Engineering Lessons from Building a Compliance-Bounded Multi-Tenant Voice Agent in Financial Services

Large-language-model agents are now widely evaluated as task solvers in fixed environments, but their deployment in regulated industries surfaces a largely orthogonal set of engineering concerns that benchmark-oriented work seldom addresses. We report eight engineering lessons distilled from building and operating a multi-tenant voice agent in the financial-service industry — a deployment that is simultaneously multi-tenant, voice-bound under a hard end-to-end latency budget, and subject to regulatory constraints whose violation is not recoverable once an action becomes externally observable. The lessons span multi-tenant prompt isolation, PII-bound retrieval under encryption, voice–LLM coordination within a perceptual deadline, plan–executor decoupling for regulatory action filtering, supervisor handoff as an explicit state machine, per-tenant evaluation customization, knowledge-gap detection with sanctioned fallback, and continual-learning data quarantine. Each is presented with its issue, design choice, rationale, and internal validation. We organize the lessons along three recurring axes — tenant isolation, long-horizon state coherence, and regulatory constraint adherence — and state an applicability boundary tying their transferability to how many of the three deployment properties a target setting shares. The lessons are design observations rather than measured results: a catalog of failure modes that are invisible in single-tenant, text-only, or unregulated prototypes but become first-order at their intersection.