When we started drafting rehab referrals with AI, citations were on the list for the usual reason. A referral packet goes to another facility and shapes a patient's care, so every statement in it has to be true. If a model writes "tolerated three hours of therapy" and the PT note says 45 minutes, that's not a typo. It's a wrong decision waiting to happen.
So we required a source for every clinical sentence. Each line in a summary, packet or consult links to the note, flowsheet, order or result it came from. If the system can't point to the chart, the statement is removed before review.
We thought of that as a safety control. It turned out to be the feature that made the product work.
Review is the real workload
The promise of an AI draft is that it saves time. The catch is that someone still has to check it, and checking a draft you didn't write can take longer than writing it yourself. You're reading for errors in sentences that sound right.
A case manager preparing a rehab referral is reading days of notes for function, therapy tolerance, medical stability and home support. Before Aurifex, that took about 45 minutes per patient. An uncited draft doesn't remove that work. It moves it: you read the draft, then you go back to the chart to see whether it's true.
Citations change what review means. When "ambulated 40 feet with a rolling walker" links straight to the PT note from Tuesday, checking it takes one click. The case manager reads the claim, glances at the source and moves on. In workflow testing, referral prep went from 45 minutes to 10, and that time came from faster review, not from skipping it.
The draft didn't save the time. The citations did, because they made the draft cheap to trust.
Citations make errors visible
No model is perfect, and we don't claim ours is. In testing, 1.5% of reviewed draft statements weren't supported by their cited source. What matters is what happens next.
With citations, an unsupported statement is easy to spot. Either it has no source, which we block, or its source doesn't say what the draft says, which a reviewer sees in seconds. Each finding is corrected before anything is sent, and it goes into the clinician review queue with a category and a severity, so the next release fixes the pattern behind it.
Without citations, the same error hides in a fluent paragraph. Reviewers get tired of hunting, start skimming, and the safety net quietly fails.
Design choices that follow from this
Once we treated citations as the core of the product, a few decisions made themselves.
- Every clinical sentence, not just the important ones. Partial citation trains people to trust uncited lines, which are exactly the ones that need checking.
- Diagnoses come from the problem list. The agents don't infer a diagnosis from a narrative note. If it isn't on the problem list or in a documented assessment, it isn't in the packet.
- Critical fields are checked, not summarized. Allergies, code status and isolation are compared with the chart before a packet can be sent.
- The source travels with the packet. The facility receiving a referral sees the same numbered sources the case manager saw.
What I'd tell another team
If you're building AI for clinical work, measure how long review takes, not just how good the draft is. A draft that's almost always right but slow to verify will be abandoned. A draft that's slightly less polished but fast to verify will be used, and use is what turns into outcomes.
Citations were the cheapest way we found to make review fast. They're also what lets a case manager send a draft under their own name with confidence.