The $320M reason access beats revenue cycle

Napkin math on why primary care panels are the next capital fight. ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌

Blake Madden Sep 28th, 2026 Read in browser

In partnership with Happy Monday, Hospitalogists! Today, I'm diving deeper into Tom, Lumeris's AI platform. One of my biggest takeaways this time around: access, not the EHR, is where health systems have to fight for differentiation now. Epic just shipped enough AI at UGM that your two closest competitors will have the same digital experience you do next month if all three of you are EHR-first and EHR-only. Let's get into it. Meet Tom. Part 2. Late on a Monday afternoon at ViVE last year, with the conference day winding down and most of the floor already drifting toward happy hours, I walked into a Lumeris meeting room and found David Carmouche and Jean-Claude Saghbini beaming at me like two school kids who'd just found the last Pokemon card pack hiding at the back of the shelf (and inside was a Mewtwo SIR). What they showed me that afternoon had a 3-letter name and a product that would reshape the trajectory of the company: primary care as a service, built as a multi-agent platform, aimed at the 100 million American adults who can't reliably get primary care today. I wrote it up that week. I've written about it 4 times since. And over the course of my journey with Tom and Lumeris to date, the company went from that room to a deployed enterprise platform at Ochsner Health in roughly 18 months, with more in the works, and no signs of slowing down. Human labor has been the binding constraint on care delivery for the entire history of this industry. Every model, panel, access strategy, payment code, has always been dealt with in the same way: “just throw more people at it.” But people are scarce, and the introduction of AI presents an opportunity to reskill and redistribute people toward human-facing, top of license roles rather than the mundane copy and paste tasks healthcare administration has bogged us with to date. Primary care is one of the first battlegrounds for reimagining what care delivery and workflows - both administrative and clinical in nature - look like in an AI-native future. And Tom is one company at the forefront working at a juncture of some of healthcare’s toughest problems. How do we solve the primary care access problem? What would care delivery look like reimagined with a potentially infinite workforce? How might we build workflows from the ground up in an AI-native future? How does access shift and consumer preferences change in primary care over time? Tom sits at the forefront of where healthcare is headed with multi-modality, AI-native primary care. So before we get into the future (which is well on its way), I wanted to take this newsletter to distill down with you the lessons I’ve gathered in working closely alongside Tom’s and Lumeris’ leadership, and how they’re rethinking primary care for hospitals and health systems. Access is the new AI battleground Healthcare AI capital has been allocated with remarkable uniformity, and for a defensible reason. Systems are funding transformation out of operating margins running 1% to 3% in a good year, with a recovery driven substantially by investment income rather than operations. Layer Medicaid reconciliation, site-neutral expansion, and 340B on top and hundreds of millions of dollars a year of exposure shows up at the enterprise level. Then AI walks in asking for money. So the gate, naturally, became hard ROI. A named P&L line, verifiable in 2 or 3 quarters, with soft-dollar savings and cost avoidance disqualified on arrival because attribution is impossible and CFOs got tired of being sold it. Consequently, health systems have spent 3 years pointing AI at the back office. Ambient documentation, revenue cycle, coding, denials, prior auth, credentialing. That work, while ongoing, is now table stakes from a roadmap / prioritization perspective especially considering the minimized risk from a liability perspective in deploying in this domain. But now leaders are thinking hard about patient-facing AI as folks across the industry talk about autonomous clinicians and whether we’re accelerating technology too quickly. Nevertheless, access (which is a really broad term encapsulating everything having to do with answering the question “how do we acquire and retain patients and assure our served population gets care when they need it?”) is the next domain. It's also arguably the only one left where a health system can still build a durable advantage or differentiated experience. Yet so many of them still cede this vessel to MyChart. Every health system in the country needs to be differentiating away from its EHR, whether or not leadership currently believes it. Good enough is not good enough. And the clock on that is short, because you now have an imperative to deliver a delightful experience to your own customers or cede volume and demand to tech-forward entrants everywhere outside the ED, lest you risk becoming a ghost kitchen of proceduralists. Ambient documentation is now recruitment table stakes — try hiring a physician into a system that doesn't have it. Autonomous coding is live. Denials, appeals, prior auth, credentialing, policy consolidation, provider data. All correct work, all being done, all converging. Here's the problem with converging. Your competitors are running the identical playbook on the identical platform, and increasingly buying it from the same vendor you are. Epic's 2026 UGM put roughly 120 adaptable AI features into Agent Factory, shipped autonomous coding in Penny, put a patient-facing assistant in the portal with Emmie, and has Ergo arriving in November as an AI-native clinician interface. Whatever you think of the quality, the strategic reality is that next month your two closest competitors will have the same digital experience you do if all three of you are EHR-first and EHR-only. Congratulations to everyone on achieving parity! I'd argue the old heuristic is dead too. Health systems used to buy best-of-breed on the theory that the EHR ships a C-plus version of everything. That was true when the comparison was an in-house data science team against a vendor's. Now everyone is prompting the same 3 frontier models, the gap between tools has compressed to something like A-minus versus B-plus, and the incumbent wins on integration alone. Which means the honest question is no longer "is this better than Epic's version." but rather "is Epic even trying to build this, and if they are, why am I paying twice." Ambient, coding, and denials are all inside the blast radius of that question. Access mostly isn't. And access is where the differentiation actually lives, because it's the one thing patients experience directly and the one thing most health systems in America aren’t very good at. Nothing here is an anti-Epic or anti-EHR argument. Epic-first is the correct default. You've probably spent 9 figures, the integration advantage is real, and a vendor's mediocre model wired directly into the chart routinely beats a superior model that lives in another tab. Use all 4 quarters of what you bought. But the system of record is, by design, a convergence machine. Its job is to make every customer function the same way, which is precisely why it cannot be the source of your differentiation. Keep it as the record and the data substrate, then create the layer that acts. Work with partners who tackle an entire domain. Like Tom does. Thinking about primary care and access differently Nearly every technology leader I talk to has arrived at the same conclusion in the same words: technology has stopped being the primary constraint in healthcare. Even with suboptimal tech, people and process (AKA, change management) are the primary limiting variable. Bolt AI onto a broken process and you get a turbocharged broken model. If you drop an agent into a care model designed around 15-minute visits 3 times a year, you've built an expensive reminder system. This model is more or less our fee for service primary care business model today. David Carmouche at Lumeris’ experience was similar. We have artificially distilled primary care into a handful of short office visits, "as if the other 361 days of the year, you know exactly what to do to stay healthy." In this antiquated model, you cannot run a proactive, AI-orchestrated model of care through that cadence. So AI-native primary care means something different. A new paradigm, model, whatever fancy word you want to use - designed for the 361 days first and treating the visit as the exception. Ochsner Health down in Louisiana experienced this problem firsthand. For some context, I had a chance to sit down with Denise Basow, Chief Digital Officer at Ochsner and remarkably transparent and candid about their initiatives. Here’s more on Ochsner’s journey based on this previous conversation which will help contextualize their local market dynamics and pain points around primary care and access. Louisiana is a hard place to run a health system. 40% of the population sits in Medicaid, reimbursed below Medicare in most cases, with a single dominant Blue holding commercial rates down. To be a growing, profitable system in that payer mix, you have to be innovative by definition. Ochsner had already run the conventional plays and run out of them: e-visits and asynchronous visits turned on in Epic, a real telemedicine program, a large urgent care network operating as coordinated overflow so primary care patients weren't getting shuttled into the ED. Then leadership handed primary care an organic growth target. When primary care ran that number through the traditional model, the answer came back at roughly 162 new primary care physicians, plus the clinics to house them. Money to build or buy new clinics aside, there weren’t 162 PCPs chomping at the bit to be recruited. What made Ochsner ready for Tom wasn't enthusiasm for AI but a prior disappointment with it in deploying to its radiologists and not seeing any productivity gains or extra reads. So, lesson learned on that hard ROI front: you can't afford to invest in relatively expensive technology that doesn't come with a business case. Which is also why Tom had to be workflow-invasive to be worth doing at all. The AI that creates ROI embeds itself in important workflows, and if it doesn't, it isn't creating much value. Lessons from Ochsner What can we learn from AI deployments at a health system like Ochsner? Trust is AI implementation currency. Before Tom existed in the building, Ochsner went after the pain clinicians actually feel. Ambient documentation to kill the 10 p.m. note-writing. In-basket reduction to give time back. Rolling out physician friendly tooling across thousands of clinicians reducing cognitive load and giving hours back built real equity with the medical staff. That equity is what bought permission to attempt something more ‘intrusive’ as rebuilding primary care around an agent (AKA, requiring more change management and adaptation). Plus, ambient rollout acts as a retention tool. Once you go ambient, you never go back. The experience is just that much better. They refused to call it a pilot. Ochsner committed to enterprise deployment up front, then stood up 2 to 3 model clinics led by physicians who already lean into innovation. A pilot is something that might fail and it’s not the right characterization for something you’re implementing as a strategic priority. A phased implementation is something you've already decided to do. Conflating the two burns change-management capital you don't get back. You only have so much of that human social capital. They let their own people tell the story. At the primary care launch town hall, 300+ people showed up, the largest primary care town hall Ochsner had ever held, and the system produced its own video laying out the plight of the primary care physician today and positioning Tom inside that. They ran two different narratives internally. One with the C-suite, with hard ROI metrics. Then one with clinicians, which was slower moving. Doctors are skeptics, doctors are scientists, doctors want proof, and they’ve been practicing in their own way for their entire lives. Nobody flips a switch and goes from 20 in-person visits a day to 7 visits plus 120 asynchronous AI-enabled encounters. That happens at a pace clinicians can absorb, or in training over time before or during residency, or it doesn't happen. It makes me think about someone telling me to stop using Excel since it’s not as good anymore even though I’ve been using it for charts my entire content creation and consulting career. I shudder to think of that future. Meanwhile, in Maine Ochsner's math extends beyond Louisiana. This is a national issue and you hear it in disparate markets that look nothing like New Orleans. Take Dr. Rob Chamberlain, chief of primary care at MaineHealth, who I hosted on a virtual session earlier this year where we chatted alongside Lumeris about the future of primary care and agentic AI. MaineHealth runs 9 hospitals and roughly 2,000 physicians, about 400 of them in primary care, in the state with the oldest population in the country. And to note, Rob isn't running Tom. But he's running the problem Tom was built for. His communities need more primary care, yet there's a shortage of practitioners and support staff whom he can engage. More than the patient who sees him, then sees him again in 4 months to follow up on diabetes with an A1c sitting at 9. More than the PCP panel he already can't staff. He's fielding calls from specialists, the ED, and his own leadership team about people who just moved to Maine and can't get established anywhere. And there's nobody to hire. Not a budget problem - a supply problem. They can't hire their way there, and even if we had all the money available to us, the people aren't out there. So what fills that gap? Almost every AI product in healthcare today waits to be asked. Scribes summarize what a clinician already did. Copilots draft what a user is already writing. Portal assistants answer when a patient opens the app. That's the reactive paradigm we inherited from consumer AI, and it shaves minutes off a workflow without laying a finger on the capacity problem. Tom inverts the dynamic. Tom wakes up with an objective and works on the patient to accomplish it, and the patient is the one being moved rather than the one initiating. Which is a fundamentally different thing to buy, and a different thing to manage. Rob brought up a framing he's been hearing for agentic AI as a care team member - the infinite intern - and I haven't been able to shake it since. Hopefully in 30 years that intern hasn't worked its way up to running the health system. Here's what an AI care team member handles in practice: Takes the history before the visit and hands you the summary, so you walk in ready for the conversation that needs a human Works the in-basket Calls Mr. Jones about FIT versus Cologuard versus colonoscopy, walks the tradeoffs, and reports back which one he picked Checks whether the hypertension script ever got picked up, and finds out what's in the way when it didn't Follows up after the ED visit, documents back into the chart, escalates when the answer isn't reassuring Rob also named another huge gap in primary care, and one that leads to a dismal patient experience: we're awesome at building care plans with patients, and then…we do almost nothing to help execute them. Every system in the country has a care gap dashboard. Very few have a mechanism that closes the gap. State of the art at a lot of organizations remains bulk outreach. 30,000 MyChart messages, a 10% response rate, and a slide that calls it patient engagement. Anyone happen to know if that patient picked up their medication? Maybe we'll find out in a few weeks if they end up back in our ER. Rob's replacement framing is simple: can we get to an abundance of engagement? Maine is also further along than most on the part everyone finds inconvenient, which is payment. MaineHealth's CEO is a family medicine doc who believes fee-for-service fundamentally doesn't work for primary care, so the system funded Trellis Health, a wholly-owned primary care organization on a subscription model contracting directly with employers. Out from under the fee-for-service yoke, free to integrate physical therapy and behavioral health and anyone else who can't drop a CPT code. Meanwhile the state is looking at primary care capitation in Medicaid, and Rob's team is rebuilding cost and revenue accounting internally so panel growth shows up as something a CFO can recognize. That's a health system choosing to compete on access before somebody else did it to them. If you aren't having that conversation, an employer coalition in your market is having it with someone. Primary care economics and consumer preferences are changing. Quickly. Here's what lets access compete for capital against revenue cycle. Carmouche's napkin, with the assumptions labeled: 200 employed PCPs at 2,000 patients each = 400,000 attributed lives Roughly $8,000 of annual healthcare spend per patient, of which the system captures about 50% depending on payer mix and network configuration = $4,000 200 × 2,000 × $4,000 = $1.6B of captured revenue Grow panels 20% and you're looking at roughly $320M of incremental captured revenue, against a software fee that rounds to a rounding error Now run it backward, because, in my opinion, this is the opportunity cost of not acting on access and changing consumer demand preferences. Health systems are facing a significant shift in demand, starting with primary care, driven by consumer preferences. Resulting in fragmentation and material competitive threat to your organizations. That erosion doesn't announce itself, either. It shows up as urgent care volumes flattening in markets where virtual-first and retail options matured. As the 28-year-old who never establishes with a PCP because a subscription service ran her labs and a chatbot interpreted them. As a referral that goes to a competitor's orthopedist because your provider directory hasn't been touched since 2019 and theirs has. And the behavior is already normalized with the stats we’ve all seen on consumer preferences with using ChatGPT and other interfaces. 30% or more adults are using AI for health queries. Now imagine if verticalized health AI like OpenEvidence goes consumer-facing too. Whoop is entering CMS’ ACCESS model. Oura is going public and already has partnerships with Counsel Health and payors. On the diagnostics side, consumer memberships now cover 160+ biomarkers through national lab networks, with $499 whole-body MRI bolted on. And let’s not forget about the managed care players investing heavily into consumer-facing navigation while employers explore direct contracting models. The world is changing. Lose that upstream relationship and you keep the clinicians, the buildings, and the capital stack while somebody else owns the customer. That's the ghost kitchen I mentioned up top, and it doesn't arrive as an event. It arrives as 5 consecutive years of slightly softer volume that everyone attributes to something else. What’s that Hemingway quote again? Pace creates the moat I spent time recently with Jean-Claude Saghbini, Lumeris' CTO, on what was billed as an open-ended conversation and turned into a great technical dive. Start with his read on the competitive landscape. For decades, the comparison for a health system's patient-facing technology was the health system across town. Now it's every consumer surface a patient touches in a day, and those surfaces upgrade on hyperscaler release cycles rather than healthcare procurement cycles. Consumers and patients are the same people. If your healthcare AI doesn't feel as good as the consumer AI your patient used 5 minutes ago, you’re considered ancient. Cleanest expression of that is Tom Native Audio, built on Gemini's native speech-to-speech stack. Real-time conversation, meaningfully reduced latency, better turn-taking, noise handling, multilingual on the roadmap. 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. Which brings me to what I'd argue is the most underrated strategic decision this company has made. The conventional version of these partnerships is: we're a Google shop, we get GCP credits, we use Gemini. Press release material. What Lumeris actually has is pre-release access to Google's frontier models well ahead of public availability, direct line of sight into the healthcare research teams writing the Med-Gemini papers, and offices across the street from each other in Kendall Square. New Tom capabilities land every 2 to 3 weeks. Native Audio. Ask Tom, the natural language analytics layer. Inbound symptom checking and administrative triage. A new mobile communication modality. Best Clinical Guidance on the clinician side. JC's hard rule is the one more vendors should have the discipline to adopt: Tom will never become another healthcare app. We have hundreds of thousands of healthcare apps in this country and have moved the national needle on neither cost nor outcomes. One more portal credential! Exactly what patients have been asking for. Instead Tom runs multi-modal the way a human teammate would, voice and text and video embedded in voice, matched to the context and to what the patient actually prefers. Tom is a platform, and the cornerstone partner in a successful AI-native access deployment. A 4-minute conversation between Tom and a patient is the demo-able tip of it. Behind that conversation: ingest data from the EHR, claims, labs, pharmacies, HIEs, remote monitoring and consumer sources; normalize it; construct a 360-degree view of the person; reason on that view to determine what should happen next; activate the right modality; hold the conversation; document it back into the electronic health record; close the loop with the care team; escalate and alert where needed; store it; surface analytics across patients and populations. Then wrap the entire thing in AI governance and safety monitoring. By the way, that last layer is Hawkeye, and it's actually an underrated differentiator for Tom that CIOs probably care a LOT about. Every Tom interaction is logged and monitored across roughly 85 dimensions: clinical guideline adherence, guardrail compliance, conversational delay, patient intonation, repetition, hangups. JC's framing is that if Tom is the agentic primary care workforce, Hawkeye is the workforce management platform watching the whole fleet in real time. Add 60+ models evaluated against clinical guardrails, a team working full time to find the unhappy paths, and 260,000 test cases running in the pipeline. Sit that next to the human baseline for a second. Nobody listens to every nurse triage call. Nobody reviews every front-desk interaction for tone. Nobody flags the practice manager who rushed a patient because she was late for lunch. Human-delivered care is operationally a black box, which gets us into a whole ‘nother argument on what we’re comparing AI to (perfection) as opposed to human error. JC pointed me to Zak Kohane's [Compared to What?], which makes the argument better than I will. We are benchmarking healthcare AI against an idealized clinician who exists nowhere, instead of against the mess that actually exists. Hold deployment to that bar and you will reject every rollout that would measurably improve care for the 100 million people currently getting none. The rural imperative CMS put $50B over 5 years behind the Rural Health Transformation Program. For scale, HITECH, which dragged EHRs into essentially every hospital in the country, ran about $26B. This is nearly 2x that, aimed at 26 million people, against a backdrop of 190 rural hospital closures since 2005 and 600+ more financially at risk. CMS has been explicit that it wants transformation rather than a subsidy for models that already failed. Innovation is the entry ticket, not a bonus. That's the opening the Collaborative for Healthy Rural America was built for. Tom sits in the center of various partners reimagining rural care as the connective tissue, orchestrating outreach, follow-up, and navigation across all of it, and plugging in local hospitals and providers who want to participate rather than routing around them. If you designed primary care from scratch for someone 2 hours from the nearest clinic, you would never start with brick and mortar. You'd start with something that looks a lot like this collaborative. For health systems, this is the rare case where mission and margin point the same direction. Rural access is in your community benefit report already. The RHT dollars are infrastructure capital with a deadline attached. And the model you stand up to serve a county 90 minutes away is the same model that defends your suburban primary care base 3 years from now. The Infinite Workforce JC told me about a health system CEO who stopped a meeting mid-sentence to ask the room what they would do with an infinite workforce. Almost nobody is asking that because it almost seems naive, or too myopic. What we’re all more comfortable with instead is the faster horse instead of the car. More or less, OpEx framing. AKA, Bob does 5 steps in this workflow, AI does 3 of Bob's 5, so therefore we need fewer Bobs. Streamlined. Boom, done. Good for another year on the P&L. Sure. This tactic saves money at the margins and produces a slightly cheaper version of the business you already run. But JC characterized this whole category as small ball, and the alternative he keeps pointing at is simple to state and hard to do: aim AI at the things we are not doing at all, instead of automating the things we already do. So here's a challenge to the executives reading this. The instinct in this industry is to let somebody else go first, wait for the outcomes data, then move in year 3 with a fully de-risked business case and a vendor who has been through 12 deployments. That instinct was correct for 30 years. It is wrong now, for exactly one reason: your patients and your employers are not waiting for your governance committee. They, quite literally, cannot afford to wait. The walls will crumble around you. Look at what changes when the labor constraint stops binding. Panel sizes move from 2,000 toward 5,000. Follow-up cadence stops being a function of open appointment slots. Preventive economics flip when outreach approaches zero marginal cost. Geographies you couldn't serve become servable. Carmouche thinks the in-office primary care visit becomes the exception, with the remaining 15% or so of encounters that need hands on a patient met by a paramedic in a van with a virtual NP on the line, or a 500-square-foot micro-clinic distributed across a county. Specialists keep their clinics. Primary care leaves the building. Access stops being measured in exam rooms and starts being measured in responsiveness and continuity - whole-house math. That's a fundamentally different business than the one most systems are budgeting for right now, and the gap between the two is going to be the most consequential strategic split in care delivery this decade. What I can't tell you is which payment mechanism wins, or what the 5th enterprise deployment of something like Tom looks like, or whether a state legislature blows the whole thing up in a fit of caution. Which makes me wonder how many organizations are going to spend the next 18 months waiting for those answers, and then find out the answer in their own market was settled by whoever moved first. These are my rambling thoughts, written in partnership with a company I've now covered closely for a year and a half, so weigh it accordingly. If you think there’s more to this conversation, especially on the payment side, hit me with it. Or better yet, get in touch with the Tom team and tell them what to do about it. Let’s get after it, fam. 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

@Blake Madden Unsubscribe