These three terms get used as if they mean the same thing, usually because all three promise the same outcome: less manual work. They solve different problems, and picking the wrong one is a common reason automation projects stall after the demo.
RPA: replaying a fixed sequence
Robotic process automation records a sequence of clicks and keystrokes against a specific screen layout, then replays it. It is fast to build and genuinely reliable for the exact path it was recorded on. It has no understanding of what it’s doing: change a field position, add a new case, or feed it an input the recording didn’t anticipate, and it breaks rather than adapts.
Workflow automation: routing structured data
Workflow automation (think Zapier-style tools, or a custom pipeline moving data between systems) is a step up: it moves structured data along a defined path with conditional branches. It’s a good fit when the shape of the input is predictable and every branch of the decision tree can be enumerated ahead of time. It still can’t handle a request written in a customer’s own words, or a case that falls between the branches someone thought to build.
AI agents: deciding, not just routing
An AI agent reads a request, decides what needs to happen, and can reach into the software a business already runs to carry that decision through to a finished outcome. The difference from the two categories above is judgment: an agent handles the request that doesn’t match any pre-built branch, because it’s reasoning about the content of the request rather than matching it against a fixed pattern.
That judgment is also why an agent needs different safeguards than a script does. A script either runs or it doesn’t; an agent can be confidently wrong, which is why the systems worth building keep a human in the loop at the decision points that matter, rather than running every action fully autonomously.
How to actually choose
Start by asking whether the inputs are structured and the paths are fully known. If yes, RPA or workflow automation is cheaper and more predictable, and reaching for an agent adds cost without adding capability. If the inputs are messy, the requests come in free text, or a meaningful share of cases fall outside whatever branches you can draw today, that’s the signal an agent is worth building. Most production systems end up as a mix: scripted steps for the deterministic parts, an agent making the calls that actually require judgment.