Treat vendor AI agents as a data source you don't control
At HIMSS26 this week, Oracle announced a clinical AI agent that drafts notes in emergency and inpatient settings. Other vendors are moving the same way. AI agents built into the clinical workflow are arriving quickly.
Many of these tools will help. Documentation takes a lot of clinician time, and a good drafting tool gives some of it back.
But each agent built into a vendor’s product also creates a new kind of data inside your systems. Most data platforms aren’t set up for it yet.
A closed loop
The vendor approach is easy to follow. Put an AI agent inside the EHR workflow for notes, documentation or decision support. Now the EHR is more than the system of record. It’s also where the model runs. The model reads clinical data, writes structured and unstructured output, and that output goes back into the same system.
That’s a closed loop, and closed loops tend to lock you in.
Ask a few questions about any note an agent wrote. Where is the record of how it was produced? Where does the model’s confidence score live? Who can trace what the model was trained on? If your analytics and AI work runs on Snowflake, dbt and the rest of a modern data stack, you’ve just gained a source you don’t control and can’t easily see into.
Your governance assumes a person wrote it
Most health systems have spent years building data governance around EHR extracts, claims and operational systems. Those frameworks assume a person created the data and that you know where it came from.
AI-written clinical content breaks that assumption. It sits in the same tables and the same FHIR bundles as content people wrote, and it flows through the same pipelines. Your dbt models, quality checks and compliance reports can’t tell the two apart.
That matters. I expect auditors and payers to look harder at AI-generated documentation. If your pipeline can’t tag each row as written by a person or by a model, you’re building compliance risk into your analytics.
It also creates three practical problems for data teams.
Lineage. Every AI-written item needs a record of which model and version produced it, from what inputs, with what confidence. Your data catalog has to track that, and your dbt models have to carry it forward. Snowflake’s ACCESS_HISTORY and OBJECT_DEPENDENCIES views help with governance, but they won’t capture this on their own. You’ll need a separate provenance trail for AI-written content.
Visibility. Vendor agents run inside the EHR. You see the output, such as notes, orders and alerts, but not the telemetry. From your data platform, you can’t watch for model drift, falling quality or invented content. The AI with the most influence on patient records may be the one you can see least.
Schema change. Agents add to your data model, with new fields, new document types and coding patterns no person would produce. Expect your EHR extract pipelines to break more often. Your ingestion jobs, whether you use dlt or something else, need to handle schema drift without someone patching staging models every sprint.
How I’d set up the platform
I wouldn’t reject vendor agents. I’d plan for several of them and treat each one as an untrusted source that gets full monitoring.
- Ingest agent output as its own source. Give it its own staging, validation and lineage instead of letting it blend into the EHR extract. In dbt, that means its own source definitions, freshness checks and schema tests.
- Tag AI content when it lands. Every record an agent touched gets the model ID, version, timestamp and confidence score. When an auditor asks about AI-generated documentation, the answer is a query away.
- Watch the output yourself. Use the data quality framework you already have. Write dbt tests on the statistical shape of AI-written data, such as average note length, ICD code distribution and how many clinical terms appear per note, and alert when those numbers shift.
- Keep your own inference layer. Run your own models on Snowflake Cortex, Amazon Bedrock or the ML platform you prefer. Let the vendor agent handle drafting at the point of care. Let your platform handle analytics, population health and operations. Don’t fold both into one vendor’s runtime.
EHR vendors want to own the clinical AI layer because it’s the most valuable part of healthcare IT. If they own it outright, your data platform ends up downstream of AI output you can’t inspect or reproduce.
The health systems that do well will treat vendor agents like any other outside data source. Useful, and checked before trusted. The system of intelligence should be yours, not your EHR vendor’s.
Here’s a concrete first step. Pick the first vendor agent you plan to turn on. Before go-live, ask the vendor what metadata comes with each output in your extracts: model, version, timestamp, confidence and a flag that marks it as AI-written. If they can’t supply it, make it a contract term before you sign.
Discussion