Call · 30 min
AI systemsMattia Esposito8 November 20266-minute read

RPA or AI agent. Fixed rules or text to understand, and the choice follows from that.

RPA is a software robot that carries out rule-based tasks in place of a person, clicking and copying as they would. An AI agent does something else: it reads text that differs every time and decides the steps. Here is the difference, and a sheet for choosing.

In brief

RPA, Robotic Process Automation, automates repetitive rule-based tasks; an AI agent uses a model that interprets texts and documents and decides which steps to take. Microsoft’s Power Automate documentation describes RPA in exactly these terms: automating repetitive, rule-based processes.

Under the AI Act the difference is also a legal one. Recital 12 of Regulation (EU) 2024/1689 excludes from the notion of an AI system those based on rules defined solely by people; an agent that infers from its input how to respond falls within the definition in Article 3.

Below is the eight-question sheet in Word and PDF, in Italian and English, with an example on two tasks, to download without leaving an email address.

Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.

What is RPA

RPA stands for Robotic Process Automation: software that carries out on screen the steps a person would, opening a program, copying a value, pasting it into another, saving. The “robot” is a program that imitates clicks and typing on the screen.

Microsoft describes Power Automate desktop flows as the tool that extends its RPA capability and automates repetitive desktop processes, designed for people who carry out rule-based tasks. This is both the strength and the limit of RPA: it works as long as rules are enough and the screen doesn’t change.

RPA or AI agent: the difference in ten rows

The difference lies in what they take in: RPA works on tidy data and stable screens, an AI agent on text written by people. Everything else follows from this, from what happens when they get it wrong to which rules apply.

CriterionRPAAI agent
How it works

Repeats steps written as rules.

A model reads the input and decides the steps.

What comes in

Data always in the same format.

Emails, documents, messages that differ every time.

If the screen changes

It stops or fills the wrong field.

It works on the content, not on where the buttons are.

How it goes wrong

Repeats the same error, which is soon spotted.

Can give a plausible but wrong answer.

Oversight

Spot checks on the results.

A person approves anything that goes outside.

Under the AI Act

Usually not an AI system.

An AI system, with the rules that go with it.

If it talks to people

It usually talks to nobody.

It must say it is a system, Article 50.

Who uses it

Needs training on the flow.

Also needs training on AI, Article 4.

An example

Copying orders from a portal into the business software.

Reading requests by email and preparing the offer.

When to choose it

Stable task, tidy data, few rules.

Text to understand, too many rules to write.

The sheet for choosing, to download

Eight questions about the task to automate: the first four point to the tool, the other four say which rules apply and who checks. The example runs two tasks from the same fictitious company through the sheet and arrives at two different choices.

Eight-question sheet for choosing between RPA and an AI agent, with the AI Act rule on rule-based systems
FileContentsLink
ItalianWord, to fill in

The eight questions, each with its pointer.

rpa-o-agente-ai-scheda-it.docx

EnglishWord, to fill in

The same sheet in English.

rpa-o-agente-ai-scheda-en.docx

ItalianPDF

The copy to print.

rpa-o-agente-ai-scheda-it.pdf

EnglishPDF

The English sheet, ready to print.

rpa-o-agente-ai-scheda-en.pdf

ExamplePDF, filled in, in Italian

Two tasks from the same company: one goes to RPA, the other to an AI agent.

rpa-o-agente-ai-esempio-it.pdf

ExamplePDF, filled in, in English

The same example, in English.

rpa-o-agente-ai-esempio-en.pdf

What the AI Act says about rule-based systems

The AI Act defines an AI system as one that, for explicit or implicit objectives, “infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions”: Article 3(1) of Regulation (EU) 2024/1689. The word that matters is “infers”.

Recital 12 says it from the other side: the definition “should not cover systems that are based on the rules defined solely by natural persons to automatically execute operations”, and it names the capability to infer as a key characteristic of AI systems. An RPA robot following rules written by people usually stays outside.

An AI agent falls within the definition, and its rules apply: the training duty in Article 4, explained on the page about the AI literacy obligation, and the transparency duty in Article 50 when it interacts directly with people. If an RPA flow calls a model to read a document, that part of the flow should be treated as an AI system.

When to use one, when the other

RPA makes sense when the task is stable, the data is tidy and the rules are few: copying orders from a portal, downloading bank statements, filling in a form that never changes. It is quick to get running and predictable, until someone changes the screen it reads from.

An AI agent makes sense when you have to read what people write: quote requests, customer emails, documents from different suppliers. There, the rules to write would be too many, and would change with every customer. The price is a different kind of error, plausible rather than repeated, which is managed with a person who approves: the way of working called human in the loop.

They often work together. The agent reads the email and prepares the data, a person approves it, and RPA enters it into a program that has no other way of receiving it. According to Istat, Enterprises and ICT 2025, 70.8% of businesses that use AI apply it to extracting knowledge and information from text documents: the part RPA can’t do on its own. The difference between an agent and a chatbot is on the page about AI agents and chatbots.

Questions and answers

What is RPA?

RPA stands for Robotic Process Automation: software that carries out repetitive, rule-based tasks in place of a person, such as copying data from one program to another, clicking and typing as the person would.

It works well when the data always arrives in the same format and the screens don’t change.

What is the difference between RPA and an AI agent?

RPA follows rules written by people and always repeats the same steps. An AI agent uses a model that reads texts and documents that differ every time and decides which steps to take, pausing where a person is needed.

The first suits tidy data, the second suits text written by people, such as emails and requests.

Is RPA artificial intelligence?

Usually not. Recital 12 of the AI Act says the definition of an AI system should not cover systems based on rules defined solely by natural persons to automatically execute operations.

If an RPA robot uses a model that infers from its input how to respond, that part falls within the definition in Article 3, and its rules apply.

When is an AI agent better than RPA?

When the task means reading and understanding text written by people, such as emails, requests or documents in different formats, and when there would be too many rules to write.

Where the result goes to a customer or involves money, a person approves it before it goes out.

Can RPA and an AI agent work together?

Yes, and often: the agent reads the email and prepares the data, a person approves, and RPA enters it into a program that has no other way of receiving data.

Each does the part it is suited to, and the boundary is decided beforehand, task by task.

Notes on sources

  1. Regulation (EU) 2024/1689, the AI Act, on EUR-Lex: Article 3(1) (definition of an AI system), recital 12, quoted from the official English text, and Articles 4 and 50.
  2. Microsoft Learn, Introduction to desktop flows in Power Automate, read on 8 November 2026.
  3. Istat, Imprese e ICT, 2025: the AI technologies most used by businesses with at least ten employees that use AI.
  4. The sheet and the example are built by Itria; the company in the example is fictitious.
·The next step

Rules repeat. Reading and understanding is an agent’s job, with a person who approves.

When the task starts from text written by a person, an AI agent can read it, work out what it asks and prepare the reply or the data, leaving them for a person to approve before they go out. The piece that reads documents recognised the right type thirty times out of thirty and got 98.7% of fields right, in 6.3 seconds each: measured on a test bench on 4 September 2026, on a test system.