Beyond the Data Warehouse: Rethinking AI Enablement in Healthcare

For fifteen years, healthcare’s answer to “how do we use our data” has been the same: build a bigger, more centralized warehouse. There was a real reason for it – an EHR only sees its own records, and crucial context, including outcomes, lives elsewhere. Assemble enough of it in one place, the thinking went, and you’d eventually have a dataset worth training a model clinicians could trust. That bet no longer makes sense. LLM capability has outpaced the value of massive retrospective data repositories — and the EHR data those repositories depend on remains largely locked away by vendors with little incentive to open it, FHIR and HL7 mandates notwithstanding. Meanwhile, the cost and context loss of feeding population-scale datasets to an LLM makes the monolithic approach impractical on its own terms.

This talk argues for a different bet, one healthcare has seen before: the same resistance-then-rapid-adoption pattern that played out with BYOD in the 2000s is playing out again, this time with data architecture. Instead of retrospective, warehouse-first analytics, the better question is how to safely AI-enable specific roles across care delivery, transactionally, using transformation pipelines rather than monoliths. Much of that opportunity doesn’t depend on EHR data at all. One example: leveraging a model to route incoming messages, propose responses, and automate follow-up.

Using AWS as the toolset — data lake infrastructure, Glue, Lake Formation, governance and cataloging, and Bedrock for compliance-conscious model integration — we’ll walk through an architecture that treats the LLM as swappable rather than fixed, and show what a secure, auditable, role-specific AI enablement pipeline actually looks like in practice.