What I'm Betting On in AI and Care Management

We have spent the last decade in RPM, CCM, and chronic disease care: hypertension, diabetes, CHF, behavioral health. The conditions that represent the bulk of U.S. healthcare spend and have been stubbornly resistant to every cost-curve promise made about them.

For most of that decade, the innovation was about the care model itself: who delivers the care, how often, through what channel, at what cost. The AI layer changes the questions. Not because it replaces the model, but because it changes what is possible inside it.

I want to be upfront about where I'm coming from. Most of my career has been on the operations side of building businesses: workflows, human systems, how organizations actually function at the task level versus how they look on paper. The last few years have been a deep dive into care management and chronic disease specifically. Over the last year I've been going pretty deep on the AI stack itself and watching what consumer adoption actually looks like at the product level — that combination is the lens I keep coming back to when I look at this space. This post is partly me working through it out loud, and partly an invitation to the people building here.

Where the ROI actually pencils today

There are four areas where I've seen AI generate verifiable, repeatable returns in chronic disease care. Not "promising pilots." Actual math.

Readmission prevention in high-acuity cohorts. This is the clearest case. AI models that draw on EHR data, claims, and social determinants can identify patients at meaningful readmission risk earlier, and with substantially higher precision, than traditional scoring models. A few health systems have reported 30-day readmission rate reductions in the 40–75% range when AI-flagged patients are paired with active care team interventions. For a 300,000-member Medicare Advantage plan, that math translates to somewhere in the range of $4–6M in annual savings for a mid-six-figure technology investment. The ROI timeline is 12–24 months. That is a real business case.

Ambient documentation. This one is not about chronic disease specifically, but it is reshaping what care teams can do with the same headcount. AI clinical documentation tools are returning 2–3 hours per day per clinician. That is not marginal. At the care manager level, that is the difference between a caseload of 80 patients and a caseload of 130. Same team, same operational cost, 60% more covered lives. For a care management company trying to hold margins in a competitive PMPM environment, this is one of the few cost-structure levers that actually works.

Prior authorization automation. The drag that PA creates on care management workflows is real and underappreciated. Platforms like Cohere Health are now auto-approving upward of 90% of prior authorization requests for the populations they cover, processing tens of millions of requests annually. The ROI for health plans is obvious. But the second-order effect for care managers is just as significant. Removing authorization delays from the care pathway is the operational equivalent of hiring additional coordinators, except the math compounds at scale rather than grinding against fixed labor costs.

Risk stratification, but only when paired with an intervention. I want to be specific here because this is where a lot of AI deployments stall. AI can now identify the highest-cost, highest-risk members with materially more precision than traditional CDPS or HCC models, finding 40% more truly high-risk individuals in some analyses. But identifying those patients saves exactly zero dollars on its own. The ROI only appears when stratification is paired with a disciplined outreach and care management protocol. The companies that are getting this right have built the AI layer and the human intervention workflow as a single system, not two separate products talking to each other over an API. That integration is where the actual value sits, and it is harder to build than either piece in isolation.

Where the bodies are buried

A few areas where I am more skeptical than the market seems to be:

Broad care coordination platforms without clear workflow ownership. There are a number of platforms selling AI-powered care coordination to health systems and payers that are genuinely impressive in demos and genuinely difficult to capture ROI from in practice. The issue is not the technology. It is that care coordination involves a lot of people doing a lot of different things, and AI that sits adjacent to the workflow rather than inside it creates another alert to manage, not a problem that goes away. The implementations I find most credible are the ones where the platform owns the workflow, not just the intelligence layer on top of someone else's workflow.

Risk stratification sold as a standalone product. A score is not a service. If a company's core offer is "we will tell you which patients are high risk," the health system or payer on the other end still has to figure out what to do with that information. That is a consulting engagement, not a scalable business. The durable companies in this space are the ones that own the intervention, not just the prediction.

AI replacing the care relationship entirely. This one comes up in board decks more than it should. There is a reasonable version of AI-automated patient outreach: post-discharge check-ins, medication reminders, appointment confirmations, triage protocols. It genuinely works for a defined set of touchpoints. Memora-style automation is real and valuable. But the idea that AI can replace a care manager in a complex chronic disease relationship, particularly with a high-risk patient, is not supported by what the outcomes data actually shows. The staffing math is appealing. The clinical reality is more complicated.

My actual bets on the next few years

The care team augmentation story wins. The replacement story loses. I expect the companies that position AI as a force multiplier for clinical labor to significantly outperform the ones that lead with headcount reduction. Partly because the outcomes evidence supports augmentation more than replacement. Partly because the clinical trust problem of getting care teams to actually use AI-generated recommendations in real-time decisions is much easier to solve when the AI is saving nurses time than when it is competing with them for authority over patient decisions. The liability picture also favors augmentation, and that matters more than most tech companies building in this space seem to appreciate.

Risk-bearing entities will be the power buyers. The ROI math for AI in care management is structurally better for organizations that own the risk than for those on fee-for-service. If you save a readmission in a FFS arrangement, you lose the revenue. If you save it in a capitated or value-based contract, you keep the margin. This is not a novel observation, but I think the implication for which buyers AI care management companies should prioritize is more significant than most GTM strategies reflect. ACOs, Medicare Advantage plans, and risk-bearing provider groups should be the primary motion, not the secondary one.

Data infrastructure will separate the real platforms from the point solutions. The care management companies that build durable AI advantage are going to be the ones with the cleanest, most longitudinal patient data: EHR integrations that actually work, device feeds that are actually clean, SDOH signals that are actually current. That data layer is slow and expensive to build and almost impossible to replicate quickly. It does not show up in a pitch deck as a headline feature. But it is the actual moat, the same way clinical workflow rigor was the moat in RPM before AI entered the conversation. Companies that are investing in data quality and integration depth right now are building something that will matter a lot in 36 months and that nobody is going to hand them later.

The winners in this space will look more like Cohere and less like most of the ambient AI field. What I mean by that: Cohere's model is deeply embedded in a high-friction, high-stakes workflow (prior authorization), owns the outcome (approval or denial), and operates at a volume where the economics are unambiguous. That is a different animal than a platform that provides AI insights to people who then go do clinical work somewhere else. I am more bullish on companies that own a defined slice of the care workflow end-to-end than on companies that are intelligence layers hoping someone else executes against their outputs.

What I want to learn in the next few months

I'm actively trying to get sharper on a few things, and I am genuinely looking to learn from founders and operators building in this space:

• How are the companies getting real workflow integration (not just API connections) actually doing it? What does the change management look like inside a health system when an AI tool goes from pilot to production?
• Where is the liability conversation landing in practice? I hear a lot of confident positioning from founders and very little specificity about how the AI governance and accountability frameworks actually hold up under pressure.
• What does the commercial motion look like for AI care management companies selling to risk-bearing entities versus health systems? The buyer, the budget owner, and the value narrative are all different. I have more conviction about the thesis than I do about the playbook.
• Which SDOH data sources are actually useful in practice versus useful in a model? I have a healthy skepticism about the gap between "we incorporate SDOH signals" and "our SDOH data actually changes what we do for the patient."

If you are building in this space and any of this resonated, or if you think I have something wrong, I would genuinely like to hear from you. Not as a sales conversation. As someone trying to develop real conviction in a category I find important and think is still being underestimated by most of the health tech market.

— Andrew