Healthcare organizations are under pressure to adopt generative AI while managing uncertainty around cost, reliability, privacy, vendor lock‑in, and accountability in clinical and medico‑legal workflows. My talk presents a real‑world case study from the Canadian Medical Protective Association, where we designed and deployed an internally hosted architecture of orchestrated, task‑specific small language models (SLMs) for medical‑legal advice workflows. The system automates Note‑to‑File summarization and multi‑label classification of medico‑legal issues across thousands of physician advice cases, achieving high-quality performance comparable to, and in some areas exceeding, human coders, with quasi‑deterministic behavior suitable for governance‑critical use. I will share our end‑to‑end approach, including de‑identification, fine‑tuning, validation against human agreement, human‑in‑the‑loop review, and operational monitoring. The session is designed to give healthcare leaders and technical teams a concrete blueprint for building regulatory‑grade, on‑premises AI that balances innovation, risk management, and strategic autonomy.