Medical AI's Real Gatekeeper Isn't Model Accuracy—It's the Hospital IT Security Review
Most Western medical AI tools are cloud-dependent, which creates a legal deadlock in jurisdictions with strict data residency laws. A local-first architecture that still supports model swapping and persistent team knowledge offers a deployment path that doesn't require rewriting privacy regulations.
China's 2026 healthcare reforms are pushing AI into primary care, but the hard constraint remains unchanged: patient data cannot leave the hospital intranet. Coco addresses this by running entirely locally, storing conversation context, case summaries, and quality-control rules on-disk without touching external servers. Its model-agnostic architecture lets hospitals swap the underlying LLM—on-premises or cloud—without losing the team's accumulated knowledge.
Three concrete workflows show the architecture at work. A township doctor gets diagnostic decision support, drug interaction checks, and auto-generated referral summaries without an internet connection. A medical records department shifts from spot-checking 3,000 monthly records to full review by codifying quality-control manuals as Skills that flag anomalies for human review. Multidisciplinary tumor board prep shrinks from two hours of cross-system data gathering to a ten-minute verification of an auto-generated timeline.
The underlying argument is that medical AI's bottleneck is not capability but trust, which breaks down into three requirements: data never leaves the building, recommendations cite current guidelines rather than stale training data, and the system survives network outages and vendor failures. Coco's local-first design hits all three without being healthcare-specific.
The piece reframes medical AI's core problem as a compliance bottleneck rather than a technical one: the question is not whether the model is good enough, but whether the IT department can sign off on the security review.
By making the architecture model-agnostic and local-first, Coco decouples institutional knowledge from any single AI vendor—a hedge against model deprecation, vendor bankruptcy, or regulatory shifts that Western hospital systems rarely build into their procurement.
The three scenarios illustrate a pattern where AI doesn't replace clinical judgment but absorbs the fragmented, cross-system data-gathering work that currently burns hours of skilled labor, which is a more realistic near-term deployment model than autonomous diagnosis.