Computer-Using Agents: Copilot Operates Any Software

In May 2026, Microsoft reached a milestone: with computer-using agents in Copilot Studio, companies can now build AI agents that operate software through its user interface - just as a human otherwise would with mouse and keyboard. The feature has been generally available since May 13, 2026 and is rolled out across all commercial Power Platform regions. That makes Microsoft the first major cloud provider to offer production-ready computer use by agents.

For small and medium-sized businesses, this is more than a technical footnote. Many SMBs work with specialized applications that offer no modern interface - legacy merchandise management, industry software, supplier web portals. That is exactly where the new technology comes in: it automates even when a classic API is missing.

What computer-using agents actually deliver

A computer-using agent looks at a screen, recognizes buttons and input fields, and performs actions on them. It reads, clicks, types and navigates through applications - regardless of whether it is a website or a desktop program. The decisive advance: these agents have become more robust. If a screen or a form changes, they do not immediately fail but adapt. For a long time, that was the biggest problem with classic screen automation.

The technical trick behind it: instead of relying on rigid screen coordinates or fixed element identifiers, the agent uses image understanding and reasoning. It "sees" the interface the way a human sees it and decides the next sensible step from that. If a button moves or a piece of software changes its layout after an update, the agent usually still finds the target. Exactly this robustness was the perennial problem of the classic, coordinate-based screen automation of the past.

Microsoft has also designed the feature for enterprise operation. Credentials can be managed more securely - for example via Azure Key Vault - and every session is logged through Microsoft Purview. Organizations can choose the appropriate model per scenario, depending on their requirements for speed, quality and cost. At launch, the generally available options include the OpenAI computer-use model and Claude Sonnet 4.5.

More than just clicks: the full package in the May update

The screen operation does not come alone. The May 2026 update of Copilot Studio bundles several new features that together form an end-to-end automation system:

• Embedded workflows: Computer-using agents can be integrated directly into multi-step processes that combine API actions, approvals and business logic.
• Agent-to-agent communication (A2A): Also generally available - agents can exchange information, delegate tasks and collaborate across systems.
• Better orchestration: A new control layer improves the execution of business processes by around 20 percent while cutting token consumption by about 50 percent.
• Model variety: Alongside GPT-5, options such as Claude Sonnet 4.5, Sonnet 4.6 and Claude Opus are available for production agents.

The improved orchestration in particular is worth real money for SMBs. More reliable execution means fewer aborted runs and less rework, and lower token consumption directly reduces running costs - a point that quickly matters as usage grows. Agent-to-agent communication, in turn, makes it possible to divide up tasks: one specialized agent does the research, a second one writes, a third enters the result into the target system. This way, surprisingly capable workflows emerge from simple building blocks, without any single agent having to do everything.

Typical use cases for SMBs

The appeal lies in everyday, unloved routines. Wherever employees transfer data from one system to the next, an agent can take over:

• Transferring order data from a supplier portal into your own merchandise management system.
• Entering invoices from the email inbox into an accounting program.
• Reconciling master data between the CRM and an older industry application.
• Compiling recurring reports from multiple applications.

The common denominator of all these tasks: they are frequent, rule-based and tie up expensive working time without creating value. This is exactly where automation pays off fastest. An employee who used to spend 45 minutes every morning transferring order data gets that time back for tasks that require judgment and customer contact - precisely what a machine cannot deliver.

Computer-using agents or a classic API - which makes more sense?

Despite all the excitement, one rule holds: screen operation is not always the best route. If an application offers a clean interface, connecting directly through that API is usually faster, cheaper and more stable. Computer-using agents play to their strength where there simply is no interface - with legacy industry software, walled-off portals or applications whose vendor does not provide for integration.

A simple rule of thumb helps with the decision:

• Is there an API or a ready-made connector? Then use it - it is more robust and usually cheaper in continuous operation.
• Is there no interface at all? Then a computer-using agent is often the only bridge short of replacing the software.
• Does the interface change constantly? Check whether the application is stable enough - with very frequent layout changes, the maintenance effort rises even for an adaptive agent.

Do not forget governance

With the power to operate any software comes greater responsibility. Microsoft has therefore also expanded administration: administrators can install first-party agents at scale and automatically assign orphaned agents to responsible owners. For SMBs this means: define from the start which agents exist, who looks after them and which permissions they have. Otherwise you quickly end up with uncontrolled sprawl that is hard to rein in.

Is now the right time to start?

If you already use Microsoft 365, the obvious platform is already in house. The move from public preview to general availability means the feature may be used in production and is actively supported by Microsoft. That noticeably lowers the risk. A sensible way to start is with a clearly defined process against which the benefit can be measured - for example the time saved on data transfer.

How to take the first step

A successful start follows no big bang but a manageable pilot project. This sequence has proven itself:

• Pick a process: Look for a routine that occurs frequently, is clearly described and is done manually today - ideally one your team already grumbles about anyway.
• Make success measurable: Record beforehand how much time the task costs today. Only then can the gain be demonstrated later.
• Start with approvals: Let the agent run under supervision at first and have it present critical steps for confirmation. Only once the quality convinces you should you grant more autonomy.
• Document cleanly: Define what the agent may do, who looks after it and how errors are reported.

Within a few weeks, this produces solid real-world experience - and a sound basis for deciding whether and where further processes are worthwhile.

Conclusion

Computer-using agents close a gap that has held SMBs back for years: automating software without a modern interface. With general availability since May 2026, the technology has left the experimental stage. What remains decisive is a careful selection of use cases and clear governance. Then the innovation becomes a real productivity gain instead of yet another never-ending project.

Want to use computer-using agents in Copilot Studio for your processes? Cryon supports you with experience in the Microsoft world - from concept and setup to sound governance. Together we identify the workflows in your company that are truly worth automating.