What do you automate when a patient’s appeal is on the line?

We chose prior authorization appeals rather than care navigation for newly diagnosed patients, billing disputes, or a general health platform. Care navigation had the highest emotional intensity, but its win condition was too diffuse to assess in weeks; a prior authorization case is approved or it is not. The narrow entry point would make a small business if adjacent workflows never opened up, but its binary outcome also made a success fee possible without a track record and let us tell whether the product worked before there was enough data to model anything.

We limited Phase 0 to AI-assisted intake and classification, manual advocate work, and a case timeline, rather than a full case copilot that drafted appeals, summarized history, extracted facts, and suggested actions, or drafting with advocate review. All of it was reachable by week two, but every week AI stayed out of drafting left advocates writing appeals by hand on a success fee that only pays when cases close. We accepted that cost because a slow advocate loses margin, while an advocate who stops trusting the classifier checks every call by hand and costs more than the automation saved. An advocate who catches the AI wrong twice stops trusting the system; one who saves two hours while staying in control tells other advocates. The model and accuracy are the same, but what it touches first changes the outcome.

We chose a ranked queue with the required action on each card over one case per screen with full depth or a status list. The depth-first version required backing out, scanning, and clicking into the next urgent case, creating eight context switches across eight active cases before the day began; the queue puts a prior authorization deadline expiring today above a document review that can wait until Thursday, while care plans, activities, and documentation stay inside the case. Too little on a card would make advocates open everything and too much would make eight stacked cards unreadable, so each has patient name, case type, next required action, and deadline. Testing showed advocates also wanted their workload on screen, not only patient status: a session summary changed “I can see what my patients need” to “I can see what my day looks like,” a difference tied to burnout.

We led profiles with aggregate outcome patterns against named insurers rather than credentials or named patient testimonials. The first profile ordered credentials, specialties, and experience like a resume, but patients managing chronic pain wanted someone who had beaten their insurer on their condition, and patient narratives would expose people. Profiles use patterns such as “15 of 17 denials at this insurer for this condition reversed on first appeal,” with ratings separated into communication, resolution, and speed. Publishing an acceptance rate gives patients a selection signal but creates liability if criteria are not legible, so the criteria sit alongside it. In testing, patients identified the rate as what separated the product from a directory: “So they’re not just listing everyone.”