Computer Use in Copilot Studio just got a lot more enterprise-ready
If you have been following Copilot Studio over the last year, you have probably noticed Microsoft pushing hard on what they call Computer-Using Agents (CUA) — agents that can actually use a computer the way a person does. Move the mouse, click buttons, type into fields, read what is on the screen.
I find this area really interesting because it solves a problem most of us have hit at some point: the system you need to automate has no API. It is a legacy app, a vendor portal, or a third-party site that just expects a human to log in and click around. Classic RPA can handle some of it, but as soon as the UI changes, the bot breaks.
This is where Computer Use comes in. And with the latest update from Microsoft (announced February 24, 2026), it has taken a big step from “interesting preview” to something I think is starting to look ready for real enterprise use.
What is Computer Use, in short?
Computer Use is a tool you can add to a Copilot Studio agent. Instead of calling an API or a connector, the agent looks at the screen, understands what is there, and performs actions — clicks, typing, scrolling, navigating menus.
You describe the task in natural language. Something like:
“Log in to the vendor portal, download the latest invoice, and enter the amount into our ERP system.”
The agent then figures out the steps. And because it is AI-powered with vision and reasoning, it does not break the moment a button moves five pixels to the right. That is the big difference from classic RPA.
What is new lately
1. You can now choose the model
Until now, Computer Use ran on OpenAI’s Computer-Using Agent model. With this update, you can also choose Anthropic’s Claude Sonnet 4.5.
The idea is that you pick the model that fits the task:
- OpenAI Computer-Using Agent — good for orchestrating multi-step web and desktop flows.
- Claude Sonnet 4.5 — better when you need stronger reasoning on dynamic UIs and dense, changing dashboards.
Nice to have the choice. In practice, I would test both on the actual workflow and see which one is more reliable for your scenario.
2. Built-in credentials and Azure Key Vault support
This was the one that really stopped many of us from going further. Authentication.
If your agent needs to log into a portal every night and there is no clean way to handle credentials, you cannot really call it “unattended”. You end up babysitting it.
Now you have two storage options:
- Internal storage — encrypted in Power Platform. Easy to set up.
- Azure Key Vault — for proper enterprise-grade secret management.
Credentials are encrypted, never exposed to the AI model, and can be reused across multiple agents and automations. So your agent can log into a vendor portal and a desktop ERP without anyone typing a password.
3. Managed Cloud PCs for scale
Until now, if you wanted to run Computer Use at scale, you needed to manage your own machines or virtual machines for the agents to run on. That gets old fast.
The update introduces a Cloud PC pool, powered by Windows 365 for Agents. These are fully managed cloud-hosted machines, Microsoft Entra joined, Intune enrolled, and built specifically for Computer Use runs.
The point is that you can handle peaks without keeping a fleet of patched, idle desktops sitting around. For evaluation, you can create up to two Cloud PC pools per tenant with 50 hours of free usage for published autonomous agents — so you can actually try it out before committing.
Should this replace your existing RPA?
No. And Microsoft is pretty clear about this too.
If you already have Power Automate Desktop flows that work fine on stable interfaces, keep them. Classic RPA is still the right tool for deterministic scenarios where the UI does not change much.
Computer Use shines where RPA struggles:
- Interfaces that change frequently
- Systems without APIs
- Workflows that need decision-making, not just clicks
A nice pattern is to combine the two. Let RPA handle the stable, deterministic parts, and let a CUA take over the messy, variable bits where reasoning is needed. So instead of rebuilding everything every time a vendor updates their portal, the agent just adapts.
Where is Computer Use available?
This is the part that will be a bit disappointing for those of us in Europe.
UPDATE:
Only a few days after i created this article Computer Use started to show up in environments located in Europe as well
Computer Use is in public preview and only available in Copilot Studio environments where the region is set to the United States.
So if your tenant is in Europe, the UK, Asia, or anywhere else — you cannot just turn it on. To try it out today, you need to:
- Have access to a Power Platform environment located in the United States, or
- Create a new environment in the US region for testing purposes.
A couple of other things worth knowing about regions:
- Claude Sonnet 4.5 is rolling out gradually across the supported regions. Even inside US-based environments, it might not show up as a model option yet. OpenAI’s Computer-Using Agent has been there from the start.
- External models must be enabled by your administrator before you can pick a non-default model.
- The rest of Copilot Studio’s generative AI features are already available globally (Europe, UK, Australia, India, Japan, Asia Pacific, South America, Switzerland and more).
If you are in Norway like me, the practical answer for now is: spin up a US environment, do your proof of concept there, and be ready to move to a regional environment later when it lights up.
How to get started
If you have access to a US-based Copilot Studio environment, you can try it today. The basic steps are:
- Open or create an agent in Copilot Studio.
- Go to Tools → Add tool → New tool and select Computer use.
- Describe the task in natural language.
- Optionally, choose a model, configure built-in credentials, and set up a Cloud PC pool.
Then publish and test.
Just make sure the generative orchestrator is enabled on your agent — Computer Use depends on it.
My take
Computer Use has felt like a “look what is possible” feature for a while. With the latest updates — model choice, proper credential management, and managed Cloud PCs — it is starting to look like something you can actually put into production as soon as it’s out of preview.
The use cases I would look at first are the boring ones nobody likes doing manually: invoice processing, vendor portal lookups, data entry into legacy systems, downloading reports from sites that have no API. Things that eat hours every week and never quite justify a full integration project.
The big caveat is still the region limitation. Until Computer Use lands in European environments, most of us will be testing in US tenants and waiting. But the direction is clearly the right one.
I will be playing more with this in the coming weeks, especially around credentials in Azure Key Vault. I will share what I find.
If you have already tried Computer Use in your environment, I would love to hear what worked and what did not. Drop a comment below.
Sources:
- Computer-using agents now deliver more secure UI automation at scale — Microsoft Copilot Blog
- Automate web and desktop apps with computer use (preview) — Microsoft Learn