What Is Robotic Process Automation (RPA)?

How it works

An RPA bot follows a defined sequence of steps in the same applications a person would use. You build that sequence in an RPA tool, often with a visual designer where you drag in actions such as “open application,” “read cell,” “type into field,” and “click Submit,” or by recording yourself doing the task once and then refining the recording.

To know where to click or type, the bot identifies elements on the screen by their underlying properties, such as an ID or label, or sometimes by their position or appearance. Some tools also use optical character recognition (OCR) to read text that appears only as an image, such as a screen streamed through remote-access software or a scanned page.

Bots usually run in one of two modes:

• Attended: the bot runs on an employee’s computer, usually when they start it or take a specific action, handling part of a task while they handle the rest.
• Unattended: the bot runs on its own, on a schedule or when triggered, typically on a company server or a cloud-hosted computer, and is usually monitored from a central dashboard that shows results and errors.

Either way, the bot does exactly what it was told to do. It does not understand the content it is moving.

When RPA is a good fit

RPA fits best when a task is:

• Rule-based: the same inputs always lead to the same actions, with no judgment calls.
• Repetitive and high volume: people do it many times a day or week.
• Built on structured, digital data: values sit in known fields, spreadsheets, or forms, not in free-text emails or PDFs with varied layouts.
• Done in stable applications: the screens and steps rarely change.

Its main practical strength is reaching systems that offer no other way in. Many older or vendor-hosted applications have no API, or one that is limited or costly to use. Because a bot works through the same screens a person does, it can move data in and out without changes to the underlying system. It is also predictable: given the same input and the same screens, a bot takes the same steps every time, and RPA tools can typically log each action, which helps with auditing.

Common examples include keying approved invoice details from a spreadsheet into an accounting system, copying order details between an online store and a warehouse system, updating customer records in two places, and downloading reports from a portal on a schedule.

Limitations to know

The biggest weakness of RPA is fragility. A bot depends on the screens it was built for, so a moved button, a renamed field, a new pop-up, or an extra login step can stop it or, worse, cause it to enter data in the wrong place. Plan for ongoing monitoring and maintenance, especially for web portals you do not control, which can change without notice.

On its own, RPA cannot interpret unstructured input such as a customer’s email, a handwritten form, or a contract, and it cannot make judgment calls. Anything outside its rules must be detected and routed, usually to an exception queue for a person. That handling has to be built in, or the bot may simply stop or make an error.

A few other points to weigh:

• It automates the process as it is. If the steps are inefficient, the bot repeats them. It often makes sense to simplify the process first.
• Bots need logins and permissions. Treat them like user accounts: grant only the access they need and review it regularly.
• Some websites restrict automated access in their terms of use, so check before automating a third-party portal.

RPA vs. API integration vs. AI automation

RPA is one of several ways to automate work, and in practice the approaches are often combined rather than chosen between.

• RPA works through the user interface, imitating a person at a keyboard. It is typically the fallback when a system has no usable API.
• API integration connects systems directly through interfaces built for software to talk to software. When a usable API exists, it is generally faster and less fragile than a bot clicking through screens, because it does not depend on how a page looks.
• AI automation uses AI models to handle what fixed rules cannot, such as reading a free-text email or pulling fields from documents with varied layouts. See AI automation for how it compares with rule-based automation.

In a combined AI workflow, an AI model might extract details from an incoming document. API calls then update the systems that offer an API, and an RPA bot enters the rest into an older application. Anything that fails a validation check or falls below a set confidence threshold goes to a person for review, an approach known as human-in-the-loop. Some AI agents can also operate a screen the way RPA does. They can handle some variation in what they see, but they are less predictable than a scripted bot, so their actions need closer checking.

Further reading

• AI ImplementationCovers SN3’s work redesigning processes, connecting systems, and automating work, relevant if you are weighing RPA against other ways to automate.