For compliance, a rule engine beats an AI model
A compliance check has to give the same verdict every time and be able to explain why. That rules out a model as the decision-maker, and rules in a deterministic rule engine.
Why determinism matters here
If you are validating a regulated document, the result cannot wander. The same input must produce the same output today and next month, and you have to be able to point at the exact rule that failed. A language model is the wrong tool for that job: its answer can change from one run to the next, and it cannot give you a stable, auditable yes or no.
What a rule engine gives you
So the core is a deterministic rule engine: a large set of explicit checks, grouped by category, each one repeatable and auditable. Run it twice on the same document and you get the same answer, with a clear list of what passed and what did not. That is what a reviewer, and an auditor, actually need.
Where AI still helps
AI is not useless here, it is just not the judge. It is good for the free-form parts: answering a question about a document, or giving a second opinion that a human can weigh. It assists the reviewer; it never decides the verdict. The rules decide.
Where I used it
This is the spine of an IVDR compliance platform that validates in-vitro diagnostics documents against regulatory and ISO standards: a deterministic engine for the verdict, with AI kept to the optional second opinion.
Questions
Why not let an AI model do the checking?
Because the same document could pass one day and fail the next. Compliance needs a result that is the same every time and that you can explain rule by rule, which a deterministic engine gives you and a model does not.
So where does AI fit?
As a helper, not a judge. It is good for free-form questions and a second opinion, but the pass or fail decision stays with the rules.