The short answer: give an agent the steps that need judgement or that run in an application with no other way in, and keep everything else on cheaper, predictable automation. At Microsoft Ignite, held in San Francisco from 17 to 20 November 2026, Microsoft confirmed that agents built in Copilot Studio can now work on Windows 365 for Agents: Cloud PCs joined to Microsoft Entra and managed in Intune like any company laptop, where an agent reads the screen, clicks, types and moves data between web, desktop, legacy and on-premises applications. The capability is generally available. It is also billed per step, and Microsoft’s own product team presents it as an addition to robotic process automation (RPA), not a replacement. For operations and finance leaders, the question is no longer whether an agent can do the task, but which tasks are worth paying an agent to do.
What is available now, and what is still in preview
| Capability | Status in November 2026 |
|---|---|
| Computer use in Copilot Studio: agents that operate websites and Windows applications | Generally available since May 2026 |
| Windows 365 for Agents as the machine that computer use runs on, billed in Copilot Credits | Available, confirmed at Ignite |
| Agents reaching on-premises applications under their own identities | Preview |
| Managed agent browsing in Edge for Business, each agent in its own browser window | Preview |
| Edge skills recorder: an employee’s browser routine recorded as a reusable skill | Preview, with Copilot Cowork |
| Agent 365 Cost Management: spend views, limits and alerts for usage-based AI | Preview |
| Windows 365 for Agents in Microsoft Foundry | Preview planned for early 2027 |
Microsoft’s documentation names automated data entry, invoice processing and data extraction as typical uses. Because the agent reasons about what it sees, Microsoft says it keeps working when buttons or screens change, the kind of change that breaks scripted automations.
What each step costs
Computer use is billed at 5 Copilot Credits per step with the standard models, which include OpenAI’s Computer-Using Agent and Anthropic’s Claude Sonnet 4.5, and 15 credits per step with a premium model. A step can contain several clicks or keystrokes; Microsoft’s own example, a four-step timesheet, costs 20 credits. Credits are sold in packs of 25,000 for USD 200 a month, or USD 0.008 each, with pay-as-you-go and prepaid options alongside. Unlike most agent features used by employees who hold a Microsoft 365 Copilot licence, computer use is not included in that licence.
Other ways of automating the same work are metered differently. Agent flows, which follow fixed steps without the model reasoning at each one, cost 13 credits per 100 actions. Reading a document with Copilot Studio’s content processing tools costs 8 credits per page. Power Automate cloud flows are covered by Power Automate licences and use no credits at all.
Take a supplier invoice keyed into an older ERP screen, assuming 15 steps: open the screen, find the supplier, fill ten fields, attach the PDF and submit. Run by computer use on a standard model, that is 75 credits, about USD 0.60 an invoice, or USD 600 a month at 1,000 invoices. If the same two-page invoice is read by content processing and posted through an API by a 20-action agent flow, it uses about 19 credits, roughly USD 0.15. The step counts are our assumptions; the rates are Microsoft’s published ones. The point stands at any volume: where an API exists, the agent is the expensive route.
Capacity limits matter too. When a tenant on prepaid credits reaches 125% of its capacity, Microsoft blocks its agents, including conversations already under way, unless overage is billed to pay-as-you-go. A month-end surge can therefore stop an automation exactly when finance needs it.
Which tool for which step
| Situation | Best fit |
|---|---|
| The application has an API or a ready connector | A workflow in Power Automate, n8n, Make or Zapier |
| A stable desktop screen, no API, high volume | An RPA desktop flow that repeats the same steps |
| Screens that change, portals without an API, occasional tasks | A computer-using agent |
| Invoices, forms and emails in varying layouts | Document extraction feeding a workflow |
| Exceptions that need judgement, such as a price mismatch | An agent that proposes and a person who approves |
Microsoft’s Copilot Studio team draws the same line: classic RPA stays with predictable work on stable interfaces, while computer-using agents take the variable parts and escalate what they cannot resolve. Most real processes mix all five rows, so the design work is in splitting a process into steps and giving each step the cheapest tool that handles it reliably.
Controls to set before an agent touches finance systems
- Its own machine and account. Microsoft recommends dedicated, isolated machines and an account with only the permissions the task needs.
- Credentials out of reach of the model. Sign-in details are stored encrypted in Power Platform or in your Azure Key Vault. Take care with the default setting: an agent shared with maker-provided credentials lets everyone who uses it act with its author’s access.
- An allow list. Limit the websites and applications the agent may act in. Microsoft notes that the list blocks actions, not navigation, so pair it with browser and application policies on the machine.
- A named reviewer. If the agent detects instructions that could be harmful, it emails a reviewer and stops when the request expires unanswered.
- A record of every run. Session replay with screenshots, step logs and Microsoft Purview audit logs, kept for a period you choose.
An agent that can use any screen a person can use needs the same scrutiny as a new employee with system access. Our article on treating AI agents as insiders sets out the wider controls.
Six steps for your 2027 automation plan
- List the candidate processes. Record volume, steps, systems touched and error rates for each, so choices rest on numbers.
- Look for an API first. Many systems that teams operate by hand already offer a connector or an API that a workflow can call.
- Keep agents for the variable step. Let rules and workflows carry the routine path and call the agent only where screens or documents vary.
- Put a person on every money-moving action. Supplier payments, bank details and credit notes should wait for approval, whatever the tool.
- Price each run before building. Multiply steps by credits and volume, set monthly limits, and decide in advance whether a full capacity should block or bill.
- Measure after go-live. Track cost per item, escalations and failed runs each month, and move steps to cheaper tools as patterns settle.
Where SoftwareBiz fits. Our intelligent process automation service starts with process discovery, then splits each process into workflow automation, system integration through Power Automate, n8n, Make or Zapier, and AI-assisted steps where rules are not enough, with a person approving sensitive actions and run logs and alerts that keep it reliable. Where invoices, contracts and forms arrive in many layouts, our generative AI and AI agents team builds the document processing that extracts their data straight into your systems. Where you would rather hand over the work itself, our finance and accounting outsourcing team runs invoice processing, approvals and payment runs inside your accounting system.
What to do this quarter
Choose one process with clear volume, such as supplier invoices or order entry, and map it step by step. Mark which steps have an API, which run on stable screens and which need judgement, then price each route with Microsoft’s published rates before anyone builds an agent. The vendors have made agents that can use any screen. Deciding which screens they should use is still your call, and it decides the bill.



