What is RPA?
It is the reason a nightly job fails the morning after a vendor moves a button three pixels to the left.
RPA
robotic process automation
Robotic process automation, or RPA, is software that repeats a sequence of clicks, keystrokes and screen reads that a person recorded once. It operates the interface the way a human would, rather than calling the system underneath it directly.
You point an RPA tool at a screen and record a task: log in, open a record, copy a field, paste it elsewhere, click submit. The bot replays that sequence on a schedule or a trigger. Nothing about the process is understood by the software. It follows coordinates and labels, not logic.
This is a workaround for a missing integration, not a design choice. A system with a documented API gets called directly, every time. RPA exists for the older system that never got one. You still find these behind a login form: a desktop application, a mainframe terminal, a portal with no export button.
Durable where there is no API, brittle everywhere else
A bot that reads a fixed layout on a system nobody touches can run for years without a single change. That is the honest case for RPA, and it still holds in 2026. Some vendor software will never ship an API. A scheduled bot beats a person doing the same ten clicks every morning.
The failure mode sits in the same sentence as the strength. A bot has no idea what the screen means, only where things sat when it was recorded. Move a field or add a confirmation dialog, and it clicks the wrong spot. Push a routine software update, and it times out waiting for something that no longer appears.
We have taken over automation built this way. One screen-scraping bot ran fine for months, then failed silently after a vendor release changed one label. Nobody noticed until the numbers downstream stopped matching. Where a direct data connection existed underneath, replacing the scrape with it removed that failure mode instead of just making it easier to spot.
- 01Records a person's clicks once, then replays the exact sequence.
- 02Reaches systems that expose no API, by operating the interface directly.
- 03Breaks on layout changes it cannot detect, because it never understood the screen.
- RecordA person performs the task once, on screen.
- MapClicks and fields are saved as fixed steps.
- TriggerA schedule or event starts the run.
- ReplayThe bot repeats the recorded steps exactly.
- BreakA moved button or new field stops the run.
Nothing in this sequence reads the screen's meaning. It reads its coordinates, so the last step is a when, not an if.
Common questions
01Is RPA dead now that AI agents exist?
No. RPA still fits a fixed, structured task on a system with no API, and an agent adds cost without adding value there. Agents took over the work RPA was always bad at: reading messy, unstructured input like email or free-text documents.
02What is the difference between RPA and an AI agent?
RPA replays a recorded sequence of clicks with no understanding of what the screen shows. An AI agent reads the content and decides what to do next. One is deterministic and blind to meaning; the other adapts, at the cost of being harder to predict run to run.
03Should we migrate an existing RPA bot to an AI agent?
Only the ones that keep breaking. A bot that has run unchanged for two years is doing its job. Migrate the bots where someone rewrites a recorded step every month, or where the underlying task genuinely needs judgement a fixed script cannot make.
04Why does RPA break so often?
Because it operates on the visual layout of a screen, not on the data behind it. A routine vendor update can move a button, rename a field, or add a dialog box. The bot has no way to notice that the meaning survived while the coordinates changed.
