The model is rarely the hard part. The hard part is the interface around it: what the user sees, trusts, and can undo.
Design for the wrong answer
AI gets things wrong. A feature that assumes a perfect answer breaks the first time it isn't. Build the recovery path first, then the happy path.