Companies that bought RPA five years ago are now being sold “AI agents” as the upgrade. Sometimes that's right; often it's expensive overkill. The distinction is simple and worth getting precise.
The one-question test
Does the workflow require judgment on unstructured input? If every run is identical — same fields, same screens, same rules — RPA or a plain script is cheaper, faster and more reliable than any LLM. Deterministic automation doesn't hallucinate. If the input varies (free-text emails, scanned documents, edge cases that currently land on a human's desk), rules break down and you need a system that can read, interpret and decide. That's the agent's territory.
Where RPA quietly fails
RPA breaks when the surface changes: a UI redesign, a renamed column, a vendor who formats invoices differently this month. Every variation becomes a maintenance ticket. Agents absorb surface variation because they read meaning, not pixel positions — the new invoice layout is still an invoice.
The pattern that actually wins: agent decides, automation executes
The strongest deployments we've shipped are hybrids. The agent does the judgment work — classify the document, extract the fields, decide the route, flag the exception. The execution layer is boring, deterministic code: API calls, database writes, posted entries. Judgment where judgment is needed, determinism everywhere else. This is also the cheapest architecture to run, because the LLM is invoked only at the decision points.
How to sort your backlog
List the workflows you want automated. For each, ask: is the input structured or messy? Do exceptions reach a human today? Would a wrong decision be expensive? Structured + rare exceptions + cheap errors → script it. Messy input or meaningful exceptions → agent with an execution layer. Expensive errors → agent with human approval on the risky tier. Most companies discover their backlog is a mix — which is exactly why the hybrid pattern exists.