How Localized Voice AI Could Reshape Clinical Triage, Ensuring Data Sovereignty and Sub-Second Latency

Éric Pinet of Unicorne has presented a compelling model for operationalizing generative AI in highly regulated sectors like healthcare. His approach moves far past the glossy 'demo' phase that often defines enterprise AI adoption; instead, he focuses on solving core infrastructure and compliance challenges—specifically data sovereignty and real-time performance. The system itself is an intricate pipeline designed to intercept and structure initial patient calls for medical clinics across Québec. Instead of relying on receptionists taking anecdotal messages, the voice AI proactively engages the caller, asking structured questions based on the clinic’s specific triage protocols. The outcome is a comprehensive summary that significantly enhances the efficiency of nurses' subsequent callbacks. From an engineering perspective, what stands out is the technical rigor applied to two common failure points: latency and security. Firstly, Pinet correctly observed that in conversational AI, even fractional delays—anything over one second—can break user trust and cause patients to demand human intervention, undermining the system's goal. The solution requires a highly optimized multi-modal pipeline (Speech $\to$ Text $\to$ Generative Model Reasoning $\to$ Speech) with built-in conversational fillers ('OK, I understand') to maintain the illusion of fluid, human conversation. The second pillar is compliance and control. By running the entire process—from call handling (AWS Connect) to voice processing (Nova Sonic) to reasoning (AWS Bedrock)—entirely within a controlled AWS environment, Unicorne ensures that patient audio data never leaves the secure infrastructure. This architecture makes meeting stringent Québec privacy rules not merely an add-on compliance step, but a foundational element of the system itself. In short, the security model dictates the product design. Unicorne’s philosophy—that infrastructure questions must precede model questions—is a critical corrective to the prevailing pattern in enterprise AI. For regulated Canadian industries, data residency and auditable logging are not secondary concerns; they *are* the product guarantee. The system's ability to seamlessly hand off calls when distress is detected or protocols are exceeded ensures that human expertise remains appropriately prioritized, building trust rather than replacing it.