🏥 An infinite workforce for primary care?
 |
|||||||||
|
|||||||||
|
|
|||||||||
I spent time recently with Jean-Claude (JC) Saghbini, Lumeris' CTO, to get more of the insider perspective on how they're scaling Tom. Something incredibly provocative he brought up was how AI-enabled primary care is creating a new paradigm for the industry. So today we're talking about this paradigm shift with AI-enabled primary care as a service, diving into:
We'll also discuss a provocative question... “How would you design primary care differently if you had an abundant, near infinite workforce?” All of these discussion points drive toward something bigger - the platform-ization and orchestration of care delivery models, at scale. Whether it gets there is the $64 billion question, and I'm excited to keep covering it as it plays out. Inside Tom: Lumeris Is Quietly Building Primary Care's Operating SystemI've been turning the same question over in my head lately: what does primary care actually look like five years from now, in a world where the labor constraint goes away? It's the question I keep circling in my own coverage — through the War of Five Kings piece on healthcare AI consolidation, through Photon Health and the ambulatory pharmacy thesis, through every earnings call where another health system CFO walks investors through their wage inflation problem. The labor crunch in primary care is structural and getting worse, and the industry's default response — recruit harder, pay more, scope-of-practice expansion at the margins — is, frankly, a losing strategy. Which is why I keep coming back to what Lumeris is building with Tom. I spent some time recently with Jean-Claude (JC) Saghbini, Lumeris' CTO, on what was billed as an open-ended conversation but ended up being a much deeper download on the technical, philosophical, and strategic decisions underneath this product than I was expecting. I came away with notes on the Google partnership, the new product cadence, the architecture of their monitoring stack — but the bigger takeaway was a clearer thesis about why Lumeris specifically is positioned the way it is. So let me share where my head is. Pace is the moat. Procurement cycles are the punishment.Here's the thing I think most health system executives still underestimate about the AI moment we're in: the bar your patients are now using to evaluate your digital experience is being set by their banking app, their Instacart, and their daily ChatGPT use — not by your peer health systems. That's a meaningfully different competitive dynamic than healthcare has dealt with before. For decades, the comparison set for a health system's patient-facing tech was the health system across town. Now the comparison set is every consumer surface a patient touches in their daily life, and those surfaces are upgrading themselves on hyperscaler release cycles rather than healthcare procurement cycles. JC framed it well — that consumers and patients are the same people, and the AI capabilities showing up in their consumer lives at near-zero latency are now the implicit benchmark for what healthcare AI ought to feel like. I've been making a version of this argument in Hospitalogy for a while, but he put a sharper point on it than I'd been articulating: if your healthcare AI doesn't feel as good as the consumer AI your patient used five minutes ago, your AI feels broken even when it isn't. The Tom Native Audio launch is the cleanest expression of this. Lumeris shipped it roughly a month after Google released the underlying capability, which is not a coincidence — they were building to it before it released. That's the cadence problem. If you're a health system building your own AI internally, or relying on a vendor that ships annually, you are already behind. And you will continue to fall behind, accelerating. Pace is the moat. It's high time the strategy conversation caught up to that reality. Tom wakes up with a job to doThis is the part of the conversation that has been rattling around in my head for days, and I haven't seen anyone in the space articulate it this cleanly. Almost every AI product we interact with — the GPTs, the Claudes, the Geminis — is reactive. We bring the agenda. We open the chat, we type the prompt, the model serves us. That's the consumer AI paradigm and that's most of what's being built for healthcare today: scribes that summarize what the clinician already did, copilots that draft what the user is already trying to write, search tools that respond when asked. Tom flips that around. Tom wakes up with an objective and works on the patient to accomplish it. Tom is doing the agentic work, the one trying to get something done — close a care gap, complete a follow-up, navigate symptoms — and the patient is the one Tom is trying to move. The more I sit with it, the more I think this is the actual fork in the road for healthcare AI. The reactive-assistant flavor is useful at the margins; it shaves minutes off clinician workflows, it cleans up documentation, it's table stakes. The proactive-agent flavor is what actually changes the economics of care delivery, because it's the only version that scales the care team itself rather than just making the existing care team slightly more efficient. And it's a dramatically harder engineering problem. When the AI is the one driving, every conversation has to navigate clinical guardrails, patient trust, regulatory boundaries, and the question of when to escalate. You're building a teammate patients have to actually trust, which is several orders of magnitude harder than building a smart Clippy. Most "AI in healthcare" announcements I've covered this year aren't in this category, and I think the gap between the two will become very obvious very quickly. An operating system, not another appI've been a little allergic to the "X is the operating system for Y" framing in healthcare, because it gets thrown around loosely and usually means nothing. But it actually fits here, and I'll tell you why. What Lumeris is shipping is a coordinated stack of agentic capabilities that share a common substrate — the Tom persona, the underlying clinical reasoning layer, the Hawkeye monitoring infrastructure, the connectivity into the health system's data and care plans. New capabilities are being plugged in on a cadence that JC told me is roughly every two to three weeks. Native Audio. Ask Tom (the natural language analytics layer). Inbound symptom checking and administrative triage. A new mobile communication modality that explicitly is not another healthcare app. Best Clinical Guidance (BCG) on the clinician side. The pipeline is dense and shipping. A couple of things I want to flag for Hospitalogy readers specifically. First, the explicit decision not to build another healthcare app is the right call, and I wish more vendors in this space had the discipline to make it. The last thing patients want is yet another portal credential. The modality Lumeris is building is a style much closer to the apps people use daily to communicate with anyone they trust. That's the move. Second, Lumeris is in market with one of the more sophisticated health systems in the country, doing patient-facing work right now. And the Collaborative for Healthy Rural America (CHRA) — which is bringing in additional partners including Instacart — is the deployment context I'm most interested in, because rural primary care is where the labor constraint has already broken. Tom in a rural setting, with a labor situation that's catastrophic rather than just tight, is where this technology gets stress-tested for real. There's a credibility question every AI vendor in healthcare has to answer: can you actually deploy this at scale? Lumeris has been operating for the past 15 years at a national scale. That's not nothing. The Google partnership is the smartest call Lumeris has madeIf I had to point to the single most strategically important decision Lumeris has made, it's the depth of the Google Cloud partnership, and I don't think that's a popular take yet. Let me make the case. The conventional story on these AI partnerships is: we're a Google shop, we get GCP credits, we use Gemini. That's the press release. The actual relationship Lumeris has built is meaningfully deeper than that, and once JC walked me through it, the strategic implications got a lot bigger. What Lumeris gets, as I understand it: pre-release access to Google's frontier models well before public availability, which is why they were able to ship Tom Native Audio in lockstep with the underlying capability rather than chasing it. Direct line-of-sight into Google's healthcare research teams — the people writing the Med-Gemini papers — which is roadmap influence at the foundation-model layer, not just the application layer. And the practical reality that Google's team has offices across the street from Lumeris in Kendall Square, which I'd argue matters more than any contract clause. Now, here's why I think this is strategically underrated. The healthcare AI market is going to bifurcate sharply over the next 18 to 24 months between vendors building on top of frontier models that are improving on a quarterly cadence and vendors who picked a model in 2024 and are now stuck. The latter group will look fine for a while, then they will get lapped, hard. The companies that win are the ones that can absorb every model upgrade as a tailwind rather than a refactor. Lumeris' bet is essentially that Google is one of the two or three companies that will still matter in frontier AI in five years — a bet that, given Google's full-stack position from TPUs to data centers to model training to underlying data assets, is one of the safer bets in tech right now. Picking the right partner at this layer of the stack is a capital allocation decision, and Lumeris allocated well. Hawkeye watches everything. Nobody watches your nurses.This is the part of the conversation where I pushed back the hardest, and where I think JC made the point that I'm going to keep making in this newsletter going forward. Every Tom interaction — every word, every reasoning step, every pause, every patient response, every conversational artifact — is logged and monitored. The platform doing the monitoring is called Hawkeye, and JC's framing was that if Tom is the agentic primary care workforce, Hawkeye is the workforce management platform watching the entire fleet in real time. Roughly 85 distinct dimensions are being monitored: clinical guideline adherence, guardrail compliance, conversational delays, patient intonation, repetition, hangups. Pick a criterion. Hawkeye flags it. Now sit with that for a second. There is no equivalent monitoring infrastructure for human care delivery anywhere in American healthcare. Nobody is listening to every nurse triage call. Nobody is reviewing every front-desk interaction for tone. Nobody is flagging when the practice manager rushes a patient because they're running late for lunch. The actual operational baseline of human-delivered care is, candidly, a black box. We don't measure it, we don't audit it, and we don't really know how often it goes wrong, because we don't look. And yet — and this is where I got energized — we are holding AI in healthcare to a standard of perfection we have never, at any point in the history of this industry, applied to human clinical labor. Every Tom hallucination becomes a public incident. Every off-script moment becomes a referendum on whether AI belongs in care delivery. Every edge case becomes proof that the technology isn't ready. Compared to what?JC pointed me to a piece by Zak Kohane at Harvard from a few years back, titled Compared to What? — and I'd encourage every Hospitalogy reader who hasn't seen it to go find it. The core argument is exactly what the title suggests: the safety bar we're setting for AI in healthcare is not being benchmarked against the actual quality of human-delivered care, which is far worse than the industry likes to admit. It's being benchmarked against an imaginary, idealized version of human care that doesn't exist anywhere. This is the conversation healthcare media has been ducking, and I want to start having it directly. Because if the bar for AI deployment is "must outperform a clinician who doesn't actually exist," we will reject every AI rollout that would — in practice, measured honestly — dramatically improve outcomes against the real-world baseline. What we're calling caution is actually a policy failure dressed up as caution, with a body count attached to it in the form of every patient who didn't get the care they needed because the labor wasn't there to deliver it. No AI vendor can credibly lead this argument, because it'll read as self-serving every time. This is a job for independent voices in the space. Count me in. What would you actually do with an infinite workforce?The single line from this conversation I keep circling back to was JC quoting a health system CEO he'd been working with. Mid-meeting, this CEO stopped the room and asked, "What would we do with an infinite workforce?" That is the question. And almost no health system I cover is actually asking it. What I see, instead, is the OpEx framing. Bob does five steps in this workflow, AI does three of Bob's five, we need fewer Bobs. That's workflow change management. It's fine. It saves money at the margins. It will produce, at best, modestly cheaper versions of the same business model these systems already operate. The real question — the one Lumeris and a handful of other companies are quietly building toward — is what happens when the binding constraint of care delivery, which has been human labor for the entire history of the industry, becomes effectively unconstrained. Panel sizes change. Follow-up cadence changes. Preventive care economics change. The set of patients you can credibly serve changes. The geographies you can serve change. The lines of business that become viable change. What you're really redesigning is the value proposition itself, not the workflow. The health systems that internalize this early are going to look fundamentally different from the health systems that don't. This is, I'd argue, the most consequential strategic split we'll see in healthcare delivery this decade. The cost-takeout shops end up with marginally cheaper versions of what they already have. The capability-unlock shops end up with businesses their peers can no longer compete against. I have to wonder, looking at my readership of health system executives, investors, and operators: which side of that line are you actually building toward right now? Look at the budget, not the strategy deck. My takeWhere I land: Tom is the first credible attempt I've seen to build primary care's operating system on top of frontier AI — with the philosophical clarity and infrastructure discipline to actually pull it off. I've covered a lot of AI in healthcare. Most of it is workflow point solutions chasing reimbursement. A meaningful slice is foundation-model wrappers chasing a use case. Almost none of it is trying to build the connective tissue that lets an entire model of care delivery run differently. Tom is in the last category, and the cadence of what's about to ship — Native Audio is already live, Ask Tom is announced, inbound symptom checking is deployed, the new mobile modality is in development, BCG is being built for the clinician side — suggests the next 12 to 18 months are going to be the period where this thesis either gets confirmed or doesn't, in production, in front of real patients, at scale. I'll be watching closely, and I'll keep writing about it as the announcements drop. Because the question of what primary care looks like in an infinite-workforce world is the question for the rest of this decade, and Lumeris is one of the very few companies actually building toward the answer rather than just talking about it. Onward, Hospitalogists. Big things coming. The operating system for primary care is here. Explore Tom at Lumeris. This essay is a sponsored post in partnership with Lumeris. I write these posts for companies with products or missions I believe can provide value-adds for Hospitalogy subscribers, many of whom work with/for ACOs, FQHCs, integrated health systems, health plans, and other risk-bearing organizations that want to learn more about potential value-based care partners. If you’re interested in a sponsored deep dive, please reach out to blake@workweek.com! Thanks for the read! Let me know what you thought by replying back to this email. — Blake |
|||||||||
|
No comments