Early on, we had to decide where Aurifex would live. The easy path for a startup is a separate web application. You control the screens, ship changes daily and don't wait on anyone's EHR team.
We chose the harder path. Aurifex shows up inside Epic, in the case-management worklist, the chart and the order workflow that clinicians already use.
Every extra click is a reason not to look
A case manager's morning is full. They're covering a unit, preparing for the discharge huddle, returning calls from facilities and chasing authorizations. A new application, however good, is one more login, one more tab and one more place to remember to check.
Tools that live next to the workflow get opened when someone remembers. Tools that live inside it get used every day, because there's nothing to remember.
So the model's expected discharge date appears as a column beside the date the clinician set. A Possible IRF flag sits in the same row. Opening a flagged patient takes you to a candidacy review inside the chart, and the PM&R consult is a pended order waiting for the attending's signature, in the place attendings already sign orders.
Why it matters
In workflow testing, reviewing a cited draft in place took referral prep from 45 minutes to 10. I don't think any feature we could have built in a separate app would do as much for adoption as putting a flag in a worklist people already open.
A prediction nobody sees can't move a discharge date.
What it costs, and why it's worth it
Building inside the EHR is slower. Worklist columns and chart activities are configured with the hospital's Epic analysts. Changes go through their change control as well as ours. We test in non-production environments before anything reaches live data.
It also forces discipline. When your output appears in a clinician's own worklist, it has to be clearly labeled as a suggestion, it can never overwrite the clinician's own fields, and it has to be right often enough to deserve the space. Those constraints made the product better.
The exception: leadership
There's one place we didn't use the EHR. Executives and finance leaders see results on a separate reporting site, because they need aggregate numbers, not patient lists, and many of them don't work in Epic every day. The dashboard shows totals and rates, with patient-level detail left in Epic behind its existing access controls.
The rule isn't "always inside the EHR." It's "wherever that person already does this work."
What I'd tell another team
Before you design a screen, watch the person who'll use it for a full shift. Notice which screens they open without thinking. That's where your product should be. The best AI feature is the one people don't have to go looking for.