Last Updated: 14 May 2026
Microsoft is moving Copilot Studio beyond simple chat. Its new computer use feature is now generally available, which means businesses can build AI agents that do more than answer questions. These agents can use apps, websites, and desktop tools in a way that looks much closer to how a person works.
Many business tasks still happen the same way every day: someone opens an app, reads a screen, fills in fields, checks a few details, and moves to the next step. Copilot Studio computer use is built for that kind of work. The agent can read what appears on screen, click buttons, choose options, and enter information into websites or desktop apps. For companies with older systems, vendor portals, or tools that do not connect easily with other software, this gives automation a much wider reach.
Computer use is a tool inside Microsoft Copilot Studio that lets an agent interact with a Windows computer. The agent does not need a clean API to complete a task. It can work through the screen, using a virtual mouse and keyboard to move through websites and desktop apps.
This is useful for tasks that still depend on people moving data from one place to another. Think of data entry, invoice checks, order processing, or copying information from an email into an internal system. These jobs are common, repetitive, and still hard to automate in many companies.
Microsoft says the feature is powered by computer using agents that combine vision and reasoning. In plain English, the agent can look at a page, understand what it needs to do next, and act. If a button moves or a screen changes, it can adapt better than older automation scripts.
That makes Copilot Studio computer use more useful for real business work, where screens change often.
Many companies still run important work through older apps. Some are internal tools. Some are portals from suppliers, banks, insurers, or service partners. These systems often do not have modern APIs, so teams must still enter information by hand.
Traditional RPA helped with this problem, but it often depends on fixed steps. If a screen changes, a field moves, or the workflow takes a different path, the automation can fail. Computer using agents take a different route. They look at the screen and decide what action fits the task.
For IT teams, this can reduce the pressure to rebuild every old workflow before automation begins. A company can keep using a needed system while giving an AI agent a controlled way to complete routine steps inside it.
The general availability release expands computer use across all commercial Power Platform geographies, according to Microsoft. That means more business customers can deploy agents under their own tenant, data location, and compliance needs.
Microsoft is also adding business controls around the feature. These include secure sign in options with stored credentials and Azure Key Vault, admin controls through Power Platform, data loss policies, environment separation, and audit trails. These details matter for companies that need more than a clever demo.
The release also includes human in the loop checkpoints. That means an agent can pause when it has low confidence, finds an exception, or needs approval. Admins and makers can also review run history, including what the agent saw, what it clicked, and why it acted.
Graebel, a global talent mobility company, is one of the clearest examples of how this can work in practice. The company handles complex relocation work for large clients, and many requests arrive through free form emails with instructions, attachments, and special cases.
Graebel used Microsoft Dynamics 365 Finance, Power Platform, and Copilot Studio to automate parts of its operations. Microsoft says the company used AI agents for tasks such as invoice processing, knowledge retrieval, and work across core business applications.
For computer use, Graebel worked with GET AI and Microsoft on agents that can interact with legacy systems through the user interface. Microsoft’s customer story says these agents can help automate older systems, third party apps, or older websites where modern connections are not available.
“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.”
Matt Brownlee, Chief Revenue Officer at Graebel
The most practical use cases are easy to picture. An agent can open a website, sign in, read fields, enter data, move between screens, and complete a transaction. That makes it useful for companies that live inside many apps but cannot connect all of them cleanly.
Here are common examples:
This does not mean every workflow should be handed to an agent. The best starting point is a clear, repetitive task with known steps, visible screens, and rules that a human already follows. Tasks with risk, money movement, legal review, or customer impact should keep human approval in the process.
Microsoft says makers can add computer use as a tool inside Copilot Studio. The basic path is to open or create an agent, go to Tools, add a tool, and choose computer use. Then the maker describes the task in natural language.
The setup also asks for details such as the tool name, description, task instructions, and model choice. Microsoft’s documentation says supported models include OpenAI’s computer use model and Anthropic’s Claude Sonnet 4.5, with admin settings needed to use Anthropic’s model.
For a safe first test, teams should choose one boring task that wastes time every day. Keep the task small. Add review points. Watch the run history. Then decide if the same pattern can be used in other areas.
Microsoft Copilot Studio computer use gives businesses a practical way to build AI agents that can complete work inside real apps. The main value is not in chat. It is in helping teams automate work that still depends on screens, forms, portals, and old systems.
The release also shows where enterprise AI is going. Companies want agents that can act, but they also need control, logs, security, and human approval. Copilot Studio computer using agents are a step toward that kind of business AI.
Computer using agents are AI agents that can operate websites and desktop apps through a screen. They can click buttons, choose menus, read page content, and type information into fields. The goal is to help businesses automate tasks that usually need a person using a computer. This is useful when a system has no API or when building a full integration would take too much time. In Copilot Studio, makers can add computer use as a tool and describe the task in natural language.
A normal chatbot mainly answers questions or gives instructions. A Copilot Studio agent with computer use can take action inside software. It can open apps, follow steps, fill fields, and complete parts of a business process. That makes it closer to a digital worker for repeatable tasks. The user still defines what the agent should do, and companies can add rules, review points, and admin controls. The main change is that the agent is not only talking about work. It can help carry it out.
Yes. That is one of the main reasons this feature exists. Many business systems still do not offer easy API access, especially older internal apps, vendor portals, and custom platforms. Computer use lets an agent work through the visible interface instead. It reads what is on screen and uses virtual mouse and keyboard actions to complete steps. This can help companies automate work without waiting for a full system rebuild, although careful testing and human review are still important.
Microsoft is positioning the feature for enterprise use by adding controls around access, credentials, logs, and admin review. The release includes support for secure authentication, Azure Key Vault, Power Platform governance, data loss policies, environment separation, audit trails, and human review for uncertain steps. Run history also helps teams inspect what the agent saw and did. These controls do not remove all risk, but they give IT teams more ways to manage how agents operate inside business systems.
The best early tasks are repetitive, clear, and already done by people through a screen. Examples include entering data, checking order status, processing service requests, moving information from emails into systems, or filling forms in older apps. A good task has known rules, stable steps, and low risk if a human can review exceptions. Teams should avoid starting with sensitive or high risk work. A small pilot is usually better than trying to automate a large process at once.
