The network had accumulated, over years, a serious body of clinical knowledge — internal protocols, decision aids, guidelines, peer-reviewed references. All of it lived in folders and binders. When a nurse or PA needed an answer mid-shift, the options were: interrupt a physician, dig through the drive, or go from memory. All three happened daily. The knowledge existed; it just wasn't reachable at the speed care moves.
This is the setting where 'just use ChatGPT' is not an acceptable answer, and everyone in the room knew it. A model that paraphrases confidently and cites nothing has no place near a clinical decision. Whatever we built had to answer only from the network's own material, show its source on every answer, and refuse to answer when the material didn't cover the question. The refusing part mattered as much as the answering part.
What shipped was a retrieval system over the network's own protocols. A clinician asks a plain-English question and gets back the relevant passage from the network's actual documents — quoted, cited, linked to the source page. Not a summary in the AI's voice. The network's own words, found fast. If the corpus doesn't cover it, the system says so and routes to the right human instead of improvising.
The work that made it trustworthy wasn't the model — it was the curation we did with the clinical leads. Which documents are current. Which are superseded. Which are site-specific and which apply network-wide. The system inherited the network's own discipline about what counts as an answer. That's why the clinicians use it: it behaves like their best colleague, not like a chatbot.
Answers that used to take minutes — sometimes hours, when the right person was off shift — now come back in seconds. Usage grew week over week without a mandate, which is the only adoption signal that means anything. The clinical leads now maintain the canonical set themselves. Next lane: staff onboarding and training, on the same foundation.