Whether to use an AI agent or RPA (Robotic Process Automation) is decided step by step, not for a whole process at once. Steps that follow a fixed procedure and must give the same result every time go to RPA (a deterministic flow). Steps that need judgment, such as reading, sorting or choosing, go to an AI agent. When a process has both kinds of steps, combine them: the agent decides and hands execution to RPA. UiPath also describes this pattern (agentic workflows), where agents handle the dynamic decisions and hand execution to predictable workflows.

RPA A person defines the procedure

  • Input has a fixed shape (forms, CSV, fixed fields on a screen)
  • The same input must give the same result
  • Stops when screens or fields change
  • Cost per user or per bot (in Power Automate)
⇄Split by step
and hand off

AI agent Decides the procedure on the fly

  • Input varies (emails, PDFs in different layouts)
  • Many exceptions, the path changes each time
  • Can carry on after a wrong action as if it worked
  • Cost grows with the number of calls or actions

Sources for each point are given in the sections below. Product feature names and prices are as shown in the official documentation and pricing pages on November 14, 2026.

1. The difference: who decides the procedure

RPA repeats screen operations and data transfers exactly as a person designed them. If the input is what it expects, it gives the same result every time (deterministic). An AI agent is given a goal and the tools it may use, and a language model chooses the next action on the spot. It can adapt to inputs in different layouts and to exceptions, but the same input does not always produce the same path or the same result (probabilistic).

UiPath sets out the difference between agents and traditional workflows (fixed flows, including RPA) in its official page "Agents and workflows". It is useful that a company selling RPA lists the agents' weak points in the same table.

AspectAgentWorkflow
Control flowPlans during the run and picks its toolsFixed sequence with fixed tools
InputUnstructured, including images and documentsStructured records or forms
Adapting to changeAdjusted on the fly by re-promptingNeeds changes at design time
ReliabilityVariable (depends on guardrails and evaluations)High if inputs stay within spec
Governance loadHigherMature policies and tools exist
Cost predictabilityMedium to low (token use varies each run)High

Source: excerpted from the comparison tables in UiPath's "Agents and workflows". The same page also describes "agentic workflows", where agents handle the dynamic decisions and hand off to predictable workflows for execution.

The point is not which one is better, but what each one takes on. With RPA, a person guarantees that the procedure is correct at design time. An agent takes on the job of deciding the procedure itself, so on the agent side you need separate checks on results, limits on permissions and cost monitoring. For the bigger picture of what AI does well and badly, see what AI can and cannot do.

2. A step-by-step decision table

Split the process into steps and ask these five questions for each one. Only the steps that match the right-hand column are candidates for an AI agent.

QuestionKeep on RPAConsider an AI agent
1. What shape is the input?Fixed forms, CSV, fixed fields on a screenEmail bodies, PDFs in mixed layouts, chat
2. Must the same input give the same result?Yes (accounting, payments, filings)Wording may differ (summaries, drafts, classification)
3. How many exceptions?Rare, and exceptions can also be written as rulesDaily, and a person thinks each one through
4. Can a mistake be undone?No, or the damage is large (payments, orders, deletions)A person can check before it is final, or it can be redone
5. Does the procedure change?Not for nowThe path changes with each counterpart or situation

Question 4 weighs the most. For actions that cannot be undone, even if the agent makes the judgment, the final execution goes through human approval and a fixed procedure. Microsoft's FAQ for computer use in Copilot Studio (AI that operates the screen) also lists making financial transactions, and recommending or scoring in hiring, healthcare or finance, among the uses the tool is not intended for.

Applied to common tasks, it looks like this.

TaskWhat to useWhy
Copy invoices in a fixed format into the accounting systemRPAThe input is structured and the same result is required
Read invoices whose format differs by supplierExtract with an agent → enter with RPAReading needs judgment, entering is routine
Route inquiry emails and draft repliesAgent + a person sendsNatural language with many exceptions, and sending cannot be undone
Employee onboarding (accounts, equipment requests)RPA or a workflowThe steps are fixed (UiPath uses the same example)
Executing payments, month-end closeRPA + human approvalCannot be undone and needs an audit trail
Operating an old internal system with no APIRPA (if the screens are stable)Stable screens can be driven deterministically (if they keep changing, trial AI screen control first)

UiPath's comparison table also lists invoice processing as a workflow case that follows a fixed path. Send only the reading step for mixed formats to an agent.

3. How each one fails

An easy point to miss when choosing is that they break differently. RPA stops. An AI agent can keep going after a mistake. A stop can be caught by an alert, but a mistake that carries on is invisible unless something checks the result.

AspectRPAAI agent (including screen control)
Typical failureA screen element (selector) is not found and that step errors outClicks the wrong place, repeats the same action, or treats the task as done without checking
CauseChanges to the screen layout, field names or pop-upsMisreading the screen, ambiguous instructions, or commands written on the screen (prompt injection)
How you noticeLogged and alerted as a failed runLooks like a normal finish unless the result is checked separately
Main safeguardsFailure alerts, a path back to a person for stopped items, checks for screen changesA verification step, human confirmation before important actions, an isolated environment and allow lists

RPA products are also adding AI to stop less often. UiPath's Healing Agent suggests new selectors, delays and handling for unexpected pop-ups when a UI automation fails, and tries recovery strategies in sequence during the run.

On the agent side, the vendors' official documentation is explicit about the cautions. Here are the names of the features that let AI operate a screen (computer use) and what the official pages warn.

FeatureWhat the official documentation warns
Anthropic: the computer use tool in the Claude APIClaude sometimes acts as if an action succeeded without checking the result. It may follow instructions found in screens or images. The page recommends a dedicated virtual machine, not giving it login details, an allow list of domains, and a human confirmation before decisions with real consequences.
OpenAI: the computer tool in the APIIt can affect real accounts and data. Treat screen content as untrusted and confirm consequential actions. Use an isolated browser or VM and an allow list of sites and actions.
Google: Computer Use in the Gemini APIIt may contain errors and security vulnerabilities. Supervise important tasks closely and avoid it for critical decisions or sensitive data. When a response asks for confirmation (require_confirmation), ask the end user. Run it in a sandboxed VM or container.
Microsoft: computer use in Copilot Studio (FAQ)The feature page says it adapts when buttons or screens change, but the FAQ warns as follows. Layout changes or loading delays can give different outcomes for the same task. Dropdowns and date pickers are hard for it. It may repeat actions or get stuck. Human supervision works probabilistically, so do not treat it as the last line of safety.

The agent inside ChatGPT and browser-based AI products are listed with their current status in how far AI can automate browser tasks. Incident types and defenses are covered in AI agent security incidents.

Microsoft's FAQ says AI screen control is also sensitive to interface changes. You cannot say that an agent never breaks when the screen changes. It just has more room than RPA to adapt, and whether the adapted result is correct still needs a separate check.

4. How costs work: fixed units vs. usage

RPA and agents are priced in different shapes. Power Automate's RPA is priced by the number of users or bots, while agents cost more with the number of model calls or screen actions. Which one is cheaper flips with volume and the number of actions per item, so estimate in both shapes.

Pricing shapeOfficial exampleAs volume grows
RPA per userPower Automate Premium: $15 per user per month (billed yearly). Includes attended RPA, which runs on a machine a person is using.Stays the same if the users stay the same.
RPA per botPower Automate Process: $150 per bot per month, and Hosted Process (on a Microsoft-hosted virtual machine) $215 per month (both billed yearly). Includes unattended RPA.Stays the same while one bot can handle the load. Add bots when it cannot.
Agent per callUiPath Unified Pricing: with UiPath-hosted models, each LLM call consumes Platform Units (0.2 for standard models, 0.4 for premium, 0.16 for basic). Calls are counted in 64k input token increments.Grows with volume and with the number of calls per item.
Agent per screen actionCopilot Studio computer use: 5 Copilot Credits per step (a group of clicks or typing), 15 with a premium model.Grows with volume × steps.
Agent per tokenClaude API computer use: each screenshot is roughly 1,000–1,800 input tokens.The longer the task, the more image tokens pile up.

Sources: Power Automate pricing (US list prices; they vary by country, region and contract), UiPath Agents licensing (on Flex plans, Agent Units: 1 standard, 2 premium, 0.8 basic), Copilot Studio computer use, Claude computer use tool.

Example estimate: entering 500 invoices a month in mixed formats with AI screen control

If each invoice takes the same number of steps as the example in Microsoft's pricing explanation (a timesheet: open, create new, fill three fields, submit = 4 steps), a standard model uses 500 invoices × 4 steps × 5 = 10,000 Copilot Credits a month. A premium model triples that to 30,000.

If you split the same work into "the agent extracts the values → RPA enters them", the screen actions move to the RPA side, and the agent only costs the extraction calls. Which is cheaper depends on the credit price and how much spare capacity your RPA bots have, so multiply with the unit prices in your own contract.

Copilot Studio agent flows consume capacity for each action they run, and when the environment has no Copilot Credits left, new runs are blocked (to keep runs going, you can turn on pay-as-you-go billing; Agent flows overview). Runs that an agent starts while a user with a Microsoft 365 Copilot license is chatting with it are not affected, but runs it starts autonomously, for example after an event trigger, are charged and can be blocked. Work stopping when the spending limit is reached is something per-bot RPA does not have. Add capacity monitoring to your operations checklist.

5. Permissions and audit

An RPA bot only performs the operations it was built for, so the breadth of its permissions equals the range of operations you designed. An agent chooses among its tools by itself, so every permission you give it is within the range it may use. Here are the points to check from the official documentation.

Four things to check before building

Available tools
A UiPath agent can use only the processes you explicitly add (Automations). In Automation Anywhere, you also add only the Processes, Task Bots, API Tasks and AI Agents you need in the Tools tab (Create and manage AI Agent). Give it only what the work needs
Whose credentials
Copilot Studio computer use uses the maker's credentials by default. Microsoft warns that if you share an agent with this setting, anyone using it can act with the original author's access. If you share it, switch to end-user credentials or use a dedicated account with narrow permissions
What it may touch
The Copilot Studio allow list blocks actions on sites and apps not on the list, but it does not stop them from being opened. Layer dedicated machines, least privilege and allowed sites and apps on the machine side too
Keeping records
Automation Anywhere's AI Governance shows logs of prompts and responses exchanged with models. In the standalone computer use tool in Copilot Studio (preview), you can review the last 28 days of runs in the Activity tab (session replays, screenshots and step logs). Make the agent's decisions and the RPA run logs traceable by the same case number

Approvals also need care. Microsoft writes that human supervision for computer use works probabilistically, so it may not pause when you want it to and may pause when it is not needed, and asks you not to rely on it as a safety fail-safe. Do not leave approval of irreversible actions to the agent's judgment. Put it in the flow as an approval step that every item must pass. The thinking behind constraining an agent's judgment is covered in AI guardrails.

6. Combining them: three examples and product features

The skeleton of a combination is always the same. The agent reads and decides, RPA executes the items that fit the rules with a fixed procedure, and items that do not fit or have low confidence go to a person.

1. Receive: input from email, PDFs or forms
↓
Agent 2. Read, classify, extract values
↓
RPA 3. Fits the rules: registers or enters with a fixed procedure
Person 3. Does not fit / low confidence: approves or decides
↓
4. Record: keep decision and run logs under the same case number, and use the items sent to people for the next improvement

Examples: how to split three tasks

  • Invoices in mixed formats

    The agent extracts supplier, amount, due date and line items, and sends items with missing required fields or mismatched totals to a person. RPA enters only the complete ones into the accounting system. Payment runs in a separate flow, executed by RPA after human approval.

  • Inquiry emails

    The agent classifies the intent and drafts a reply. Routine processing such as order lookups or address changes goes to RPA or a flow. A person checks before sending, and irreversible actions such as refunds pass an approval step.

  • Old system with no API

    If the screens are stable, drive them deterministically with RPA. Only when screens change so often that maintenance cannot keep up, trial AI screen control first, and add a result-check step before expanding it to production. Microsoft also describes computer use as a way to work when there is no API.

UiPath: the agent calls processes as tools

  1. Add a process as a tool: in the agent's Design tab, under Tools, select Add tool, then pick a process from the Automations tab. It must be a project in the same solution or a process published to a shared Orchestrator folder. Inputs and outputs are limited to the types the docs list, such as strings, numbers and booleans (Automations).
  2. Build a path to people: Escalations use Action apps in Action Center to ask a person to validate, approve or help manually. Assignment can go to a specific user, by workload, round robin and other options (Escalations).
  3. Connect the whole process: Maestro manages agents, robots and human tasks as one flow.

Microsoft: Power Automate and Copilot Studio

  • One AI step inside a flow

    Add AI Builder's Run a prompt to a cloud flow to classify or summarize, branch on the result, and run a desktop flow (RPA) with Run a flow built with Power Automate for desktop. People define the flow and AI is only one step, so this is the most deterministic setup (Use your prompt in Power Automate, Trigger desktop flows from cloud flows).

  • The agent calls a flow

    On the agent's Tools page in Copilot Studio, choose Add a tool → Flow to add an agent flow. The flow needs the When an agent calls the flow trigger and a Respond to the agent action, asynchronous response turned off, must be published, and must respond within 100 seconds (Add an agent flow as a tool). Microsoft describes agent flows as deterministic: the same input always gives the same output.

  • AI for screen control too

    Copilot Studio computer use operates a Windows screen from natural-language instructions. The standalone tool version, reusable across flows, is in preview, and Microsoft states that preview features are not meant for production use (Add standalone computer use tools).

Automation Anywhere: AI Agent Studio

AI Agent Studio in Automation 360 consists of AI Agents, AI Skills, AI Guardrails and AI Governance, and requires the Enterprise Platform license. You can add existing Task Bots and Processes to an AI Agent's Tools, so your current bots can serve as the agent's "hands". In the official walkthrough (an agent that processes insurance claims), the instructions tell the agent to use Human in the Loop to ask for help when claim details are incomplete.

If you build an agent yourself, see the AI agent frameworks comparison for choosing tools, and business automation use cases for examples.

7. Rollout in six steps

UiPath's "Best practices for building agents" recommends building agents as components embedded in existing business flows rather than as standalone bots. Turned into product-neutral steps, that approach looks like this.

  1. Split the current process into steps and mark the ones that need judgment: judgment means reading unstructured input, comparing context, or choosing among reasonable options. Unmarked steps go to RPA or a deterministic flow.
  2. Lock down the routine steps in RPA first and define their inputs and outputs: assume they will be called by an agent and keep arguments to simple types. If Copilot Studio will call them, split them into units that return within 100 seconds.
  3. Give the agent one role: one responsibility, such as "extract values from invoices" or "classify the intent of inquiries". UiPath's same guide also says a narrow role is easier to validate.
  4. Test with an evaluation set before production: mix in not only typical cases but broken formats, missing information and confusing inputs. How to test is covered in the AI agent evaluation guide.
  5. Draw the path back to people and the permission lines: send missing information, tool failures, ambiguous decisions and rule exceptions to a person. Make approval a mandatory step for irreversible actions, and if the agent controls screens, limit the apps and sites it may touch.
  6. Go live small, then improve from the logs and costs: start with one department and one document type. Look at the decision records, the RPA run logs and capacity or credit use in the same units, and use the items sent to people to improve instructions and rules.

Summary

  • How to decide: split the process into steps, not the whole thing, and ask about input shape, whether the same result is required, how many exceptions there are, whether mistakes can be undone, and whether the procedure changes.
  • How they break: RPA stops when screens change, while an agent may carry on after a wrong action. Add a result check to the agent's steps.
  • Cost: RPA is priced per user or bot, agents per call or action. Estimate both as volume × actions × unit price.
  • Permissions and audit: give tools explicitly and minimally, check whose credentials it runs with, and make decisions and runs traceable per case.
  • Combining: when a process has both kinds of steps, the agent decides, RPA executes, and irreversible actions pass a human approval step.

FAQ

Q. Should we replace RPA with AI agents?

That is not a reason to replace it. For routine steps that need the same result from the same input, deterministic RPA makes both results and costs easier to predict. Hand only the steps that need judgment to an agent, and keep the routine steps on RPA.

Q. Do we have to combine them?

No. If every step is routine, RPA alone is enough. If the work is made only of judgment steps (such as classifying inquiries), an agent alone is enough. You need the combination when judgment steps and routine steps sit side by side in one process.

Q. If RPA stops when screens change, is AI screen control better?

Not necessarily. Microsoft writes that AI screen control also gives different outcomes when layouts change or pages load slowly. Use RPA for stable screens, trial AI screen control for screens that keep changing, and add a result-check step either way.

Q. UiPath or Power Automate: which should we build on?

Building on your existing RPA assets and company platform means less rework. If you have processes in UiPath Orchestrator, call them as tools from a UiPath agent. If Microsoft 365 and Power Platform are central, Copilot Studio with Power Automate flows is the natural choice. Their pricing shapes differ, so estimate both from your volume and number of actions.