Robotic Process Automation (RPA)

Robotic process automation (RPA) is software that automates repetitive, rule-based computer work by having software bots operate applications the way a person does: opening programs, clicking buttons, typing into fields, copying data between systems, and following the same steps every time. Typical jobs include invoice entry, account reconciliation, order processing, employee onboarding, and report generation, often in older systems that offer no API. The robots are entirely software and have nothing to do with physical robotics. Blue Prism, founded in the UK in 2001, is widely credited with coining the term, and UiPath, Automation Anywhere, and Microsoft Power Automate are among the most widely used platforms.

How RPA Works#

Developers build a bot in a visual designer, often by recording a person's actions and then editing the steps. To find each button or field, the bot uses a selector, a stored description of the element's attributes such as its window, control type, and name, and falls back to screen coordinates or image matching where no such attributes exist. Attended bots run on an employee's desktop when that person starts them, such as filling a form during a customer call. Unattended bots run on servers on a schedule or from a work queue, managed by an orchestrator that assigns jobs, stores credentials, and logs every action for audit:

The same precision makes bots brittle. A renamed field, a moved button, or a new pop-up breaks the selector, and the process stops until a developer updates it, so large RPA deployments carry steady maintenance work as the underlying applications change.

RPA and AI#

Rules handle structured data; anything messier needs machine learning in front of the bot, a combination vendors call intelligent automation and Gartner calls hyperautomation.

• Document understanding: Optical character recognition (OCR) and extraction models turn scanned invoices, receipts, and forms into structured fields that a bot enters into the finance system, a common step in document analysis pipelines.
• Computer vision for screens: Virtual desktops such as Citrix often expose only an image of the remote screen, so selectors cannot see its controls. UiPath's Computer Vision activities send an image of the window to a neural network that identifies and labels every button, text field, and checkbox, which is object detection applied to user interfaces.
• Language models and agents: Large language models (LLMs) classify emails, draft replies, and resolve the exceptions that used to stop a bot, while bots keep the deterministic steps. Vendors now market this mix as agentic automation, with AI agents deciding what to do and bots doing it.

Visual inspection results can start these automations too. In Ultralytics Platform Agents, a Gate condition on a YOLO26 result, such as a detection count or a confidence range, can feed a Webhook block that sends one JSON POST per matching result to a public HTTPS endpoint, so a defect detection can start a flow in any automation tool that accepts incoming web requests.

RPA vs. Related Terms#

• Agentic workflows: Multistep processes in which LLMs plan, call tools, and check results, handling the unstructured inputs that scripted bots cannot.
• Computer use agents (CUAs): AI agents that operate software from screenshots by deciding each click as they go, so a moved button does not break them. They run slower and less predictably than a recorded RPA script, which repeats exactly the same steps on every run.
• API integration: Connects systems directly through their programming interfaces, which is faster and survives interface redesigns. RPA fills the gaps where no API exists or building one would take too long.