Oracle just dropped its Clinical AI Agent for note generation in ED and inpatient settings at HIMSS26. Kala Bio is shipping a commercial AI agent product in 14 days. The $180 billion agentic AI healthcare market is no longer a forecast — it's a sprint.
But here's what nobody in the HIMSS hallways is talking about: every vendor-embedded AI agent is a new data silo waiting to happen.
The Vendor Play
Oracle's move is textbook. Embed an AI agent deep into the clinical workflow — note generation, documentation, clinical decision support — and suddenly the EHR isn't just a system of record. It's the inference layer. The model consumes clinical data, generates structured and unstructured outputs, and those outputs flow back into the same system.
This is a closed loop. And closed loops create lock-in.
When Oracle's AI agent generates a clinical note, what happens to the provenance metadata? Where does the model's confidence score live? Who audits the training data lineage? If you're running your analytics and AI workloads on Snowflake, dbt, and a modern data stack, you just acquired a new data source that you don't control and can't easily observe.
The Governance Gap
Most health systems have spent the last three to five years building data governance frameworks around their EHR extracts, claims data, and operational systems. Those frameworks assume human-generated data with well-understood provenance.
AI-generated clinical data breaks that assumption.
When an AI agent authors a note, it introduces a new category of data: machine-authored clinical content that lives alongside human-authored content in the same tables, the same FHIR bundles, the same downstream pipelines. Your dbt models, your quality checks, your compliance reporting — none of it was designed to distinguish between the two.
This isn't theoretical. CMS is already signaling that AI-generated documentation will face different scrutiny in audits. If your data pipeline can't tag AI-generated versus human-generated clinical content at the row level, you're building compliance risk directly into your analytics layer.
Three Data Engineering Problems You Didn't Have Last Month
The clinical AI agent wave creates concrete challenges for data teams:
New lineage requirements. Every AI-generated artifact needs provenance metadata: which model, which version, what inputs, what confidence score. Your data catalog needs to track this. Your dbt models need to propagate it. If you're using Snowflake's ACCESS_HISTORY and OBJECT_DEPENDENCIES views for governance, you now need a parallel provenance chain for AI-authored content that those views will never capture natively.
Observability blind spots. Vendor-embedded AI agents operate inside the EHR's walled garden. You get the outputs — notes, orders, alerts — but not the telemetry. You can't monitor model drift, data quality degradation, or hallucination rates from your data platform. You're flying blind on the most consequential AI in your organization.
Schema evolution at AI speed. AI agents don't just consume your data model — they extend it. New fields, new document types, new coding patterns that no human would produce. Your EHR extract pipelines will break more often. Your dlt ingestion jobs need to handle schema drift gracefully, or you'll be patching staging models every sprint.
The Architecture Response
The answer isn't to reject vendor AI agents. Oracle's clinical note generation will genuinely reduce physician documentation burden. The answer is to architect for AI agent pluralism — a data platform that treats every AI agent as an untrusted data source with first-class observability.
Ingest AI agent outputs as dedicated data sources with their own staging, validation, and lineage tracking. Don't let them hide in the EHR extract as undifferentiated rows. Give them their own dbt source definitions, their own freshness checks, their own schema tests.
Implement AI content tagging at the ingestion layer. Every record that touches an AI agent gets flagged with model metadata — model ID, version, timestamp, confidence. This is non-negotiable for compliance and will save your team hundreds of hours when CMS auditors start asking questions about AI-generated documentation.
Run your own observability on agent outputs. Monitor for drift, anomalies, and quality degradation in AI-generated content using your existing data quality framework. Build dbt tests that assert statistical properties of AI-authored data — average note length, ICD code distribution, NLP entity density — and alert when distributions shift.
Keep your inference layer independent. Run your own models on Snowflake Cortex, Bedrock, or your preferred ML platform. The vendor AI agent handles point-of-care generation. Your platform handles analytics, population health, and operational intelligence. Don't collapse these into a single vendor runtime.
The Platform War Is Here
What's unfolding at HIMSS26 is the opening salvo of a platform war. EHR vendors want to own the clinical AI layer because it's the highest-value surface area in healthcare IT. If they succeed, your data platform becomes a downstream consumer of AI outputs you can't inspect, audit, or reproduce.
The health systems that win the next five years will treat vendor AI agents like any other external data source: valuable, but not trusted by default. They'll build the governance, observability, and architectural independence to leverage these tools without being captured by them.
Your Snowflake instance should be the system of intelligence. Not Oracle's. Not Epic's. Yours.
The land grab is on. Build the architecture that gives you leverage — or get comfortable being a passenger.