Part 5: What Would I Build?

My grandfather, an engineer on NASA's Apollo missions, used to joke that rocket science was the simple part. You point the rocket in the right direction and light it. The hard part was getting thousands of engineers, contractors, and government agencies to point it in the right direction.

Over the course of this series, I've argued that American healthcare faces a similar coordination problem but without the shared mission control that made Apollo possible. We have consolidated fragmentation: powerful entities optimizing locally while the patient falls through the cracks between them. We have a value-based care model that works when it works - Patient 17 is proof - but is strangled by 18-month contracting cycles, data that is months old or missing entirely, and incentive structures that punish investing in the long term health of the American people.

So what would I build?

After nine years in this industry, across two companies, approximately ten insurance contracts, and tens of thousands of patients. If I could start over with everything I've learned, I would build two things:

1. A standardized contract and data broker

The single greatest barrier to scaling value-based care is not clinical - it's contractual. Every insurance company considers their method "standard." None of them agree with each other. A startup running three contracts with three payers delivers identical care but is functionally running three different companies with different target outcomes because the measurement methodology - not the medicine - determines the result. No matter how we would love clinical impact to drive contracting, it just rarely works out that way.

In Part 4, I proposed two paths: a centralized broker that provides standard terms and hosts data in an isolated compliant warehouse, or fully open-sourced contract templates and data interchange formats. I still believe either would advance the industry. But let me be more specific about what this would actually look like.

The contract layer. A standardized VBC contract template — open-sourced or federally maintained — with fixed structure and variable terms. The structure covers attribution methodology, quality measures, data sharing requirements, and financial reconciliation. The variable terms are the blanks: which patient population, which geography, what capitation rate, what shared savings split. Two parties should be able to negotiate the blanks in weeks, not years, because the structural questions are already answered. If this works, VBC providers should be able to contract with Medicare nationally.

The template would need to accommodate the real complexity I described in Part 4 — pharmacy carve-outs, regression-to-mean adjustments, comparison cohort methodology. But these would be solved once, openly, rather than reinvented behind closed doors in every negotiation. The Tuva Project is already doing this for claims data transformations. The same approach should extend to the contracts themselves.

The data layer. The broker hosts all required data exchange in an isolated, compliant environment. Insurance companies upload member files. VBC companies query those files via API without copying the underlying data. This collapses months of SFTP setup, security negotiations, and format reconciliation into a single integration. It also dramatically reduces the security burden — one environment to certify, rather than a bespoke security architecture for each payer relationship.

Today, getting to "can we even evaluate this contract?" requires HITRUST R2 and SOC 2 Type 2 certification — a process that takes years and consumes a significant fraction of a small company's total resources. A centralized broker with pre-certified infrastructure would eliminate that barrier for every new entrant.

Why this matters for patients. The 18-month contracting cycle doesn't just burn venture capital. It delays care. Every month spent negotiating data formats and security exhibits is a month where patients who could benefit from value-based care are stuck in a system designed around billing codes and quarterly financials. Compressing contracting from one year to one month means reaching patients a year earlier with more affordable services.

2. A universal patient data layer

In Part 3, I showed that the data powering value-based care is fundamentally broken. Claims data arrives three months late at best. When a patient changes insurance plans — which happens to 30%+ of Medicaid members annually — their medical claims history is zeroed out. The new insurer starts from nothing.

We spent weeks at Cityblock just correcting records. Not through software — through relationships. A community health partner would sit with a member for hours, building trust, then working through medication reconciliation, confirming diagnoses, updating contact information. This process was clinically valuable and humanly important. It was also an absurd duplication of effort — every VBC company in the country doing the same manual correction work because no shared data layer exists.

The patient data layer I would build has three properties:

It follows the patient, not the plan. When someone changes insurance — whether they switch jobs, lose Medicaid eligibility, or age into Medicare — their health record should not reset to zero. A universal patient identifier, paired with modern data-sharing agreements like TEFCA, could create a single longitudinal record that any authorized provider can access regardless of which insurer is currently paying.

It includes data that matters for health, not just billing. Claims data was designed for financial reconciliation, not clinical decision-making. It tells you what procedures were billed. It does not tell you whether the patient took their medication, whether they feel better or worse, whether they have stable housing, whether they've been exercising, or whether they just completed a substance recovery program and are at the most receptive moment of their lives for a preventive health investment.

The data sources that could fill these gaps already exist. Wearable devices track activity and sleep. Pharmacies track refill adherence. Patients themselves — if asked in the right context, by someone they trust — will share what their doctor never hears. At Cityblock, 60-80% of what we learned about a patient came from building a relationship, not from reading their chart.

It treats the patient as the owner. Most health data sharing efforts have put companies before consumers — optimizing for regulatory compliance rather than patient experience. The patient should be able to see, correct, and control their record. This isn't just a privacy principle; it's a data quality strategy. Patients are the only people who know their complete health story. Empowering them to correct the record — and incentivizing them to do so with the promise of better, more personalized care — could solve the data accuracy problem that no amount of claims reconciliation ever will.

What AI actually enables

I've been careful throughout this series to argue that AI cannot fix broken data. That remains true. An AI model trained on three-month-old claims with 30-50% coverage gaps will produce confident-sounding garbage. You cannot tell whether the AI is right or hallucinating when the underlying data is this flawed. The industry's rush to deploy AI without fixing the data foundation is my single biggest worry.

But once the data foundation is right — once there is a longitudinal patient record with real-time signals and patient-reported context — AI changes what's possible in two important ways.

AI makes the existing care model faster. Summarizing a patient's complete record before a visit. Flagging medication interactions across data from multiple prior insurers. Identifying gaps in preventive care — the missed screenings, the overdue lab work — and routing them to the right care team member. Drafting clinical documentation so providers spend more time with patients and less time typing. These are not revolutionary applications. They are operational improvements that reduce the 4.79 support-staff-per-doctor ratio and give clinicians back time for what matters: the relationship with the patient.

AI enables a fundamentally different care model. With rich, real-time data, some things become possible that were never possible before. Continuous risk monitoring that detects a deteriorating pattern before the crisis — not from claims filed months later, but from declining activity tracked by a wearable, a missed pharmacy refill, a change in sleep patterns. Predictive outreach that reaches a patient at the moment they're most receptive, not three months after they've been discharged. The ability to answer questions no one in healthcare can currently answer: did you take your medication today? Do you feel like your health is getting better or worse?

The advertising industry answers harder versions of these questions every day for the purpose of selling shoes. Healthcare should be able to answer them for the purpose of saving lives.

The coordination problem

My grandfather's generation solved a coordination problem of extraordinary complexity. They put a human being on the moon with less computing power than a modern smartphone. They did it by creating shared infrastructure — a single mission control, shared telemetry, a common language for every contractor and agency involved — and then pointing thousands of independently brilliant people at a single goal.

Healthcare has no shared infrastructure. It has three fortresses — insurers, hospitals, and providers — each with their own telemetry, their own definition of success, and their own financial incentive to guard their walls. A patient is assigned to a VBC program based on claims data that may be wrong. Their care team builds a relationship over months. Then the patient changes plans, the history resets, and a new team starts from zero. The long-acting injectable that was keeping them stable gets switched to a generic pill to save the next insurer money. The patient stops taking it. The ER visit follows. The cycle repeats.

A standardized contract broker and a universal patient data layer won't solve everything. Healthcare is genuinely harder than rocket science — the variables are more complex, the stakeholders more numerous, the history more entangled. But these two pieces of infrastructure would do what shared infrastructure always does: they would lower the cost of coordination enough that the brilliant, motivated people already working in this system could finally focus on the actual science of health instead of fighting the structure around it.

The rocket science was always the easy part. The coordination is what we haven't built yet.