When you start building AI for hospitals, the list of possible problems is endless. Documentation, coding, scheduling, staffing, prior authorization, patient messages, readmissions. Every one of them is real, and every hospital leader you meet will tell you about a different one.
We picked one: getting patients to the right next place of care, sooner.
Why discharge and placement
Three things made this the right choice for us.
The cost of delay is concrete. A patient who is medically ready but waiting for a rehab bed, an authorization or home equipment is occupying a bed the hospital isn't paid extra for. Those excess days show up in length of stay, a number every hospital already tracks.
The work is document-heavy. A strong inpatient rehab referral needs evidence from days of notes: function, therapy tolerance, medical stability and home support. Reading and summarizing documents is where language models are genuinely useful, as long as every statement is cited.
The decision has a clear owner. Case managers own the referral. Attending physicians own the consult. PM&R and the receiving facility own the admission decision. That made it straightforward to design where AI drafts and where people decide.
Narrow scope makes results measurable
Because we chose one decision, we could choose one set of outcomes. Length of stay and case-mix-adjusted excess days, rehab acceptance and referral prep time. In testing, referral prep went from 45 minutes to 10.
A broad "AI assistant for clinicians" would have produced a long list of small improvements, each hard to measure and easy to dismiss. A narrow product produces a short list of results a CFO can check.
Narrow scope makes the product easy to explain
When a case manager asks what Aurifex does, the answer fits in a sentence: it predicts when each patient will be ready to leave and where they should go next, flags likely rehab candidates and drafts the referral for you to approve.
That clarity matters for trust. People are rightly cautious about AI in clinical work. A tool with one clear job, one clear limit and one clear owner for each decision is much easier to say yes to than a general-purpose assistant.
Saying no is the hard part
Focus means turning down good ideas. We've been asked to draft discharge summaries, to predict readmissions and to handle other referral types. Some of those may come later, where they share the same data and the same decisions. Each one has to earn its place by making the core decision better, not by making the product bigger.
What I'd tell another team
Pick the decision where delay costs the most, the evidence lives in documents and the owner is clear. Then build everything around making that one decision faster and better, and measure it honestly. Breadth can come later. Proof has to come first.