If you've looked into automating a back-office process, you've probably met two terms: robotic process automation (RPA) and AI agents. Vendors often talk about them as if one replaces the other. In practice they solve different problems, and choosing the wrong one is the most common reason automation projects stall.
What RPA is good at
RPA is software that repeats the exact clicks and keystrokes a person would make: open this screen, copy this field, paste it there, press save. It is fast, predictable and cheap to run when three things are true:
- The input always looks the same, such as a fixed export file or a structured web form.
- The rules never need judgment: if X, do Y, every time.
- The screens it works on rarely change.
The weakness is brittleness. When a supplier changes their invoice layout, a website moves a button, or an email arrives in a slightly different format, a script doesn't adapt. It stops, or worse, it types the wrong value into the right field.
What AI agents add
An AI agent uses a language model to read messy input, like emails, PDFs, scans and free text, and decide the next step within rules you set. Instead of clicking through screens, it usually works through the systems' APIs. A well-built agent can:
- Read an invoice in any layout and pull out the supplier, amount and purchase order.
- Notice when something doesn't match, like a missing due date, and flag it instead of guessing.
- Route a request to the right person with a short summary attached.
- Hand the decisions that matter to a person for approval.
The trade-off is that agents are probabilistic. They can be wrong, so they need an evaluation set drawn from your own examples, logging of every action, and a human approval step wherever a mistake would be expensive.
A quick way to decide
- Look at 20 real examples of the input. If they're all structured and identical, RPA or a simple integration is probably enough.
- If the examples vary in layout, wording or quality, you need something that can read, which means an AI agent.
- Ask what happens when it's wrong. If the cost is high, design a human approval step from day one, whichever tool you choose.
- Check whether the systems have APIs. Working through APIs is more reliable than working through screens, for both approaches.
Why most real processes use both
Take accounts payable. Pulling a daily export from the ERP is a perfect job for a simple script. Reading 200 supplier invoices in 40 different layouts is a job for an AI agent. Deciding whether to pay an invoice that doesn't match its purchase order is a job for a person, with the agent preparing everything they need to decide quickly.
What we'd tell a friend: Start with one boring, high-volume process, measure how long it takes and how often it goes wrong today, and automate that first. You'll learn more from one working pilot than from a year of planning.