Why we started Pelica Health.

We left Google to start Pelica because the most expensive failures in American healthcare are not clinical. They are operational. They look like a duplicate outreach call to the same patient by three different teams in the same week. They look like a diagnosis that fell off a 2024 problem list and is now untrumped under V28. They look like a 47-step closure window for a single Stars measure that nobody on the floor can complete during a busy clinic morning.

You can hire more nurses, more coders, more analysts. None of it solves the underlying problem, which is that the data is fragmented, the workflow is fragmented, and the AI is glued onto the side of both. Healthcare has spent fifteen years digitizing records and the last three years bolting language models onto everything. The work itself has not gotten meaningfully easier.

What we kept seeing

Catherine spent years at Google building generative AI products and, before that, doing clinical AI research at Stanford School of Medicine and working as a medical assistant at the SF Free Clinic. Lalit spent a decade at Google and YouTube building large-scale distributed systems and AI agents. We came at the problem from very different angles, but we kept landing on the same observation: healthcare operations teams are running 12 spreadsheets per problem, because that is what every vendor's UI eventually collapses into.

The shape of the failure is consistent. A risk adjustment team has a list of suspect HCCs in one system, MAO-004 reject reasons in a second, encounter scheduling in a third, and a payer-specific portal for the actual submission. A quality team has hybrid measure gaps in one tool, ECDS-eligible measures in another, supplemental data in a third, and a chart-chase queue in a fourth. A pharmacy team can see PDC trends but not the EHR notes that explain them. None of these teams can see what the others are doing to the same member.

The buyer is not asking for "more AI." The buyer is asking for "fewer tabs."

What we are building

Pelica is the operating system for value-based care. Three layers, by design:

• One canonical record per member. Claims, EHR, pharmacy, lab, ADT, payer SFTP drops, and call recordings. Normalized. Resolved across NPI, TIN, contract, and member identity. Continuous, not snapshotted.
• An AI workforce of role-specific copilots: Risk Adjustment, Quality & Stars, Pharmacy & Part D, Provider Network, Care Management, plus a cross-cutting AI Data Analyst. Each one reads and writes to the same record, so the right hand actually knows what the left hand is doing.
• An action layer that closes the loop with members and providers: outbound voice, EMR overlays, a provider portal, and a coder workspace.

This is not a thin agent on top of a CRM. It is the system of record itself, with the AI sitting inside it.

Why now

Three things converged in 2025. First, V28 entered its final phase, and CMS made it clear that risk score normalization is permanent. Second, NCQA published the ECDS roadmap, which will increase supplemental data volumes 35x to 75x per measure by 2029. Third, foundation models finally got good enough to read clinical text reliably and operate inside structured workflows without breaking. The first two changes make the old vendor stack untenable. The third one makes the new stack possible.

So we started building. A physician-led IPA in New York managing 175,000 patients on a risk-bearing contract took the bet with us. The team reached 100% adoption, recovered roughly 10 hours per user per week, and hit 96% medication adherence on Pelica. Across Pelica deployments, gap closure has improved 41%, outreach capacity has tripled per coordinator, RAF has lifted by roughly +0.4 in two quarters, and customers have retired 12 separate tools and 5 point vendors. That is the story we want to repeat, contract by contract.

An invitation

We are early. The platform is real, the customer outcomes are real, and the road is long. If you are operating a risk-bearing contract and you are tired of running your team off twelve spreadsheets, we want to hear from you. If you build healthcare data systems and you want to work on the part of the problem that actually matters, we are hiring.

The rest of this blog will be short, operator-grade notes on the actual work: V28 traps, RADV chain-of-custody patterns, what we are seeing in PDC adherence, the ECDS data engineering reality. No thought leadership. Just what shipped this week and what we learned doing it.

The pieces of the bet, in more depth

If you want the operator-grade detail behind each claim in this letter, the posts below pick up where the letter stops:

• The exit ramp from analytics dashboards: AI agents vs analytics dashboards in value-based care.
• What an autonomous gap close actually does, step by step: What an AI agent does to close a care gap.
• The cost we are subtracting: The real cost of vendor portal sprawl in VBC.
• Why the BI ticket queue is the wrong place to ask questions: Replace the BI ticket queue with AI questions.
• The terms we use without defining: Glossary of value-based care terms.

Frequently asked questions

What founders, operators, and engineers usually ask after reading this letter.

Sources and further reading

• CMS: Medicare Advantage and Part D rate announcements (for the V28 phase-in trajectory and roughly 3% MA risk score impact)
• NCQA: The future of HEDIS (the ECDS transition and the retirement of hybrid measures by 2029)
• CMS: Medicare Star Ratings program (the structural reason quality bonuses move the unit economics of a risk-bearing contract)
• American Medical Association: 2025 prior authorization physician survey (on the operational drag of administrative work on clinical teams)
• Y Combinator: Pelica company page