They get mixed up constantly in meetings. "We want to add AI" usually means "we want this to happen by itself", and quite often what is actually needed is old-fashioned automation, which costs a quarter as much and fails far less.
The difference fits in one sentence: RPA follows rules, AI interprets.
What RPA is and when it is enough
Rule-based automation — what people call RPA — executes predefined steps. If this happens, do that. Copy this field from here to there. When a file arrives in this format, process it like so.
It is the right answer when the process meets three conditions:
- Inputs always arrive in the same shape.
- The rules can be written down without using the word "depends".
- The result is verifiable: either right or wrong, no shades of grey.
Examples: moving orders from email into the ERP when they always arrive in the same format, generating the Monday report from three sources, alerting the manager when stock drops below a threshold. Cheap, fast and — most importantly — predictable: it does exactly the same thing every time.
What AI adds and when it is worth it
Artificial intelligence comes in when the input is unstructured or the rule cannot be written. It is not that it is "more powerful": it solves a different problem.
- Interpreting free text. Understanding what a customer email is about when they write however they like.
- Extracting data from documents. Pulling the taxable base out of invoices that arrive in forty different layouts.
- Classifying without an explicit rule. Deciding whether an incident is urgent when nobody can quite articulate what makes it so.
- Predicting. Estimating which product will run out or which customer is about to leave, based on history.
In exchange, AI has a property you have to accept from day one: it does not always get it right, and it does not get it right the same way every time. A good classification system might run at 90-95% accuracy. That is excellent, and it still means that out of every hundred cases, five to ten need a person.
If a process tolerates no unsupervised errors, the question is not which model to pick: it is that the process needs a human in the loop, whatever it costs.
The usual answer: both
In practice almost every project we deliver combines the two. AI handles the part that requires interpretation, and rule-based automation handles the rest, which is usually most of it.
A real example of the structure: an email arrives with an invoice attached. AI reads the PDF and extracts supplier, amount and date. A rule checks the supplier exists and the amount falls within their usual range. If everything adds up it is recorded automatically; if something is unusual or the extraction was unclear, it goes to human review with the reason flagged.
That split is what makes the system reliable: AI provides what rules cannot, and rules bound what AI can get wrong.
How to tell which case you are in
One question is usually enough:
Could you write down, on one page, the instructions for a new person to do this task without asking you anything?
If the answer is yes, you need rule-based automation and you should be paying rule-based automation prices. If writing them down produces sentences like "well, if it is one of our regular customers, then…", that is where AI belongs.
If you are not sure, write to us with the specific process and we will tell you which of the two it is. It is a ten-minute answer and it can save you several thousand euros of badly framed project.