Cloud Flows, Desktop RPA or AI Agents? How to Choose the Right Automation

Microsoft now gives you three ways to automate the same task. A cloud flow can connect your systems through their APIs. A desktop flow can click through an application the way a person would. An AI agent can read the request, decide what to do and act on it.

Choosing between cloud flows vs desktop flows used to be the only question. Now AI agents are in the mix, and many organisations are tempted to start there. That is usually a mistake. Most processes people want an AI agent for do not need one. They need a reliable flow, with AI used for the one or two steps that genuinely require judgement.

This guide explains what each option is good at, what it costs to license and how to combine them.

The quick comparison

Cloud flows: the default choice

Cloud flows are the workhorse of Power Automate. They run in Microsoft's cloud, are triggered by an event, a schedule or a button, and move data between systems using connectors.

Use a cloud flow when:

• The systems involved have connectors or APIs
• The rules can be written down
• You need the same result every time

Watch out for:

• Licensing. Most Microsoft 365 business plans include cloud flows with standard connectors such as Outlook, SharePoint, Teams and Forms. Premium connectors, including SQL Server, Azure and Dynamics 365, need Power Automate Premium (Microsoft Power Platform licensing FAQ).
• Generative actions are still in preview. Microsoft is adding AI-driven steps to cloud flows, where you describe the intent and the AI chooses the actions. Microsoft states that preview features are not meant for production use (Microsoft Learn: generative actions).

For 20 practical examples, see our guide to Power Automate workflow examples.

Desktop flows (RPA): for systems without an API

Robotic process automation (RPA) controls an application's user interface. It clicks buttons, types into fields and reads values from the screen. In Power Automate this is called a desktop flow.

There are two modes:

• Attended: runs on a person's machine while they are logged in, often started by them. Useful for speeding up a task someone is already doing.
• Unattended: runs on its own on a dedicated machine or virtual machine, with nobody watching. Useful for overnight or high-volume processing.

Use a desktop flow when:

• A legacy system has no API or connector, and replacing it is not an option yet
• The screens and steps are stable
• The volume justifies the build and maintenance effort

Watch out for:

• Brittleness. A software update that moves a button can break a desktop flow. Budget for maintenance.
• Licensing. Attended RPA needs Power Automate Premium. Windows includes a free version of Power Automate for desktop, but it is for personal use only, with no sharing, scheduling or support (Microsoft Power Platform licensing guidance). Unattended RPA needs a Power Automate Process licence for each bot, listed at £115.30 per bot per month paid yearly, or a Hosted Process licence if you want Microsoft to provide the machine (Power Automate pricing). Microsoft changes prices and licensing terms regularly, so check current figures with them before you budget.

RPA is often a bridge, not a destination. If you depend heavily on RPA to work around an old system, that is a sign the system is a candidate for technology modernisation.

AI agents: for judgement, not routine

An AI agent uses a language model to understand a request, decide which tools to use and carry out the task. In the Microsoft stack, agents are built in Copilot Studio. They can call agent flows, which are automations built in Copilot Studio or Power Automate that the agent runs as tools (Microsoft Learn: agent flows).

Agents can now also operate applications directly. Computer use in Copilot Studio, now generally available, lets an agent work with websites and desktop apps by selecting buttons and entering text, described in natural language rather than scripted step by step (Microsoft Copilot Studio blog). This overlaps with RPA, and it can cope better when screens change. The trade-off is less predictable behaviour and a cost for every action.

Use an AI agent when:

• Input is unstructured: free-text emails, documents, chat messages
• The task needs interpretation, such as "is this a complaint, a query or an order?"
• People need to ask questions in plain language, for example about HR policies or project status
• The path through the task varies from case to case

Watch out for:

• Variable results. The same request may not produce exactly the same output every time. For anything financial, regulated or customer-facing, add human review.
• Usage-based cost. Copilot Studio is billed in Copilot Credits. Agent flow actions, for example, are charged at 13 Copilot Credits per 100 actions (Microsoft Learn: Copilot Studio billing rates). In the UK, Microsoft lists capacity packs of 25,000 credits at £153.80 per pack per month (Copilot Studio pricing). A busy agent doing routine work a cloud flow could handle is an expensive choice.
• Governance. Agents act on your systems with real permissions. They need owners, access controls and monitoring, just like staff.

How to decide: five questions

Work through these in order for each process.

1. Do the systems have connectors or APIs? If yes, start with a cloud flow.
2. Can the rules be written down completely? If yes, a cloud flow or desktop flow will be cheaper and more reliable than an agent.
3. Is there no API, but the screens are stable? Use a desktop flow. If the screens change often, consider computer use, but test carefully.
4. Does one step need judgement over unstructured input? Keep the flow, and add an AI step for that part only, such as classifying an email or extracting fields from a document.
5. Does the whole task need conversation, or a path that changes every time? Then an agent is the right tool, with flows behind it doing the reliable work.

The best answer is usually a combination

The strongest automations use each tool for what it is good at. Take supplier invoice processing at a mid-sized business with an older finance system:

1. Cloud flow: a new email arrives in the invoices mailbox and the flow saves the attachment
2. AI step: document processing extracts supplier, date, amount and VAT
3. Cloud flow: the extracted data is checked against purchase orders; anything that does not match goes to a person for review in Teams
4. Desktop flow: approved invoices are entered into the legacy finance system, which has no API
5. Cloud flow: the supplier record is updated and finance gets a daily summary

Only one step uses AI, and only one uses RPA. Everything else is a predictable cloud flow. That keeps the process reliable, auditable and affordable.

Read more about extracting data from documents in what is Azure Document Intelligence.

Do not skip the process review

Whichever tool you choose, automating a messy process makes the mess faster. Map the current steps first, remove what adds no value, and then decide which tool fits each remaining step. Our value stream and process improvement approach covers this.

Manage bots and agents like a workforce

As automations multiply, they start to look like a second workforce. Each one has access to systems, does work on your behalf and can make mistakes. Treat them that way:

• Give each automation a named human owner
• Grant the least access it needs, using service accounts rather than personal logins
• Use separate environments for development, testing and production
• Set data loss prevention policies so flows and agents cannot move data where it should not go
• Monitor runs, failures and costs, and retire automations nobody uses

We explore this idea in managing non-human resources and in how AI agent teams can support a marketing department.

Where Cyber Samurai helps

We help UK organisations choose and build the right mix of cloud flows, desktop RPA and AI agents, and run them safely.

• Automation Opportunity Assessment: reviews your processes and recommends the right tool for each step, with licensing and effort
• AI Readiness Assessment: checks whether your data, security and governance are ready for agents
• AI Management as a Service: ongoing governance, monitoring and support for AI agents in production

Learn more about our automation and artificial intelligence expertise, or see how this fits our wider operational excellence approach.

Book a consultation