Use rules when the policy is known
I use rules when the product needs a clear, stable answer and the cost of a wrong action is high. Rules are often the better product choice when the policy is known, the inputs are structured, and a reviewer needs to understand why the system allowed or blocked something.
In course recommendations, a model can help rank relevant options, but it should not override prerequisites or recommend a course a student has already completed. At Oracle, I worked on an AI-powered course recommendation and personalization system where the experience combined ranking with rules around eligibility. Available student and course attributes helped address cold start when behavioral history was thin. Separately, rule-based recommendations offered a fallback if the model path was unavailable.
Put rules at decision boundaries
A launch should not proceed because a dashboard score looks healthy if a required owner, check, or evidence record is still missing. The Launch Control sample on this site is illustrative, but the principle comes from product and program delivery: define the exit criteria, make ownership visible, prepare monitoring and rollback, and require the right sign-offs before a release.
Rules are not a concession that makes a product less intelligent. They make the system legible. They protect a user from an answer that sounds plausible but ignores a constraint the business already understands.
Use AI where interpretation is needed
I would reach for AI when the problem requires interpretation across messy language, a large evidence set, or patterns that a fixed rule cannot reasonably capture. Even then, I keep the boundary explicit. The system can summarize, retrieve, rank, or propose; a known policy can still decide whether it is permitted to act.
The useful question is what kind of judgment is needed, what evidence supports it, and what should happen when the system is uncertain. If the answer is a stable policy, I start with a rule. If the answer requires interpretation, I use AI with review, measurement, and a recovery path.