RPA explained: robotic process automation and when it makes sense

RPA, the acronym of robotic process automation, indicates software that uses other programs like a person sitting on a computer: it opens the management system, reads a value on a supplier's portal, writes it in an Excel sheet, presses Save and moves to the next line. The robot has no arms or sensors. It is a program that repeats a sequence of actions on the interface, established by who configures it.

It serves when two programs have to exchange data and there is no direct link between them: no API, no useful export, only screens. In an SME it often happens, with management software installed many years ago, portals of large customers, programs opened in remote desktop. Where a direct link there is, instead, RPA is almost always the most fragile road.

The comparison to artificial intelligence-based automation is in the article RPA vs AI automation.

What is RPA (robotic process automation), no pun intended

Robotic process automation means process automation with software robots. Microsoft presents Power Automate desktop streams just like RPA features: they automate rule-based repetitive tasks, even on legacy applications such as terminal emulators, and interact with the computer through interface elements, images, or screen coordinates (Microsoft Learn).

The difference with a classical integration lies in the front door. Two integrated programs speak through a channel designed for machines; the robot enters the employee's door, the screen, and so also works on software never designed to dialogue with others.

The robot, however, does not read or interpret. If the “Total” field is empty, it copies the empty value; if a document different from the usual one arrives, it stops or makes a mistake. Everything it can do was programmed by someone beforehand.

How an RPA robot works: selectors, recorders, exceptions

To press the right button the robot must first find it. RPA platforms use the selectors, item descriptions (type, name, window that contains it) saved during configuration. UiPath documentation explains that recognizing an item by its on-screen location is possible but unreliable, and that with some web applications, where layouts and attributes change, selectors should be arranged by hand (UiPath, About Selectors).

A typical configuration goes by four steps:

1. Recording or drawing of the stream. You perform the operation with the recorder on, or you compose the stream with ready actions: open the application, read the cell, write in the field.
2. Selectors and Variables. Each click becomes an action on a recognized item, while the data that changes at each turn (order number, customer, amount) comes from a file or table.
3. Exception handling. You decide what to do when something goes wrong: window that does not open, missing field, session expired. Without this part the robot freezes at the first unexpected.
4. Execution and Log. The robot starts on demand, at time or when a file arrives, and notes every step: when it stops, the log says where and why.

RPA awaited and unattended: who starts the robot

In UiPath documentation automation awaited part at the user's initiative, usually asks for his interaction and runs on his computer; that unattended is managed by a central console and works independently, often in the background or out of hours, on a separate machine (UiPath, introduction to robots).

Unattended mode has requirements that are discovered late. For Power Automate, Microsoft indicates that the Process plan is needed, that the service opens a remote desktop session with the screen locked, and that on computers with Windows 10 and 11 the robot does not start if there is an open user session, even if it is locked (Microsoft Learn). Basically the robot needs a computer, or a virtual machine, all of its own.

When should RPA in an SME

Before leaving, verify that the process has these features:

• It repeats the same many times: same steps, same fields, same order.
• The starting data is already digital and sorted, such as an Excel or a table on a portal.
• Decisions are written as rules ("if the amount exceeds the threshold, it signals"), without having to interpret a text.
• The screens rarely change: the business management software is not updated every month, and the portal keeps the same structure.
• The software provider has confirmed that there are no APIs, webhooks or automatic exports.
• Someone who knows the process can follow the robot and replace it when it stops.

When one of these conditions is missing, the project gets complicated. Free texts and documents always different ask for a part of AI, screens that change often ask for continuous maintenance, an API available makes the robot superfluous. To choose the first process to work on, the guide on what to automate in an SME helps make the list.

RPA, API, n8n and Make: what a difference there is

The most common mistake is to put a robot to click where a direct call would suffice. A API is a channel designed for one program to ask for or send data to another without going through the screen. Tools like n8n work like this: the HTTP Request node, for example, queries any application or service that offers a REST API (n8n documentation). For Make and Zapier, and to choose between the three, there is the comparison Make vs Zapier vs n8n.

From the most solid to the most fragile link, the order to follow is this:

1. Native Integration between the two software, already offered by the supplier.
2. API or webhook, orchestrated with n8n, Make or custom code.
3. File Exchange per hour, for example a CSV deposited every night in a folder.
4. RPA on the interface, when none of the previous roads are available.

The first move is to ask the management software provider what it exposes. How it occurs, and what changes between webhooks, scheduled calls and exports, is told in the article on how integrate business management software with n8n. Often the answer takes the robot out of the way.

Where RPA breaks down: changing screens and credentials

An RPA robot depends on things you do not control. The business-software provider releases a version with a button moved, the client portal changes its design, or a new warning window appears: the selector fails to find the element and the workflow stops. Microsoft also points out a less obvious case: in unattended runs, the remote desktop session may have a different resolution from the one used to create the workflow, and the robot may fail to find an element or interact with the wrong one if it uses the mouse and keyboard (Microsoft Learn).

Then there are the credentials. The robot enters programs with a user account and a password, so it treats data like an employee would. The GDPR calls for risk-appropriate technical and organizational measures, including the ability to ensure confidentiality and integrity of systems (Art. 32 GDPR). For a robot it means a user of its own with the minimum permissions, the password in a protected archive and a register of executions available.

RPA, Artificial Intelligence and AI Agents: Where the Border Passes

Classical RPA is not artificial intelligence, and the AI Act explicitly states this: the definition of an AI system is not about rule-based systems defined solely by individuals to perform operations automatically (recital 12 of EU Regulation 2024/1689). A robot copying fields following instructions written by you is here. If it processes personal data, the rules of the GDPR still apply.

The picture changes when the robot is added a part that reads different PDFs each time, ranks an email or chooses the next step. Istat, in the survey of companies with at least 10 employees, counts the automation of AI-based workflows among artificial intelligence technologies: in 2025 it is used by 2.9% of companies (Istat, Enterprise and ICT 2025). When it's best to pair AI with a stream, and when RPA is enough, you find it in the comparison RPA vs AI automation.

With AI agents the boundary is another. An RPA robot always repeats the same sequence; an AI Agent receives an objective and chooses the steps on its own. Some models already use computers through screens, a mouse and a keyboard, such as Anthropic’s computer use: the documentation recommends a virtual machine with minimal privileges, no sensitive credentials and human confirmation before actions with real consequences (Anthropic documentation). Always check the official page: these functions change often.

Four examples of RPA in an Italian SME

The scenarios below are examples built to show where RPA works and where it doesn't.

The orders of representatives in a distribution company. A spare parts distributor receives orders from agents in Excel and copies them into the management system, installed on site and without API. A manned robot reads the file line by line, fills out the order mask and returns the number generated to the Excel; the employee who starts it checks only the lines with an unknown article code. On the subject is the guide to the data entry automation, with a demo in The Lab.

A big customer's portal. A small supplier to large retailers must download orders and confirmations every morning from its customer’s portal, which does not offer APIs to small suppliers. An unattended robot logs in before the office opens, downloads the documents into a shared folder and sends a message if the portal does not respond. The biggest risk is a redesign of the portal: every new release must be followed by a test.

The last stretch to the administrative program. A technical study collects requests from the site in a modern CRM, connected with n8n via API, but must also register each new order in the old administrative program, which has no other entrance doors. n8n prepares a file every evening with the orders of the day, the robot inserts them and reports the rejected lines. The fragile part remains confined in one passage.

Passive bills, where the robot is not needed. A company's administration thinks of a robot that downloads supplier invoices from a portal. The electronic invoices between Italian operators, however, pass from the Interchange System of the Revenue Agency, which receives the files, controls them and forwards them to the recipients, including private individuals, and dialogues with the software through accredited channels such as SdICoop and SdIFTP (billopa.gov.it). If the manager or accountant already receives them from there, the robot would be a useless turn. What you can automate after receiving it explains the guide on the Invoice automation with AI.

How much does RPA cost and what it takes to leave

A robot’s cost does not end with its licence. Include the time to map the process and configure the workflow, the machine it runs on if unattended, and especially maintenance, which lasts as long as the robot.

A few tools are enough for a test. Microsoft indicates that the use of Power Automate with a Microsoft account is available at no additional cost (Microsoft Learn); connecting to cloud streams, sharing, and centralized management requires an organization premium account instead (comparison between accounts). For dedicated RPA platforms it always checks the manufacturer's official list, because licenses and plans often change. For the total cost, including the part that does not stand in the license, there is the guide on how much it costs to automate a process.

Mistakes that make an RPA project fail

• Automating a process that no one has written down. If each employee makes the steps in their own way, the robot copies the version of who configured it, including errors. First you write the sequence, then you automate.
• Choosing the robot when there was an API. A direct integration does not depend on the graphics of the programs, so it requires less intervention when the updates come out.
• No one responsible. A robot with no name next to it stops silently: it needs someone to receive the alerts and try it after each update.
• The robot with the administrator's user account. Convenient in configuration, risky in operation: if the flow is wrong, it is wrong with all the permissions.

Where to start with the first robot

Choose the case closest to yours.

In any case, two pieces of information are needed: the sequence of the steps as it is done today by those who work there, and what each program involved exposes. With those, the choice between integration, files and robots becomes almost obvious.

If there is a job in the company made of screens and copy-paste, Giallo Studio starts from the step map and checks the direct roads first: the robot only enters the stretch where it really needs to be.