4.6 (standard) and Claude Opus 4.6 (premium) are marked experimental. Anthropic models need an admin to allow access to external models for the environment.
  • Billing is by step through the Agent action feature: 5 Copilot Credits per step on a standard model, 15 on a premium model. A step can bundle several low-level clicks or keystrokes. Computer-Using Agent usage isn't covered by the Microsoft 365 Copilot user subscription licence.
  • Credentials for signing in to target sites and apps are stored either in Power Platform internal storage or an Azure Key Vault you provide, scoped by login domain or process name. Password fields work on websites and most Windows app frameworks, but Electron, Java, Unity, CLI and Citrix or virtualised surfaces might not be supported.
  • Access control lets you allow-list the websites and applications the model may act on. It stops interaction with anything off the list, but it doesn't stop the model opening it.
  • Human supervision routes an email to a named reviewer when the tool detects potentially harmful instructions, with a response-time limit after which the run stops.
  • Standalone computer use tools for agents and agent flows are documented separately as a preview capability.

When to use / skip

It earns its keep on long-tail processes — the ones with real business value but not enough volume to justify weeks of selector work, or applications that change layout often enough that a conventional desktop flow needs constant repair. It suits autonomous background agents far better than conversational ones. Skip it for high-volume, high-stakes, well-defined processes: a properly built desktop flow is cheaper per run, deterministic, and easier to audit. Anyone pitching computer use as a wholesale replacement for RPA hasn't done the credit arithmetic.

Configuration decisions

  • Which model tier, since the premium option triples the credit cost per step.
  • Maker-provided credentials versus end-user credentials — the former means anyone the agent is shared with acts with the maker's access on that machine.
  • Whether stored credentials sit in Power Platform internal storage or the client's own Key Vault.
  • The allow-list of sites and applications, and whether the automation is confined to a dedicated machine.
  • Who receives human-supervision requests, and whether that person can realistically judge them.

Gotchas

  • Cost scales with steps, and steps scale with how vague your instructions are. A poorly written instruction that wanders around the UI costs real money on every run.
  • Sharing an agent configured with maker-provided credentials hands the maker's access to everyone who can use it. That's stated plainly in the docs and still catches people.
  • Access control doesn't prevent navigation, only interaction. The model can still open Bing from an allow-listed browser; it just can't do anything there.
  • Human supervision only works if the reviewer is the person who initiated the run. Route it elsewhere and they'll see a request with no context they can verify.
  • Non-determinism is the point and also the risk. Two runs of the same instruction may take different paths, which makes reconciliation and audit harder than with a desktop flow.

Consultant notes

  • Model this as consumption, not licensing. Estimate steps per run and runs per month, and show the client the credit number before anyone gets excited in a demo.
  • The security guidance is worth following literally: a dedicated, least-privilege machine with a restricted application set. This is a model with a mouse on the client's network.
  • Instruction quality is the deliverable. Treat prompt writing with the same rigour as flow design, including version control and change review.
  • Position it alongside desktop flows rather than instead of them. A sensible estate has deterministic flows for the core volume and computer use for the awkward edges.

Model availability and credit rates here move quickly — verify against the current Copilot Studio billing page before quoting anything

Was this accurate?
Related
Power Automate for desktop02-desktop-flows-rpaCopilot Studio integration04-ai-and-agent-flowsPrompt actions and custom prompts04-ai-and-agent-flowsManaged Environments08-governance-and-securityPower Automate vs Azure Logic Apps11-decision-guidesAgent flows04-ai-and-agent-flows