Why Direct Primary Care is the Right Beachhead for Healthcare AI

Direct Primary Care is experiencing remarkable growth — up 600% since 2017 — despite economic headwinds pushing most practices toward corporate consolidation. Over 2,000 DPC practices now serve 400,000+ Americans with a model that seems counterintuitive: fixed monthly fees ($75-150), no insurance billing, and physicians managing just 600-800 patients instead of 2,500+.

This isn't a retreat from modern medicine. It's a laboratory for what healthcare could become when incentives align around patient outcomes rather than billing optimization. And it's precisely where clinical AI needs to prove itself first.

Here's why: the data DPC physicians generate is fundamentally different from the massive datasets powering current healthcare AI — and that difference matters more than size.

The Data Quality Advantage: Clinical Truth vs. Billing Artifacts

Epic's COMET model — trained on 300 million patient records—represents the current paradigm: scale through aggregation of EMR data. But there's a hidden problem with this data. It was generated in a system where documentation serves billing requirements first and clinical reasoning second.

When a physician in a fee-for-service system documents a visit, they're answering: "What do I need to record to justify this billing code?" The clinical narrative becomes optimized for reimbursement, not for the next physician who needs to understand this patient's longitudinal story.

DPC physicians document differently because they answer different questions: "What do I need to remember about this patient in six months?" Their notes capture clinical reasoning — the why behind decisions — because there's no billing code to justify. They track what actually matters for continuity: patient preferences, previous therapeutic trials, contextual factors affecting adherence.

This produces training data with fundamentally different properties. For AI to genuinely augment clinical reasoning rather than automate billing, it needs to learn from physicians thinking about medicine, not physicians thinking about reimbursement.

The Administrative Burden DPC Physicians Still Carry

DPC doctors didn't eliminate administrative work — they just changed its nature. Instead of fighting insurance denials, they're managing:

Medication logistics: Tracking GoodRx prices, coordinating with discount pharmacies, managing prior authorizations for the few things insurance still covers for their patients.

Lab and imaging coordination: Negotiating cash-pay rates, tracking results across multiple systems, following up on abnormal findings without integrated EMR infrastructure.

Specialist referrals: Finding specialists who accept cash-pay patients, translating clinical context without standardized referral systems, ensuring continuity when specialists use different EMRs.

Chronic disease monitoring: Manually tracking HbA1c trends, blood pressure logs, medication adherence—all the longitudinal synthesis that should be automated but isn't.

Patient communications: Responding to portal messages, medication refills, care coordination—all the asynchronous work that happens after the visit ends.

A typical DPC physician spends 2-3 hours daily on this administrative work. They traded insurance complexity for operational complexity. They're still not practicing medicine at the top of their cognitive capacity.

Building Intelligence That Serves Clinical Reasoning

This is why I'm founding Ignite—to build AI infrastructure specifically designed for how DPC physicians actually practice medicine. The technical architecture we're planning reflects a fundamental commitment: clinical truth over billing optimization.

Training on Clinical Narratives, Not Billing Data

We're designing our models to learn exclusively from DPC clinical narratives. This isn't just a data source decision—it's a philosophical one. When AI trains on documentation focused on patient well-being, disease progression, and treatment efficacy rather than ICD-10 optimization, the patterns it identifies reflect genuine clinical signals: actual disease progression, subtle status changes, nuanced treatment responses. Not coding strategies designed to maximize reimbursement.

Architecture Choices That Enable Longitudinal Care

We're building around Mamba architecture—based on Structured State Space Models—specifically because it processes extraordinarily long sequences with linear computational scaling. A patient's health record spans years, creating sequences vastly longer than transformer-based systems can handle efficiently. Mamba's linear scaling means computational cost doesn't explode as patient histories grow, making it financially viable for a single DPC physician to manage 600 patients with true longitudinal intelligence.

This matters because real clinical reasoning requires processing complete narratives. Our system won't just flag that Mrs. Johnson's creatinine is elevated. It will synthesize: "Mrs. Johnson's creatinine has been trending up over 6 months (1.1→1.3→1.5). She started lisinopril 4 months ago. She mentioned increased ibuprofen use for knee pain last visit. Her BP has been more variable. Consider: medication adjustment, NSAID counseling, recheck labs in 2 weeks rather than routine 3 months."

That's clinical reasoning, not pattern matching. It requires processing the entire narrative of longitudinal care—exactly what DPC physicians provide and exactly what current EMRs fragment into discrete, disconnected encounters.

Automating Administrative Burden, Not Clinical Judgment

We're also designing administrative coordination agents using MacroSwarm principles—autonomous systems that handle medication price-checking, lab result tracking, referral follow-up coordination. Not to replace physician judgment, but to eliminate the cognitive burden of tracking tasks that don't require clinical reasoning.

The goal is simple: let physicians spend their 45 minutes on comprehensive diabetic counseling instead of 30 minutes on post-visit administrative coordination.

The Economic Case: Making Relationship-Based Care Scalable

Here's the reality most Americans face: the average family health insurance premium is $25,000 annually, with a $5,000+ deductible. Most Americans never reach their deductible, meaning they're paying premium prices for coverage they rarely use while still paying out-of-pocket for routine care.

A DPC membership ($75-150/month) paired with catastrophic coverage would cost less while providing better primary care access. But DPC practices can't scale efficiently because physicians still spend 2-3 hours daily on administrative overhead that has nothing to do with clinical reasoning.

This is where AI changes the economic equation. If we can reduce that administrative burden by 70%—through automated medication price-checking, intelligent lab result synthesis, and coordinated specialist referral tracking—we make relationship-based primary care economically viable at scale.

That's what Ignite is being built to do: create the intelligence layer that makes regenerative, relationship-centered medicine financially sustainable. Not by replacing physicians, but by amplifying what they do best—think deeply about their patients' health.