An insurer's eligibility rules are written as explicit thresholds in an underwriting manual, but applications arrive as free-form text with attachments. The team proposes putting the whole manual in the prompt and letting Claude decide eligibility. Compliance requires every decision to be explainable and reproducible. Which design is most appropriate?

Answer: A — Have Claude extract applicant facts with source citations, apply the codified rules in code, and send ambiguous cases to an underwriter

(a) puts language work (extracting facts from messy applications) with the model and deterministic threshold logic in code, which is exact, reproducible and auditable, with humans on genuine ambiguity. (b) produces explanations but the decision remains probabilistic and not reproducible. (c) and (d) add capability or redundancy to a task that should not be the model's to decide.