Building an On-Device Prior-Authorization Copilot at Build with Gemma NYC
Google’s Gemma team, together with Bond AI, held Build with Gemma NYC, a one-day hackathon at One World Trade Center built around one of healthcare’s hardest constraints: keeping patient data off the cloud.
More than 200 builders came to work with Google’s on-device Gemma 4 models. My team built DermAuth, a fully on-device prior-authorization copilot for dermatologic biologics. The problem it targets deserves more attention than a single hackathon can give it.
As a board-certified dermatologist and fellowship-trained Mohs surgeon who has published on AI in dermatology since 2012, I’ve spent much of my career studying how artificial intelligence can support clinical care without replacing clinical judgment. Prior authorization sits squarely between diagnosis and treatment, and with every delay, there is also a delay in patient care.
Why Prior Authorization Is So Difficult
Prior authorization is one of the most consistent sources of friction between a physician’s treatment plan and a patient actually receiving that treatment. According to the American Medical Association’s 2025 prior authorization survey, 95% of physicians say prior authorization delays access to necessary care, and 79% report that patients abandon treatment altogether because of it. Physicians complete an average of 40 prior authorizations a week, consuming roughly 13 hours of physician and staff time each week. More than one in four physicians (26%) report that prior authorization has led to a serious adverse event for a patient.
Source: American Medical Association, 2025 Prior Authorization Physician Survey
Dermatology carries its own version of this problem. A 2025 review published in the Journal of the American Academy of Dermatology examined 100 consecutive prior-authorization denial letters for biologics and JAK inhibitors at one health system and found that 75% did not identify who made the decision, and 41% gave no information about step-therapy requirements or covered alternatives.
For patients on biologics, often the only effective option for moderate-to-severe psoriasis, hidradenitis suppurativa, or other severe skin disease. These delays mean more time in active disease, more visits spent on paperwork instead of treatment, and, for a majority of patients, treatment abandoned before it ever starts.
Why On-Device AI Matters Here
Healthcare carries some of the strictest privacy constraints in software, and Gemma 4 was built with exactly that constraint in mind: it can run locally, meaning patient data never has to leave the device. That capability was the organizing idea behind the hackathon, which split builders across three tracks.
Three tracks, one constraint:
- Voice for Care: scribes, intake, triage, reminders, and accessible, multilingual patient communication.
- Agentic Care Copilots: agents that automate healthcare admin like prior authorization, referrals, and scheduling.
- On-Device Private Health Tools: fully offline tools where patient data never touches the cloud.
Introducing DermAuth
Drawing on years of firsthand experience with biologics prior authorization in my own practice, my team built DermAuth to sit inside that third track. It uses on-device voice intake to create a structured summary, checks that summary against payer step-therapy criteria, and drafts a citation-linked prior-authorization letter for physician review. If information is missing or uncertain, DermAuth flags the gap and generates follow-up questions in the patient’s own language, using Gemma 4’s multilingual support.
DermAuth is designed as a decision-support tool, not a replacement for physician judgment. It drafts a letter for physician review. It does not submit decisions on its own.
What This Could Mean for Patients
If tools like DermAuth make it out of the hackathon and into a clinic, the impact would show up in exactly the places the AMA and JAAD data point to. A structured, citation-linked prior-authorization letter addresses the accountability gap those denial-letter reviews identified. A request that is complete and well-documented the first time is less likely to come back as a denial that offers no real explanation. Flagging missing information before submission, rather than after a denial, could shorten a process that currently stretches a days-long decision into weeks of back-and-forth.
Because it runs on-device, none of that requires sending a patient’s full chart to a third-party server just to check a step-therapy requirement. And because Gemma 4 supports multiple languages, follow-up questions can go back to a patient in their own language instead of defaulting to English-only intake. That detail matters for the patients most likely to fall through the cracks of an already opaque process.
None of this replaces physician judgment. But if even a fraction of the 13 hours a week practices spend on prior authorization can move from paperwork back to patients, and if fewer of the treatments abandoned each year are lost to missing documentation rather than genuine ineligibility, on-device AI has a real role to play in closing that gap.
The Bigger Picture
Prior authorization is one of the more visible pain points in dermatology, but it’s really a proxy for a bigger question: how much administrative work in healthcare can move onto the device a patient is already talking to, without that data ever leaving it? Build with Gemma NYC offered one clear answer over the course of a single day. DermAuth aims to keep testing it, with the goal of turning a hackathon prototype into something that actually shortens the distance between a diagnosis and a patient’s first dose.
Watch a video of the DermAuth demo:
