In the 1980s this was what "AI" meant commercially: an expert system — a human specialist's knowledge written out as explicit IF/THEN rules, plus a small engine that chains through them. There is no model here and nothing was learned; a person sat down and typed the rules. What it gives you in exchange is total transparency. When it reaches a conclusion it can show you exactly which rules fired, which were ruled out, and which single answer killed them — something no neural network on the layers below can do.
The knowledge base is a list of rules, each one a set of conditions and the diagnosis plus treatment they imply — the whole "intelligence" of the system is that list, written by hand. The inference engine walks the rules in order and evaluates conditions lazily: it only asks you about a symptom at the moment a rule actually needs it. Two optimisations do all the work. First, every answer is cached, so a symptom is never asked twice even though six different rules may depend on it. Second, the engine short-circuits — the instant one condition comes back false the rule is abandoned and its remaining questions are never put to you, which is why a twelve-rule consultation usually takes five or six questions. Because the reasoning is just a trace through explicit rules, the system can always explain itself, and its ceiling is equally explicit: it knows precisely what someone typed into it and not one thing more.