Documentation, cited.
A multi-site provider deployed a citation-grade clinical documentation assistant inside its own cloud — every answer grounded in a source, every retrieval scoped to the clinician's access rights.
Answers clinicians could not verify.
Clinicians were spending a meaningful share of each encounter searching prior notes, lab results, and care-plan documentation scattered across systems. A first attempt at an AI search assistant produced fluent answers — but with no way to check them against the record, and no guarantee it respected which patients a given clinician was authorized to see.
For a healthcare system, an ungrounded answer is not a minor defect. It is a clinical risk and a compliance exposure at once.
Retrieval that inherits access, not just relevance.
We treated permissioning as a first-class part of retrieval, not a filter bolted on after. Every query carries the clinician's access scope; the index can only surface documents that scope already permits. On top of that, every answer is required to cite the specific note or result it draws from — an answer with no supporting citation is suppressed rather than shown.
This meant building the evaluation harness around groundedness from day one: for every test question, we measured not just whether the answer was correct, but whether it was traceable to a real, permitted source.
Deployed inside the provider's own cloud.
The retrieval index, the model calls, and the citation store all run inside the provider's existing cloud environment — nothing clinical leaves the perimeter they already govern. An access-control layer sits in front of retrieval and re-checks permissions on every query, not just at login.
Trusted enough to use daily.
Clinicians adopted the assistant for daily documentation lookup once the citations made answers checkable in seconds rather than minutes. Compliance signed off once the access-control re-checks were demonstrated against real audit scenarios, not just described in a document.