Three technologies, three problems
The three are often presented as competitors. They are not. They solve different problems, and most working workflows use two or three together.
- Fixed rules (if-then logic, decision tables, validations) implement knowledge you can already write down explicitly. They are fast, free to use, fully predictable and fully testable.
- RPA operates a system as a person would: screens, buttons, fields. It exists because some systems have no usable integration.
- AI judges input you cannot define beforehand: free text, a document with an unknown layout, a conversation.
Criterion 1 — How variable is the input?
This is the clearest criterion. If you can capture input in a schema — fixed fields, values and order — processing belongs with rules. A model reading a fixed form adds nothing and introduces a chance of error you did not have.
When input becomes free text, or a document whose layout differs by supplier, rules reach their limit. You can write thirty rules for thirty layouts, and at the thirty-first you start again. That is when a model becomes cheaper than maintaining rules.
Criterion 2 — Is there an integration?
RPA is a fallback rather than a technology preference, and that is no criticism: sometimes it is the only route. If the system has an API or existing integration, use it. If it works only through the screen, RPA is the glue.
Allow for the familiar costs: an RPA script breaks at every interface change, runs under someone's account and is hard to test. Treat it as temporary infrastructure with an owner, not a finished solution.
Criterion 3 — Who must be able to explain the outcome?
If a regulator, accountant or client needs to follow why an outcome is what it is, prefer rules wherever possible. A rule is the explanation. A model produces an outcome you must substantiate afterward.
That argues not against AI, but for a division of tasks: let the model turn unstructured input into fields, and let a rule decide based on those fields. Then the judgment is explainable while input remains flexible.
The combination that usually wins
A working workflow often looks like this:
- A model reads unstructured input and produces fields, with an explicit uncertainty score.
- Rules validate those fields: does the supplier exist, is the total correct, is the amount within the authorization limit?
- A person approves where rules flag uncertainty or a deviation.
- An integration or RPA step executes the approved outcome in the target system, with a log of every step.
Mixing these four layers gives you a model that must remember thresholds and an RPA script that must interpret free text. Both fail in the most expensive way: silently.
When to use none of the three
When the process itself is wrong. If input becomes reliable as soon as someone adjusts the source form, an adjusted form is cheaper than every technology above. And if the process happens a few times a year, manual work is simply the cheapest solution — it is fine to say that aloud.