Most Retrieval Augmented Generation (RAG) systems for high-stakes classification mistakenly route every ambiguous case directly to a large language model (LLM), which fails under audit and scrutiny. The author advocates for a different design philosophy, emphasizing auditability, cost, and consistency for regulated enterprise settings where wrong answers have significant consequences. An all-LLM pipeline incurs invisible costs: difficulty in auditing decisions, high inference costs and latency at scale, and inconsistent performance on cases that should be deterministic.A cascade architecture addresses these issues by treating the LLM as an escalation path, not the front line. Stage one is deterministic, resolving clear-cut cases with rules and exact matches, clearing the majority of volume without LLM calls, ensuring full explainability. Stage two employs a retrieval layer for cases unresolved by stage one, pulling specific evidence like prior decisions or contextual documents, with retrieval quality being paramount. Stage three is the LLM call, reserved only for the genuinely ambiguous cases that stages one and two could not resolve. This approach significantly cuts inference costs and improves consistency on deterministic cases.For the LLM stage, an asymmetric risk prompt is crucial, acknowledging that the cost of different error types is not equal. This means instructing the model to escalate uncertainty, providing calibrated examples with consequences, and requesting a confidence score alongside the classification. The confidence score acts as a second cascade point, routing low-confidence cases to human reviewers.Evaluating such a system requires specific adjustments: retrieval quality must be measured independently from final classification accuracy, and the evaluation set needs to oversample stage-three cases. LLM as judge evaluation is effective if the judge prompt incorporates the same asymmetric risk framing. Finally, a feedback loop from confirmed outcomes back into the retrieval corpus is essential for continuous improvement in handling ambiguous cases. The broader lesson is that in high-stakes domains, the valuable engineering work lies in deciding what parts of a decision should never involve the model.