Reducing hallucinations and safety design

Doctronic became the first AI in the world legally licensed to practice medicine through Utah's AI Learning Lab regulatory sandbox in December 2025. In this episode of BUILDERS, I sat down with Matt Pavelle, Co-founder and Co-CEO of Doctronic, to learn how he and his co-founder (a physician) launched an AI-powered primary care chatbot in September 2023, validated demand through Facebook chronic condition groups and minimal Google Ads spend, and navigated uncharted regulatory territory to offer $4 prescription renewals for chronic conditions—targeting the medication non-adherence problem that causes 125,000 preventable deaths and costs $100B annually.

Why friends with excellent health insurance still couldn't get medical answers quickly Building clinical accuracy into GPT-3.5 when context windows were small and hallucinations were rampant The tactical launch: Google Ads plus Facebook chronic condition groups in September 2023 Architecting safety: RAG with tens of thousands of physician-written clinical guidelines The study: 99.2% agreement rate between AI treatment plans and human doctor reviews across 500 patients Navigating Utah's AI Learning Lab: the only regulatory sandbox that mitigated medical licensing laws Securing AI malpractice insurance through Lloyd's Market—a first in the industry The three-phase oversight model: 100% human review, then 10%, then spot checks Expansion strategy: targeting other state regulatory sandboxes and international governments

Launch with the minimum feature set that proves your core hypothesis: Pavelle shipped Doctronic in September 2023 without user accounts—chats disappeared when closed unless users saved them manually. Within days, user requests for persistent chat history validated demand. The insight: your MVP should test one assumption, not solve every user need. If you're hesitating to launch because features are missing, ask whether those features are actually required to validate your hypothesis or just things you assume users want.

Use specificity to unlock early adoption in skeptical markets: Rather than targeting "healthcare" broadly, Pavelle posted in Facebook groups for specific chronic conditions, offering a free AI backed by clinical guidelines. Half the groups banned them for commercial activity, but the other half engaged immediately. The lesson: in regulated or skeptical markets, narrow targeting with explicit safety mechanisms (clinical guidelines, physician co-founder credibility) converts better than broad positioning. Identify where your skeptics congregate and address their specific objections upfront.

Design system architecture to prevent failure modes, not just tune models: Doctronic's safety architecture separates AI decision-making from prescription execution. The LLM asks questions and determines renewal safety, but deterministic code outside the AI verifies the prescription exists, checks dosage accuracy, and confirms the schedule. Even if adversarial prompting compromises the LLM, the deterministic layer prevents bad outcomes. Founders building high-stakes AI products should architect multiple independent verification layers rather than relying on prompt engineering or temperature tuning alone.

Target regulatory pain points with quantified deaths and costs: Pavelle approached Utah with specific numbers: 125,000 preventable deaths annually from medication non-adherence, 30-40% caused by renewal friction, and a $100B economic burden. These statistics—combined with Utah's rural population and physician shortage—made the problem impossible to ignore. When approaching regulators, lead with mortality and cost data that make inaction untenable, not just efficiency gains or convenience improvements.

Regulatory sandboxes require proof of safety methodology, not just technology demos: Utah's AI Learning Lab didn't just grant Doctronic permission—they required a three-phase oversight structure where human physicians review 100% of initial prescriptions in each medication class, then 10%, then ongoing spot checks. Pavelle also secured AI malpractice insurance through Lloyd's Market before launch. The insight: regulatory innovation offices want risk mitigation frameworks, not promises. Build and fund your oversight methodology before approaching regulators, and treat insurance underwriting as a third-party validation of your safety claims.

Publish clinical validation studies before scaling—they become your regulatory and sales asset: The study showing 99.2% agreement between Doctronic's AI and human physicians across 500 patient encounters became the foundation for regulatory conversations and public trust. Founders in regulated spaces should budget for formal validation studies early—these aren't marketing expenses, they're the permission structure for everything that follows. Work backward from what regulators and enterprise buyers need to see, then design studies that generate that specific evidence.

// Sponsors: Front Lines — We help B2B tech companies launch, manage, and grow podcasts that drive demand, awareness, and thought leadership. www.FrontLines.io

The Global Talent Co. — We help tech startups find, vet, hire, pay, and retain amazing marketing talent that costs 50-70% less than the US & Europe. www.GlobalTalent.co

//

Don't Miss: New Podcast Series — How I Hire Senior GTM leaders share the tactical hiring frameworks they use to build winning revenue teams. Hosted by Andy Mowat, who scaled 4 unicorns from $10M to $100M+ ARR and launched Whispered to help executives find their next role. Subscribe here: https://open.spotify.com/show/53yCHlPfLSMFimtv0riPyM