The useful distinction is not old versus new. RPA fails when the work requires understanding; agents fail when the work requires exactness. Most real processes contain both, which is why the productive question is which step is which.
Rules first, model only where judgment is needed
A workflow should use the cheapest mechanism that is correct. Deterministic routing, validation and system updates belong in rules — they are faster, cheaper and testable. Reserve the model for the step where a human would otherwise have to read something and decide. Workflows built this way have far fewer places to go wrong, and each of those places is identifiable.
Both need a defined failure path
The difference shows up when something is unexpected. Classic RPA stops, and the queue silently grows until somebody notices. A well-built agent workflow escalates with context to a named owner. Neither is safe by default: whichever you choose, the exception path and the person who owns it have to be designed on purpose.
For the related distinction between agents, chatbots and plain workflow automation, see the guide linked below.
01Real scenario
An insurer receives emailed documents, extracts fields and enters them into a legacy system with no API. Language understanding suits an agent step; deterministic screen entry may still require RPA.
02Inputs and evidence
Collect representative emails and documents, target-system screens, validation rules, exception history and the cost of a wrong write. Separate unstructured judgement from repeatable clicks before choosing technology.
03Actions to test
Use AI only to classify or extract into a strict schema, validate with deterministic rules, then let RPA perform the stable legacy-system entry. Test layout changes, missing fields and low-confidence extraction explicitly.
04Security and controls
Set confidence thresholds, preserve source documents, log every field and require approval for sensitive records. RPA credentials stay restricted, and no model receives permission to improvise screen actions.
05Pilot KPI
Measure field accuracy, straight-through completion, review workload, screen-break incidents, recovery time and cost per processed case. Report AI and RPA failures separately so one layer cannot hide the other.
06Errors and human hand-off
Queue low-confidence documents, validation mismatches and changed screens for a person with the source and partial work attached. Stop the writer when layout drift appears instead of allowing repeated clicks against an unknown interface.