Why One Dev Team Chose Hard-Coded Java Rules Over AI for Refund Decisions

A software development team building an LLM-powered support agent deliberately chose deterministic Java rules over AI judgment for processing refund eligibility decisions. The team identified four core risks with letting a language model decide refunds: inconsistent outputs, vulnerability to prompt-injection via user-submitted data, unnecessary per-token costs for simple logic, and inability to provide auditable explanations. Their chosen approach encodes published refund policy — covering delivery status, payment status, and return windows — as plain Java code tested with JUnit and carrying no AI dependencies. Edge cases that fall outside the rules are escalated as policy work items requiring human review, ensuring decisions remain traceable and reviewable. The team documented the choice as an Architecture Decision Record, noting the decision can be revisited if data ever shows a machine-learning approach measurably outperforms the rule-based system.