Copilot Studio computer use vs Power Automate Desktop

You have automated data entry into a sales system, and report building, with Power Automate Desktop. In May 2026, computer use in Copilot Studio became generally available: describe the task in words, and AI operates websites and desktop apps. So should your existing flows move to computer use?

Both are Microsoft products, and both operate screens. The difference is whether the steps are decided once and repeated, or decided by AI on every run. This article is for people in charge of automation at companies on Microsoft 365. It sets out how the two work, the machine they run on, pricing as of October 2026, speed, what happens when a screen changes, and how to choose and combine them. How Power Automate compares with ChatGPT and Claude driving the browser is covered in Power Automate or an AI browser agent?

The short answer: Power Automate Desktop for set steps, computer use for screens that need judgment each time

• Power Automate Desktop: repeats the flow you built, the same way every time. It calls no AI while it runs, so it is fast, and the license costs the same however often it runs. It tends to stop when a screen changes
• Computer use: AI looks at the screen and decides the next action at every step. It gets past screen changes more easily, but each step costs 5 Copilot Credits (about 4 US cents) and waits for the AI to answer
• They run on the same machine: to run computer use on your own PC, you install Power Automate for desktop and register the PC as a machine in Power Automate. One machine runs one task at a time
• Computer use cannot learn a working run and repeat it: per Microsoft’s documentation as of October 2026, AI decides the steps on every run, and every step is billed
• Microsoft recommends combining them: keep RPA for fixed processes on stable screens, and give computer use only the parts that change

How they differ: are the steps decided when you build, or on every run?

In Power Automate Desktop, a person builds a flow by arranging actions, or by recording them. A run simply plays the flow in order, without calling AI. Buttons and fields are found by the clues recorded when the flow was built (UI element selectors).

In computer use, you describe the task in words. On every run, AI takes a screenshot, works out the next action from the current state and the actions so far, then clicks or types. Microsoft describes this as a loop of perceiving, reasoning, and acting until the task is done. Each pass counts as one step, and a step can include one or more clicks or keystrokes.

	Power Automate Desktop	Computer use
Who decides the steps	A person, when building the flow	AI, on every run
How you build it	Arrange actions, or record them	Describe the task in words, and pass values that change as inputs
AI during a run	Not called (unless self-healing steps in)	Called at every step
Speed	Fast	Waits for the AI at every step
Same input, same actions?	Yes	Not always
When a screen changes	Tends to stop. AI self-healing is in preview	Often reads the screen again and carries on
Cost	A license, the same however often it runs	Copilot Credits per step
Run record	Run logs	Session replay built from screenshots, and step logs

The machine: computer use also runs on a Power Automate “machine”

Computer use can run in three places.

Where	Status, October 2026	What it is
Hosted browser	Preview	An environment Microsoft provides, ready to try with no setup. It is not joined to your tenant, so it cannot reach internal systems or your own desktop apps. One person can have one session at a time, for at most three hours, and it may be unavailable at busy times
Cloud PC pool (Windows 365 for Agents)	Preview	Cloud PCs Microsoft hosts, joined to your Microsoft Entra tenant and managed with Intune. The number of PCs scales with the load
Your own machine	Generally available	Windows PCs you provide and manage. The standard setup for production

Running on your own machine takes the same groundwork as Power Automate Desktop.

• Install Power Automate for desktop: version 2.61.132.25266 or later, with the Power Automate browser extension for web tasks
• Register the machine: sign in to the Power Automate machine runtime, register the PC to the environment, and turn on “Enable for computer use” in the machine’s settings in Power Automate
• One task at a time per machine: runs sent to the same machine wait in a queue and run one by one
• Settings for unattended runs: for example, when a Microsoft Entra account signs in to a PC with Network Level Authentication on, you need an MFA exception or certificate-based authentication. These are the same kinds of settings unattended desktop flows need
• Some setups don’t work: desktops with more than one monitor aren’t supported. Sign-in with stored credentials may not work in apps built with Electron or Java, or in Citrix

Moving to computer use does not remove the work of managing Power Automate machines. Microsoft recommends dedicating PCs to computer use, using an account with the least privilege needed, and deciding which sites and apps it may operate.

Whether your own machine needs a bot license (Power Automate Process) to run computer use is something we could not confirm from Microsoft’s documentation as of October 2026. Check with your reseller before you roll it out.

Cost: pay per step, or a monthly license?

Computer use spends Copilot Credits by the step.

• 5 credits per step: with a standard model. OpenAI’s Computer-Using Agent and GPT-6 Sol, and Anthropic’s Claude Sonnet 4.5, are generally available. Claude Sonnet 4.6 is experimental
• 15 credits per step with the premium model: Claude Opus 4.6, which is experimental
• A credit costs about 0.8 US cents: a pack of 25,000 credits is $200 a month. A step is about 4 cents with a standard model, and 12 cents with the premium model. Pay-as-you-go through an Azure subscription is also available
• Not included in Microsoft 365 Copilot licenses: even in agents used by employees with a Copilot license, computer use spends credits
• Agents stop when credits run out: once a tenant uses 125% of its prepaid credits, agents in environments without pay-as-you-go are turned off

Power Automate Desktop costs a monthly license (US list prices, per month, billed annually).

• Premium: $15 per user. Runs desktop flows while a person is at the PC
• Process: $150 per bot. Runs desktop flows with nobody at the PC. One bot runs one flow at a time
• Hosted Process: $215 per bot. Process plus a virtual machine Microsoft hosts

Take entering orders, one by one, into a desktop sales app. Count each order as five steps: open the new order screen, enter the customer, choose the item from a list, enter the quantity and save, and read the order number. With computer use, an order costs about 20 cents at the pack price.

Orders a day (20 days a month)	Computer use (5 steps, about $0.20 an order)	Power Automate Process (1 bot)
10	about $40 a month	$150 a month
30	about $120 a month	$150 a month
50	about $200 a month	$150 a month
100	about $400 a month	$150 a month
300	about $1,200 a month	$150 a month

Cost flips at about 750 orders a month, or 37 a day. With the premium model, an order costs about 60 cents, and the cost flips at about 13 a day.

The estimate rests on these assumptions.

• Steps can grow: Microsoft says computer use can repeat actions or get stuck. Extra steps spend credits too
• The agent’s other activity is not counted: triggers and calls to other tools also spend credits
• Only licenses and credits are compared: the PCs, and the time of the people who build, fix, and check the work, are not included
• Process stays the same as volume grows: but one bot runs one flow at a time. If the volume does not fit in the time you have, you add bots

An estimate for a 20-step browser task is in Power Automate or an AI browser agent?

Speed: the number of AI calls becomes the difference in time

Microsoft does not publish how long a computer use step takes. But the design makes it clear that every step waits for the AI to answer. That is slower than RPA, which plays set steps.

To show the size of the gap, here is a measurement from Descarty’s Kitewell. It compares five actions that enter an order into a stand-in sales app, first with AI deciding each action from the screen, then with the actions AI had learned repeated as they were.

How it ran	Time for the five actions	Per action
AI looked at the screen and decided each action (order 1)	41.7 seconds	6.6 to 10.6 seconds
Learned actions repeated (orders 2 and 3)	5.8 seconds	0.9 to 1.5 seconds

We measured this on October 27, 2026, on macOS, with a fast, lightweight AI model. With AI deciding each action, the task took about seven times as long. The seconds depend on the model and the screen, but as long as AI is called for every action, the gap remains. Reading the order number from the screen after saving used AI in both cases and took 2 to 4 seconds.

Reliability differs too. Microsoft says computer use performance varies widely by task and application, that changes in layout or loading time can change the outcome, and that controls such as dropdowns and date pickers can be hard for it.

When a screen changes: Power Automate Desktop’s self-healing versus computer use

Situation	Power Automate Desktop	Computer use
The screen has not changed	Calls no AI	AI decides every step, and every step is billed
A button or field cannot be found	Self-healing (preview) uses AI to find the element again and keeps the run going	Reads the screen again and carries on
The next run	The flow itself is not fixed. A person fixes it	Uses nothing from the last run. AI decides again

Self-healing in Power Automate Desktop entered preview in March 2026. It needs an organizational Premium account and Power Automate for desktop 2.66 or later. Outside the United States, you also have to allow data to move across regions. The conditions are covered in Power Automate Desktop alternative.

Neither one carries what it handled into the next run. A Power Automate Desktop flow stays as it was until a person fixes it. Computer use decides from scratch on every run, even when nothing has changed.

Combining them: desktop flows for the set parts, computer use only where screens change

Microsoft presents the two as a combination, not a replacement.

• Keep RPA for fixed processes: in a February 2026 blog post, Microsoft recommends keeping classic RPA for fixed processes on stable screens and adding computer use where RPA falls short. Its example leaves the RPA as it is and gives computer use only the parts of the screen that change
• Call desktop flows from agents for step-by-step work: Microsoft’s plan for April to September 2026 (2026 release wave 1) lets Copilot Studio agents call desktop flows directly, for tasks that need precise, step-by-step execution
• Make computer use a building block: the standalone computer use tool (preview) publishes one screen task for use from several agents and agent flows

Sorted by example:

Task	Better fit	Why
Enter 100 orders a day into an internal sales app	Power Automate Desktop	High volume, and the screen does not change
An unexpected confirmation dialog sometimes appears mid-entry	Power Automate Desktop, with computer use for the dialog only	The set part stays fast and cheap, and AI takes only the part that changes
Collect invoices a few times a month from supplier portals that each look different	Computer use	Low volume, and every portal is different
Handle Outlook email or SharePoint files	Neither: a Power Automate cloud flow	Connectors exist, so no screen needs operating
Confirm a payment or an order	Hand neither of them the final step	Microsoft says computer use is not intended for financial transactions

A third option: repeat the actions AI learned once

What this sorting leaves is work with high volume on screens that change. With Power Automate Desktop, the fixes pile up. With computer use, the cost and the waiting pile up on every run.

That work suits a different design: AI learns the actions from a run that worked and repeats them from then on, and relearns only when the learned actions can’t proceed. Orders 2 and 3 in the speed test ran this way, and the five actions that entered each order called AI not once. While the screen stays the same, it runs as fast as RPA. When the screen changes, AI fixes it.

Repeat learned actions with Kitewell

Descarty’s Kitewell is an AI RPA that AI builds and AI fixes. Write the steps in words, and AI operates browsers and desktop apps.

• Repeats what worked: AI learns the actions and repeats them on later runs with that run’s values. Repeated actions don’t call AI. Steps that read values from a screen, and checks written in words, use AI on every run
• Relearns only when it can’t proceed: AI reads the screen again and learns the actions again. Run records show what was relearned
• Desktop apps on Windows and macOS: AI learns and repeats the actions on both
• Excel, row by row: runs once per row of your existing workbook and writes results such as the order number back to the same row
• People check what matters: a person’s approval can sit before any hard-to-undo action

You can watch the real recording of the three orders from the speed test on kitewell.app. Descarty’s migration tool reads Power Automate Desktop flows and moves them to Kitewell with the same steps. To build a new one, tell the assistant what you want in words, and AI puts the workflow together. Copying an Excel list into a desktop app is shown in Entering data into sales and accounting desktop apps.

Kitewell can be downloaded from kitewell.app for macOS and Windows. For work at your company, you can verify it with us as an early access partner. To connect in-house systems that have no API, see on-premises integration.

Frequently asked questions

Do I need Power Automate Desktop to use computer use?

Yes, if it runs on your own machine. You install Power Automate for desktop, register the machine in Power Automate, and turn on computer use for it. The hosted browser and the Cloud PC pool run in environments Microsoft provides, so they need no machine of your own. Both are in preview as of October 2026.

Can computer use record a run and repeat it without AI?

Not according to Microsoft’s documentation as of October 2026. After a run, you can watch a session replay built from screenshots, but that is for review and audit. On every run, AI decides the steps, and every step spends credits.

Should I replace my Power Automate Desktop flows with computer use?

Flows with high volume on screens that do not change run faster and cheaper if you keep them. Microsoft also recommends keeping RPA for fixed processes. If you replace anything, start with the changing parts of flows that often stop when a screen changes. Why they stop is covered in Why RPA bots break when screens change.

Can I use computer use at no extra cost with a Microsoft 365 Copilot license?

No. Some agent features cost no credits for employees with a Microsoft 365 Copilot license, but computer use is not one of them. Every step spends credits.

Summary

• Power Automate Desktop repeats the steps you built the same way every time, and computer use has AI look at the screen and decide at every step. That difference shows up in speed, cost, and how much results vary
• Computer use on your own machine requires Power Automate for desktop and a registered machine, and runs one task at a time per machine
• For entry work of five steps an order, computer use costs more than one Process bot at about 37 orders a day
• When a screen changes, self-healing recovers only during the run, and computer use has AI decide every time. Neither carries what it handled into the next run
• Use desktop flows for the set parts and computer use only for the parts that change. For high-volume work on changing screens, use a design that repeats the actions AI learned and relearns only when it can’t proceed

Sources

• Microsoft: Automate web and desktop apps with computer use (October 20, 2026)
• Microsoft: FAQ for the computer use tool (September 23, 2026)
• Microsoft: Configure where computer use runs (May 7, 2026)
• Microsoft: Troubleshooting and limitations of computer use (August 26, 2026)
• Microsoft: Add standalone computer use tools to agents and agent flows (preview) (June 4, 2026)
• Microsoft: Billing rates and management (August 3, 2026)
• Microsoft: What’s new in Copilot Studio (October 5, 2026)
• Microsoft: Copilot Studio pricing, Power Automate pricing
• Microsoft: Overview of Power Automate 2026 release wave 1 (April 20, 2026)
• Microsoft: Improve complex UI automation with computer‑using agents (February 24, 2026)
• Microsoft: Self-healing for desktop flows