AI agents vs RPA: when to use which, step by step

“Routine work goes to RPA, work that needs judgment goes to AI agents.” That is the usual way to split them, and it is not wrong. But real business processes mix both kinds of work in a single flow, so picking one tool per process strains whichever one you pick.

Take invoices. Reading invoices that come in a different format from each business partner suits AI, entering them into the accounting system suits the same fixed steps every time, and checking amounts that don’t match is a person’s job. This article explains how to split AI agents and RPA by step within a process, the criteria for deciding, and the ways to combine them.

The short answer: split by step, not by process

• Steps that need judgment go to AI
• Steps that should be the same every time run as fixed steps (RPA, or steps AI has learned once)
• Before steps that are hard to undo, a person checks

Gartner, too, advises starting with AI agents where decisions are needed, automation for routine workflows, and assistants for simple retrieval. It also notes that many use cases positioned as agentic today don’t need an agentic implementation.

How AI agents and RPA differ

Aspect	RPA	AI agent
How it works	Repeats recorded operations exactly	Decides the next operation toward a goal on every run
Who decides the steps	A person, when building the robot	The AI, on every run
Input it handles well	Data and screens in a consistent format	Documents in mixed formats, email text
When a screen changes	Often cannot find its target and stops	Looks again on the spot, so it rarely stops
Speed	Fast	Waits on the AI for every operation
Results	The same every time	Not always the same
What it costs	Licenses, and people’s time to build and fix	AI usage fees on every run, and people’s time to check results

Gartner says AI agents will not replace RPA for now, citing consumption-based pricing that drives up costs and a higher risk profile, from unpredictable results to compliance concerns.

Six criteria for deciding

For each step, think through these six.

Criterion	Suits fixed steps (RPA and the like)	Suits AI
1. Input format	Consistent (CSV files, system screens)	Mixed (PDFs, email text, each business partner’s own format)
2. Judgment	Can be written as rules	Means reading the context
3. Volume and frequency	High, daily	Low, as needed
4. Variation in results	Not acceptable (entering amounts, matching)	A person can check (drafts, suggested categories)
5. Impact of a mistake	Goes outside the company, hard to undo (pair with human approval)	Noticed and fixed in-house
6. How often screens change	Rarely (in-house systems)	Often (business partners’ websites, SaaS)

The criteria can pull against each other. A common case is work you want to run in volume every day with the same steps (criteria 3 and 4 point to fixed steps) on screens that change often (criterion 6 points to AI). Entering data on business partners’ websites or in SaaS is the typical example. For this, the option is automation in which AI learns an operation and repeats it, covered below.

Example: splitting invoice processing by step

Here is the work of reading invoices and entering them into an accounting system, split step by step with the six criteria.

Step	Who does it	Deciding criteria
Collect incoming invoices	Fixed steps	Collected from the same place every time (3)
Read the issuer, amount, and due date	AI	Each business partner uses a different format (1)
Match the amount against purchase order data	Fixed steps	A rule can decide it, and no variation is acceptable (2, 4)
Check invoices that don’t match	A person	The payment decision and the responsibility belong to people (5)
Enter the invoice into the accounting system	Fixed steps (AI only when the screen changes)	The same operation every time, it must be exact, and the screen may change (4, 6)
Record the results and report failures	Fixed steps	Done every time, without gaps (3)

AI handles only the part that makes sense of different formats, and what to do when a screen changes. People check only the invoices that don’t match. The use case Read invoices into the accounting system shows the same flow with an estimate of the effect.

Three ways to combine them

Once the steps are split, decide how to connect them. There are three broad ways.

1. Call AI from inside an RPA scenario

Keep your RPA robots as they are and hand only the reading or sorting steps to generative AI. WinActor, for example, gained an AI integration license in October 2025 that lets a scenario instruct generative AI to draft email replies or pull data out of images.

You keep the robots you have, but RPA’s weak point remains: it tends to stop when a screen changes.

2. Call automation from an AI agent

The AI agent makes the decisions and hands fixed processing to existing automation or APIs. With protocols such as MCP, an AI agent can start automation from inside a conversation.

This suits work a person asks for on the spot. Processing that runs at a set time every day is more reliable when it starts from scheduled automation. You also need a record of who started what, and when.

3. Put them as steps in one workflow

Fixed steps, AI steps, and human approval steps sit in a single workflow. Scheduling, records, approvals, and failure alerts are managed in one place, and a person can check the AI’s judgment before the next step. For processes that run every day, this is the easiest to manage.

A third option: AI learns an operation and repeats it

Screens that change often, yet the same steps every day: for this combination, AI that copes with screen changes is spreading into RPA products too.

Approach	How it handles screen changes
UiPath Healing Agent	When a UI operation fails, it recommends fixes such as new selectors. It can also recover the run while it is going (self-healing)
Power Automate Desktop self-healing (preview)	When a UI element can’t be found, AI identifies the most likely element during the run and continues. It works at run time only, and fixing the flow is left for a person to do later
AI RPA (Kitewell)	AI learns the operation and repeats the same steps on later runs. Only when the learned steps no longer work does AI read the screen again and relearn the operation, which later runs then repeat

The differences are in what happens to the AI’s fix on the next run, and in who fixes the automation. Power Automate Desktop’s self-healing recovers the run in progress, and fixing the flow is still a person’s job. If you will keep using your current RPA product, trying its own feature first is the shortest path.

When to let ChatGPT, Claude, or Copilot Studio operate the screen itself, and when to keep Power Automate, is covered in Power Automate or an AI browser agent?

Frequently asked questions

Is RPA obsolete?

No. For work that repeats fixed steps in volume, quickly and the same way every time, RPA and other fixed-step automation still fit best. Gartner, too, says AI agents will not replace RPA for now. Is RPA dead? covers why RPA spread, why its growth stalled, and today’s market numbers.

Can AI agents alone automate a business process?

Sometimes, for work with a person nearby who checks the result. For processes that run in volume every day, have deadlines, or include operations that are hard to undo, splitting by step is more reliable. What to check before you start is in 7 things to check before adopting AI agents for business.

What should we do with the RPA robots we have?

Robots that run without trouble can stay as they are. Start with the robots that keep stopping and the work still done by hand because it needs judgment, and move them to step-by-step workflows. The steps are in RPA migration steps.

Splitting the work with Kitewell

Descarty’s Kitewell is AI RPA that AI builds and AI fixes. It lets you build the third way of combining them in one piece of software.

• Steps side by side: screen operations, questions and judgments by AI, batch processing of Excel lists, API calls, and human approval sit in one workflow
• Screen operations learned and repeated: clicks and entries that worked are remembered and repeated the same way on later runs. AI relearns only when a screen changes, and the run’s record shows what it relearned
• People check AI’s judgment: in batch processing, each AI judgment shows its confidence and quotes its evidence, and rows below your threshold go to a person to check. Approval can be set on any step
• AI agents as steps: fifteen AI agents, including Claude Code and Codex, run as steps in a workflow. Send the work back and the agent tries again with your comments
• Called from AI agents: ChatGPT, Claude, and Claude Code can run workflows and approve steps through MCP

You can start by choosing one manual process or one robot that keeps stopping, and verifying it with us as an early access partner. See Early access partners for details.

Summary

• Split AI agents and RPA by step within a process, not by process
• Give steps that need judgment to AI and steps that should be the same every time to fixed steps, and put a human check before steps that are hard to undo
• When unsure, weigh six criteria: input format, judgment, volume and frequency, variation in results, impact of a mistake, and how often screens change
• There are three ways to connect them: call AI from RPA, call automation from an AI agent, or put both in one workflow. For daily processes, the third is easiest to manage
• For routine work on screens that change often, automation in which AI learns an operation and repeats it fits

Sources

• Gartner: Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (June 25, 2025)
• ITmedia @IT: AI agents will not replace RPA for the time being (January 16, 2026, contributed by Gartner, in Japanese)
• NTT Advanced Technology: announcement of the AI integration license (September 4, 2025; in Japanese)
• UiPath: What is Healing Agent?
• Microsoft: Self-healing for desktop flows (preview)