Why today’s winning companies treat distribution as architecture
I’ve spent my career building partnership ecosystems, at Google and a range of health and tech companies. Companies have long celebrated partnerships as a key to growth, and yet I’ve never seen strong partnerships matter more than they do at the intersection of healthcare and AI.
The pattern I keep running into is that, in Healthcare AI, the model is rarely the thing that wins. What wins is distribution, getting the technology inside the clinical workflow where the systems already live.
In Healthcare AI, distribution is a partnership problem.
Consider what it takes to implement AI in a hospital. The model has to read and write to the EHR. It has to clear HIPAA and the health system’s governance review. It has to slot into the call center, the patient portal, and the scheduling stack - tools the organization bought years ago and typically resists replacing. None of that is a modeling problem; it’s all integrations and alliances.
And that’s just one organization. Zoom out and it’s worse. A single provider often runs dozens of systems that were never designed to talk to each other, and a patient’s information scatters across all of them, and across the payers involved in their care. Every one of those systems exists, in theory, to serve the person at the center of the equation: the patient. Yet because they don’t connect, no one, not the provider and not the payer, reliably has the right information at the right moment to actually help that person. That fragmentation is the real problem in healthcare, and it’s exactly why distribution is so valuable. The product that moves the right information into the workflow, across systems built to resist it, is solving something the whole industry hasn’t done nearly as successfully as it needs to do.
I first learned this while working on Google Health, where I led Ecosystem Partnerships. The team and I built a 150-plus-partner ecosystem, created “Connects with Google Health” (a co-marketing and technical-integration program for third-party developers), and managed partnerships with Stanford, Cleveland Clinic, CVS, Walgreens, and Fitbit that added value to the platform. After three years, we relaunched the product and usage got a boost, but it never reached the broad adoption Google was after, and in 2011, Google decided to shut it down the following year. That was tough to communicate to our partners, many of which had built their entire business around Google Health.
Google Health “v1” failed for a variety of reasons but, in the end, the usage curve never cleared Google’s high bar. I’ve reflected on this many times over the years. The integrations made Google Health usable and trusted, but the product itself stood outside the systems people already used and only a modest number of consumers came and stayed. The part that was hardest to swallow had nothing to do with the business: we had set out to make a real difference in people’s health, and we reached far fewer lives than we believed we would. It remains one of the greatest disappointments of my career.
The lesson wasn’t that the partnerships failed. It was that even strong partnerships can’t save a product sitting in the wrong place. The companies winning in Healthcare AI right now treat lessons like this as architecture.
I’ll look at three partnerships I think get it right, each solving the distribution problem a different way:
Abridge & Epic went deep: embedded inside the one platform that dominates the hospital, so the product lives where clinicians already work.
Hyro & Five9 went wide: plugged into the contact-center and EHR tools health systems already own, so it switches on quickly across many stacks.
OpenEvidence & NEJM/JAMA owned the trusted inputs: licensed the content clinicians rely on and went straight to the doctor, skipping the platform fight entirely.
Abridge: Deep inside Epic
Abridge builds ambient AI that listens to a clinical visit and turns it into structured notes. There are several credible ambient-scribe models now; the underlying science is no longer rare. Part of what has set Abridge apart is that its product lives inside Epic’s workflow instead of beside it, so the documentation is part-in-parcel with the clinician’s work, without needing to open a second app.
That one decision has drastically minimized adoption friction, which is much of why Abridge now runs in more than 300 health systems, among them some of the most recognized names in American medicine, and is on track for 100 million patient-clinician conversations this year. Being first and deepest inside the dominant EHR in U.S. hospitals was a procurement advantage no rival scribe could erase by training a slightly better model.
Where your product runs matters more than how good it looks in a vacuum. The partnership that puts you inside the system of record is worth more than another point of accuracy.Then, in February 2026, the landlord entered the business. Epic switched on its own native ambient charting, built with Microsoft, which listens during a patient visit, drafts the note, and queues the orders, all inside the very workflow Abridge’s position depends on.
It’s the clearest proof yet that distribution is the prize in this market. The platform owner didn’t respond to ambient AI by training a better model for someone else to deploy. It claimed the workflow itself.
Depth in someone else’s platform is essentially rented land, and in February the rent went up. In June, at a keynote in New York, Abridge CEO Shiv Rao unveiled what the company now calls an AI-native clinician intelligence platform: a strategic investment from Eli Lilly, a clinical foundation model built with NVIDIA, and an expansion well beyond documentation into billing, payer workflows, and trial screening. That’s a deliberate climb up the value chain, into territory a native charting feature doesn’t reach. Deep, it turns out, is not a resting state. It’s a race to become infrastructure the platform can’t replicate before the platform replicates you, and that race is run entirely in partnerships with EHRs, health systems, and, increasingly, pharma and chipmakers.
Hyro: Wide across the stack
Hyro takes patient-access work (calls about scheduling, refills, and routing) and turns it over to healthcare-specific AI agents. The model matters, but the more instructive story is the go-to-market.
In May 2026, Hyro joined Five9’s AI Agent Connect program as an accredited ISV, the only vendor to do so, as far as I can determine. That makes Hyro something a health system can switch on inside the contact-center platform it already runs, which Hyro says can turn a typical two-week integration into a single one-hour meeting.
Add its Epic integration, and Hyro stops being just another vendor to evaluate and instead becomes a capability in tools you already own.
The tradeoff is the mirror image of Abridge’s. Wider buys reach and speed: no single gatekeeper, fast yeses, presence across the call center, the website, and SMS. What it doesn’t buy is an irreplaceable position, because easy to switch on can also mean easy to switch off. Deep risks dependency while wide risks replaceability. The craft is matching the strategy to the structure of the market you’re in: deep where one platform owns the workflow, wide where the buyer’s stack is plural. Both are partnership decisions before they’re product decisions.
Meet the buyer inside their existing ecosystem. The fastest path to a health system saying “yes” is often that a partner is already integrated.
OpenEvidence: Own what the others can’t copy
OpenEvidence is a clinical-evidence tool that answers a physician’s question at the point of care and shows its work, citing the literature behind every answer. Plenty of companies can build a medical chatbot. What is harder to copy is the position OpenEvidence built first: it went straight to the journals clinicians already trust and made licensed, citable evidence the core of the product.
It is the official AI partner of the New England Journal of Medicine and JAMA, with content agreements reaching other respected publishers and societies as well. That means its answers are grounded in the sources a doctor already considers authoritative. But the journals are signing with others too (Abridge struck its own NEJM and JAMA content deals earlier in 2026), so the licensed-content advantage is real, but no longer OpenEvidence’s alone.
The durable moat turns out to be what the content bought - first-mover trust, the official-partner brand, and a free product physicians already reach for by habit.
Then it gave the product away. OpenEvidence is free to any NPI-verified U.S. clinician, and it spread the way consumer software does, doctor to doctor, until it reached more than 40% of U.S. physicians and was soon handling upwards of 15 million clinical consultations a month. No procurement cycle, no platform gatekeeper, no integration project. The distribution channel is the clinician’s own habit.
That position is now drawing fire. In spring 2026, OpenAI launched ChatGPT for Clinicians, free to any verified U.S. clinician and built around cited clinical search, and Google kept pushing its medical models. The biggest labs are now aiming at the same doctor, which tells you the value of being the tool a clinician reaches for is no longer a secret.
What a newer entrant cannot quickly buy is the head start: OpenEvidence is already the default for a large share of the physicians these rivals are courting. That lead is a distribution advantage, not a model advantage, and that is exactly the point.
You don’t always have to win shelf space inside someone else’s platform. Freely put the evidence people already trust in their, become the habit, and they’ll walk you through the door themselves.
Why this is structural, not a phase
It’s tempting to read all this as a temporary state that ends once models commoditize. I think it’s the reverse. Healthcare is uniquely hostile to standalone software because regulation is heavy, EHRs are dominant and sticky, buyers tend to be risk-averse, and the existing stack is treated like a sacred cow.
Or, at least, an immovable cow.
Those conditions don’t loosen as models improve; they make the integration-and-alliance layer more decisive, because that layer is what turns a capable model into something a clinician will actually use at 7 a.m. on a Tuesday.
None of this says the model doesn’t matter. It is important, of course, but not sufficient. The partnership gets you into the workflow, and model quality keeps you there at renewal. But among genuinely comparable models, and in these categories there are several, distribution decides who gets the chance to prove it.
What I’ve learned
I’ve watched the pattern repeat everywhere I’ve worked over the years, most recently in building AI-centric partnership strategies across a portfolio of digital-health partners, and, on my own time, in designing and running a working system of more than thirty specialized AI agents to manage a range of professional and personal operations.
That recent project has taught me what integration actually demands of AI: state, context, handoffs, orchestration.
A few years ago, I had a consulting engagement where I evaluated the landscape of healthcare AI engines for a company that integrated many of them into a single product, comparing the models against one another, deep in the weeds of what actually separated them. I came away wholeheartedly convinced that, in Healthcare, the integration and partnership layer decides more than the engine does, the same lesson now playing out across the field, with higher stakes and stickier incumbents.
The models I’m using for this project are impressive in how much value they add to my life, professionally and personally. But it’s their partners’ seamlessly integrated products that multiply the value I get out of them.
As I see it, the teams that treat partnerships as a revenue function bolted onto engineering tend to plateau. The ones that treat the partner ecosystem as core architecture (who they integrate with, which platforms they live inside, whose stack they plug into) tend to grow exponentially.
So if you’re building or buying healthcare AI right now, push past “How good is the model?” Ask where it runs, who already trusts the rails it runs on, and who owns the room your buyer sits in. Those are partnership questions, and they are the ones that predict adoption.
There’s one more question, underneath all the others. Nearly everyone I’ve met in Health Tech came to it because, as cliched as it can sound, they make a difference in people’s lives. I generally believe them; that’s the case for me. Like so many others I talk to about this desire, I’ve also felt what it costs to fall short of that. The partnerships and the plumbing aren’t the point, they’re how the point reaches the patient, and if the right information never arrives, the smartest model in the world has helped no one.
Chad’s Business Channel is where the professional writing lives: articles on healthcare AI, partnerships, and go-to-market, written for the people building and buying in that market. Each piece publishes on LinkedIn as well, where most of the conversation happens. Read the full collection on LinkedIn.