RPA's core idea is to record or script a sequence of UI actions — open this app, click this field, copy the value, paste into that other app — and replay it on schedule or on trigger. The 'robot' is a background process that has access to the user's screen and input devices (or runs headless in a virtual desktop). Unlike workflow automation which connects apps via APIs, RPA is app-agnostic: if a human can do it, an RPA bot can do it.
RPA emerged specifically to address the enterprise integration crisis of the 2010s. Large banks, insurers, and government agencies ran hundreds of legacy systems — mainframes, green-screen terminals, old client-server apps — that had no APIs and would cost millions to replace. Rather than rebuild, RPA vendors offered a shortcut: let bots drive the existing UIs. A bank could automate loan processing without migrating off its 1990s core banking system. This is why the big three RPA vendors (UiPath, Automation Anywhere, Blue Prism) grew to billion-dollar valuations — they monetized the enterprise technical-debt overhang.
Technically RPA has two flavors. Attended RPA: a bot runs on a human's desktop and augments their work — for example, opens all the apps needed for customer onboarding when the rep starts a call. Unattended RPA: a bot runs headless in a virtual desktop (usually a Windows VM), processes queues of work 24/7 with no human involvement. Unattended is where RPA produces the big labor savings and is what most enterprise programs deploy.
The fundamental trade-off: RPA is powerful but brittle. A bot trained to click a specific button at specific coordinates breaks when the UI changes — a new field added, a redesigned dialog, a different Windows theme. Enterprises with large RPA portfolios spend significant effort on maintenance: monitoring for failures, updating scripts after UI changes, handling 'exception' flows the bot wasn't designed for. The industry phrase is 'RPA tech debt'.
2020-2023 saw RPA vendors add AI and ML to make bots more adaptive: computer vision to find buttons by appearance rather than coordinates, NLP to classify documents, ML to predict when a bot will fail. 2024-2026 is the LLM integration wave — UiPath AI Center, Automation Anywhere Co-Pilot, Blue Prism Bots+LLM — which is gradually shifting RPA from scripted action sequences to reasoning agents that can handle UI changes on their own. This convergence with AI agent tech is sometimes called 'intelligent automation' or 'agentic RPA'.
The new entrants — Anthropic's computer use, OpenAI's Operator — are effectively 'LLM-native RPA': no recording, no scripts, just ask the model to do the UI task in English. For small businesses, these are already cheaper and more flexible than traditional RPA. For large enterprises with thousands of bots, the transition is more gradual — LLM costs at enterprise scale still exceed scripted RPA costs, and reliability isn't yet at the 99.9% level required for critical workflows. Expect the two approaches to merge over 2026-2027.
For small businesses and solo founders, traditional RPA is usually overkill. A Tycoon
AI employee with API tools (via MCP or Composio) plus occasional Claude computer use for edge cases covers 95% of what a small business needs, without the implementation complexity of a UiPath program.