RPA vs AI: the difference, five ways to combine them, AI RPA

“What is the difference between RPA and AI?” The usual answer is that RPA repeats the steps a person set, exactly as set, while AI reads documents and screens and makes judgments. That is not wrong. But now that generative AI and AI agents are everywhere, the line between the two has moved. AI now builds RPA robots, and fixes robots that stopped because a screen changed.

This article sets out how RPA and AI differ, then the five ways to combine them and how to choose for each kind of work.

The short answer: RPA repeats set steps, AI reads and judges

• RPA repeats the steps a person set, quickly and the same way every time. It tends to stop when a screen changes
• AI reads documents and screens in any format, pulls out values, classifies, and judges. Its results are not always the same, and each run costs a usage fee
• There are five ways to combine them: AI reads and RPA enters, RPA calls generative AI, AI builds and repairs RPA, an AI agent operates the screen every run, or AI learns an operation and repeats it (AI RPA)
• What decides is how much the screens change, and who fixes the automation. For daily work on screens that change often, such as other companies’ web systems, AI RPA fits

How RPA and AI differ

Aspect	RPA	AI
What it does	Operates screens and moves data, following set steps	Reads documents, screens, and text; pulls out values, classifies, and judges
Who decides the steps	A person, when building the robot	AI, from the instructions it is given and what it has learned
Input it handles well	Uniform data and screens that don’t change	Invoices and purchase orders in every format, email bodies
Results	The same every time	Not always the same
When a screen changes	Tends to stop, unable to find what it operates on	Reads the screen again and finds the target
Speed	Fast	Waits for an AI response each time
Cost	Licenses, and the time of the people who build and fix robots	A usage fee per run, and the time of the people who check results

RPA does the work of a person’s hands. AI takes on part of the work of a person’s eyes and head: reading and judging. In real work, reading, judging, and entering are mixed in one flow, so one of the two rarely finishes the job alone.

The “AI” combined with RPA comes in three kinds

When people talk about combining RPA and AI, the AI can mean different things.

Kind	What it does	How it works with RPA
AI-OCR	Reads text from paper and PDF documents and turns it into data	RPA enters what it read into business systems. The oldest combination
Generative AI (large language models)	Reads text to summarize, classify, and extract; writes text and code	Called from inside RPA scenarios, and also used to build the robots themselves
AI agents	Decide the next operation toward a goal and carry it out	Read and operate screens, so they can do RPA’s work, thinking it through on every run

In MM Research Institute’s 2024 survey in Japan, 31% of companies using RPA combined it with generative AI, fully or in part.

Five ways to combine RPA and AI

1. AI reads, RPA enters

AI-OCR or generative AI reads invoices and application forms, and RPA enters the results into business systems. This has been the most common combination for documents that arrive on paper or as PDFs.

Reading has become more accurate. But the entering side is still RPA, so it stops when the business system’s screens change.

2. RPA calls generative AI from its scenarios

The existing robots stay as they are, and only some steps go to generative AI, such as classifying email, drafting replies, or extracting data from images. In October 2025, WinActor, a leading RPA product in Japan, released an AI integration license that lets a scenario give instructions to generative AI.

Existing robots keep working. RPA’s weakness, stopping when a screen changes, remains.

3. AI builds and repairs RPA

You give instructions in words and AI builds the RPA workflow. UiPath Autopilot builds Studio workflows and test cases from natural language. Copilot in Power Automate for desktop can also build desktop flows from natural language, but flow building is a preview feature, available only in environments in the United States and in English. For screen operations, Copilot places a marker, and a person records the steps with the recorder.

On the repair side, features that let AI recover an operation during a run, when a screen change would stop it, have arrived as UiPath Healing Agent and Power Automate Desktop self-healing (preview).

Building takes less work. But what you get is a conventional RPA robot, and it works the same way as before. With both recovery features, a recovery during a run does not carry over to the next run, and a lasting fix is left to a developer.

4. An AI agent operates the screen every run

The screen operations themselves go to an AI agent, such as ChatGPT, Claude, or Copilot Studio. AI reads the screen and thinks on every run, so it rarely stops when a screen changes, and nobody has to write out the steps in detail.

In exchange, it is slow, because every operation waits for an AI response; each run costs a usage fee; and the results are not always the same. Gartner says that for processes run at high frequency, an AI agent can cost up to 30 times as much as an RPA robot. It suits one-off work requested on the spot.

5. AI learns an operation and repeats it (AI RPA)

AI reads the screen and learns the operation, and later runs repeat the learned steps the way RPA does. Only when the learned steps no longer work does AI read the screen again and relearn the operation, which later runs then repeat. Descarty calls this shape AI RPA.

While it repeats learned operations, it doesn’t wait for AI, so it is fast, and it costs no AI usage fee. Steps that read documents or make judgments use AI on every run. It keeps RPA’s speed and the certainty of the same steps every time, while robots stop less on screen changes and are easier to fix.

Comparing the five

Approach	When a screen changes	Who builds and fixes it	AI usage fee per run
1. AI reads, RPA enters	Entry stops	RPA developers	Only for reading
2. RPA calls generative AI	Stops	RPA developers	Only for the steps that call it
3. AI builds and repairs RPA	May recover during a run, but the recovery doesn’t carry over to the next run	AI drafts; developers finish and fix	None on normal runs; used for recovery
4. An AI agent operates every run	Searches again on the spot	AI copes on the spot; catching its mistakes needs a separate check	For every operation
5. AI learns and repeats (AI RPA)	AI relearns the operation, and later runs repeat it	The people who do the work build and fix it in plain words	None for repeating learned operations

Examples of combining them

Work	What AI does	What runs as set steps	How often the screens change
Invoice processing	Reads issuer, amount, and due date from invoices in each supplier’s format	Matches them against purchase data and enters them in the accounting system	The in-house accounting system rarely changes
Entering purchase orders	Reads Excel purchase orders that arrive by email	Enters the orders in a sales system with no API	Changes with each system update
Routing inquiries	Reads each inquiry, classifies it, and drafts a reply	Passes it to the right person and logs it in the inquiry list	Rarely changes
Taking in orders from Web-EDI	Relearns operations when a screen changes	Signs in to each retailer’s Web-EDI, takes in the orders, and enters them in the sales system	Other companies’ screens change without notice

For entry into in-house screens that rarely change, approaches 1 and 2 are enough. For daily work on other companies’ screens, such as retailers’ Web-EDI, approach 5, AI RPA, fits. The same flows, with estimates of the effect, are in the use cases Read invoices into the accounting system, Read Excel purchase orders from email and enter them, Classify inquiries and draft replies, and Take in orders from retailers’ Web-EDI.

How to choose for each kind of work

• Fixed operations in volume on in-house systems whose screens rarely change: keep RPA. If there is reading involved, add AI with approach 1 or 2
• Reading paper, PDFs, or email bodies and entering what they say: give the reading to AI (approaches 1, 2, and 5)
• Operating other companies’ web systems or SaaS screens every day: the screens change without notice, so approach 5, AI RPA
• One-off research or operations requested on the spot: approach 4, an AI agent
• The other system offers API or CSV integration: use the official integration rather than the screen

How to split the work step by step within a process is in AI agents vs RPA, and how RPA, AI agents, and AI RPA differ is in RPA problems, and how AI RPA fixes them.

Frequently asked questions

Which is better, RPA or AI?

They are good at different things, so they aren’t really rivals. RPA is good at repeating set steps quickly and the same way every time; AI is good at judgments such as reading documents in any format or classifying by content. Most work involves both, so combine them step by step.

With AI, do we still need RPA?

For now, yes. Gartner, too, says AI agents will not replace RPA for the time being. Having AI think through every run of work done in volume every day is slow and costly. What changes is how RPA is built and fixed. Where RPA is heading is covered in Is RPA dead?

What is AI RPA?

RPA in which AI reads the screen and learns the operation, and later runs repeat the learned steps. Only when a changed screen blocks the way does AI relearn the operation. It keeps RPA’s speed and repeatability, stops less on screen changes, and can be fixed in plain words.

Can AI build RPA robots?

Yes. UiPath Autopilot and Copilot in Power Automate for desktop (preview) build RPA workflows from natural-language instructions. But what they build are conventional RPA robots, which still tend to stop when a screen changes. With AI RPA, AI doesn’t just build the automation; it also relearns operations when a screen changes.

Where should we start combining generative AI and RPA?

Pick one process still done by hand and split it into steps: reading, judging, and entering. Give the reading and judging to AI, and the entering to set steps. If one of your RPA robots keeps stopping, starting with that robot is the shortest path.

RPA and AI in one workflow with Kitewell

Descarty’s Kitewell is integration and automation middleware that connects systems through APIs, files, databases, and screens. It works the fifth way, as AI RPA.

• Screen operations learned and repeated: write the steps in words, and AI reads the screen, learns the operations, and repeats the same steps on later runs. AI relearns only when a screen changes, and the run’s record shows what it relearned
• Reading and judging in the same workflow: AI reading documents and making judgments, batch processing of Excel lists, API calls, file and database operations, and human approval sit in one workflow
• You choose the AI model: a cloud API such as Azure OpenAI, Amazon Bedrock, or Vertex AI, or a model you run in-house
• Rebuild your existing robots: describe the steps of a UiPath, WinActor, BizRobo!, or Power Automate Desktop robot in words, and AI builds the workflow again on the real screens

Start with one manual process or one robot that keeps stopping. A demo shows it at work, and a free connection test confirms Kitewell reaches your systems. You can also try Kitewell free for 14 days.

Summary

• RPA repeats set steps quickly and the same way every time; AI reads documents and screens and judges. Most work involves both
• The AI combined with RPA comes in three kinds: AI-OCR, generative AI, and AI agents
• There are five ways to combine them: AI reads and RPA enters, RPA calls generative AI, AI builds and repairs RPA, an AI agent operates every run, and AI learns an operation and repeats it (AI RPA)
• In-house systems whose screens don’t change can stay on RPA. For daily work on other companies’ screens, AI RPA fits

Sources

• MM Research Institute: RPA usage in Japan 2024 (July 24, 2024; in Japanese)
• NTT Advanced Technology: announcement of the AI integration license (September 4, 2025; in Japanese)
• UiPath: Autopilot for developers
• Microsoft: Get started with Copilot in Power Automate for desktop
• UiPath: What is Healing Agent?
• Microsoft: Self-healing for desktop flows (preview)
• ITmedia @IT: AI agents will not replace RPA for the time being (January 16, 2026, contributed by Gartner, in Japanese)