Faced with the same document they do two different things
Rule-based automation does what you told it to do, always in the same way. If the same invoice goes past it a thousand times, it records it the same way a thousand times
Artificial intelligence tries to work things out instead. It looks at a text it has never seen and pulls out what it reckons is the document number. The right words there are it reckons, because this is an estimate, not a certainty
They are two different tools for two different problems. Calling them by the same name is handy for the one selling, less so for the one buying
What rule-based automation does well
A rule is a sentence someone in your office could say out loud. If the supplier is already in the records use that one, if the lines do not add up to the total stop, if the amount goes over a certain threshold ask for a signature
The strength is not the speed, it is that you know in advance what is going to happen. You can read the rule, you can argue with it, you can change it, and if one day the system does something you did not expect there is a written line that explains why it did it
The limit is just as clear. A rule holds as long as the world around it stays the same, so if the supplier changes the form or the portal moves a button the rule no longer covers the case, and when it does not cover it the right behavior is to stop instead of guessing, as told in the guide on how you stop an automation from doing damage
What artificial intelligence does well
Artificial intelligence is useful exactly where a rule cannot reach, which is when the data arrives messy and has to be interpreted
A crooked scan with a stamp over the number, an email written by hand by a customer who orders in words instead of codes, a table with no gridlines inside a badly made PDF, the description of a bank transaction written a little differently every time. In cases like these there is no fixed position to read, there is a meaning to rebuild, and rebuilding meaning is exactly what these tools are for
Anyone who tells you artificial intelligence is never any use is overdoing it from the other side. Where the data is messy from birth, it is the road that works best
For accounting data you want the rule
When what is at stake is an accounting entry, a goods receipt, a price that ends up inside an order, what you want is not a brilliant system. It is a predictable one
A mistake in a rule shows up straight away and you fix it in one place. A mistake in interpretation is sneakier. It goes fine on a pile of documents in a row, then on one of them it takes one date for another and nobody notices, and there is no line you can read that explains why
The two go one after the other
The way they work together is less romantic than the way it gets told. Interpretation sits at the entrance, and the rule sits right behind it
The messy document is read as well as it can be, then what has been read goes through the checks. The supplier has to exist, the VAT number has to be formally valid, the lines have to add up to the total, the date has to make sense. Whatever does not pass the checks never enters the business software, it ends up in a list a person looks at
Interpretation can propose, the rule decides, and this holds even when the reading part is good, because trust is not given by eye, it is verified with a reconciliation that either adds up or does not. The same reasoning applies to reading scanned documents, as in the guide on why a computer sometimes misreads a document
How to choose, without being impressed by the word
Look at the incoming data before you look at the technology. If it already arrives with the fields in their place, like an e-invoice in XML, take the road of rules and do not look back, it is sturdier, cheaper to maintain and easier to explain to the people in the office. If it arrives as a photograph, or as a sentence written by a person, then interpretation really is needed
And if somebody offers you the word before they have even looked at where your documents come from, there is only one question to ask, which problem in my process does interpretation solve that a written rule would not solve better