In my last post, I chronicled the rise of Nurse Practitioners, a medical profession whose fifty-year journey to widespread utilization offers a glimpse into the formidable regulatory and legal hurdles that agentic clinical systems will have to navigate to reach patients at scale. For those that missed it, the TL;DR for NPs was a state-by-state slog that continues to this day; clinical AI has the FDA to contend with, too.
But, there are reasons to be sanguine: in August, the FDA’s Center for Devices and Radiological Health (CDRH), put out a discussion paper with a proposed framework for how they could approve non-deterministic clinical systems by assessing “competency” using a similar approach to how human doctors are vetted. This month, ARPA-H announced the three main participants in ADVOCATE, a program that provides significant funding to a slew of AI innovators building gen AI systems to manage cardiovascular disease and incentivizes hospitals to support them as data partners and clinical trial sites.
With regulatory winds blowing favorably for the moment, we may see elements of “spontaneous deregulation.” What if the states with the largest patient loads lean into AI-delivered care (of their own accord or incentivized by federal policy)? What if the FDA successfully reinvents its process for autonomous clinical AI with expedited clearance pathways?
Many founders are already building for a world where clinical AI overcomes regulatory barriers in the not-too-distant future, and the early contours of care delivery’s future business models are starting to take shape. What these startups tend to have in common:
- Their business model bets on a shift from a variable cost-heavy clinical staffing model to a more software-like operational profile, with higher technology (and perhaps regulatory) costs, free from some of the supply constraints of clinical labor.
- Their clinical model revolves around the ability of AI systems to learn from their own “real world evidence” (interactions with thousands of patients and their clinical outcomes, with purposeful A/B tests of different interventions).
- Their operational model leans on AI for its power as a patient-facing interlocutor, helping to reduce the length of a subsequent human encounter (when needed) to a matter of minutes or seconds through intelligent triage.
Many startups are building with the assumption that AI models will continue to improve. But these startups are orienting their care model and clinical engagement so that, as regulatory barriers fall (in particular, the authority to prescribe), their AI agents are ready to take on much more of the workflow.
With this future in mind, here are nine of the categories we’re seeing emerge:
Hospital-focused models
1. The agentic front door for hospitals
Primary care is the entry point to care for most patients, yet it is compensated poorly. For hospitals, it plays an essential role in navigating patients to surgeons and specialists within the system. Scarce as specialist appointment slots are, primary care availability is also tight (ever try booking your annual physical less than twelve months in advance?). This is where AI-native care delivery companies can step in to act as a triage layer for patients gated by availability and at risk of “leaking” to another provider. The pitch to hospitals: use AI-enabled triage (and some day AI-delivered care) to engage with every patient who attempts to book a visit, providing AI-forward urgent care and on-demand, lightweight primary care, with the goal of identifying a health issue that requires specialist attention, resolving an issue that never required a specialist in the first place, or keeping the patient engaged in their health until a more suitable appointment slot opens up. As a bonus, it can help to turn primary care itself profitable.
2. Disease management sidekicks
Hospitals are also invested in closely managing patients with complex conditions. For one thing, under state and federal quality programs, they can be penalized for ER readmissions and other adverse clinical events. For another, complex patients with unmanaged needs often tie up facility beds which affects other patients and poses financial risk when reimbursement is bundled by episode. This is exactly the category of challenges that remote patient monitoring and hospital-at-home programs have been targeting for years. The AI-native version of this has the potential to reach more of these patients and at an even lower cost. This could take the form of an agentic sidekick that acts at the behest of a hospital-based clinician with the ability to autonomously initiate and renew medications within a predetermined scope, escalating back to a human if symptoms become more complex.
Value-based care models
3. VBC-in-a-box
Value-based care has had many successes, but the number of providers in true risk arrangements is still a fraction of what it could be. Among the major reasons why: most providers don’t have the financial or operational infrastructure to provide ongoing support to patients outside of billable visits. But what if instead of staffing a clinical team to outreach patients and check in on their health, an army of agents could be deployed? Providers could dispense with the usual process of risk stratification and instead engage all patients regardless of acuity. This is one promise of clinical AI that has sparked a lot of recent innovation giving rise to AI-native MSOs as well as patient-facing agentic platforms that can be activated without disrupting a practice’s operations. Just as “business-in-a-box” is assuming a new AI-native form, “VBC-in-a-box" has become much more viable thanks to AI.
4. AI-native risk managers
Augmenting providers with AI “sidekicks” (see model #2) can make care much more accessible and is one of the most important ways the technology will permeate healthcare. But the only way to capture the full value of the outcomes is to own the patient or member relationship and take responsibility for their medical costs. That’s the focus of a number of companies building today, from AI-native health plans that offer integrated care to their members in order to reduce costs, to specialty groups that take on risk by pairing AI-enabled intake with in-person touch points and agentic follow-ups to surround patients at every step.
Employer-focused models
5. EAP 2.0
Part of the allure of therapy is the human connection itself, and no AI system, at least at this point, can replicate this effect. Which is why in the short-term, AI therapists are likely to have the most traction in low acuity, episodic settings, one of the clearest cut of which is the EAP (Employee Assistance Program). Virtually all mid-to-large employers offer an EAP benefit, a hotline or tool for employees in need of short-term counseling for personal and workplace-related issues. Usage of these programs remains very low (~10% or less on average according to many sources). An AI-native EAP has the potential to be more personalized (though not as personal), with the ability to scale to more employees at a lower cost (clinical agents aside, voice AI and AI-fueled marketing tactics can also be used to boost employee activation).
6. DPC 2.0
For years, the closest you could come to “all-you-can-eat” care was a subscription to a direct primary care practice (DPC), which offers patients unlimited access to a doctor for a fixed monthly fee. DPC has picked up steam among employers in recent years and this is likely to accelerate: OBBBA lets employers pay for DPC membership with pre-tax dollars just as they do health insurance. Meanwhile, there is a sizable body of evidence that DPC actually reduces costs by heading off more expensive ER and specialist visits. However, DPC panels tend to be small (~400-800 patients instead of the usual ~2,000-2500), which means it may never become a model that fully scales. A growing number of founders are setting their sights on subscription-based primary care models powered by clinical agents, with the goal of bringing concierge medicine to employers of all sizes and budgets. They’ll often integrate lab panels to broaden the clinical portrait that agents can help interpret and take action on. Many are pursuing this idea as a D2C service, too.
D2C models
7. Programmable clinics
Patients are increasingly bypassing traditional healthcare entry points and discovering new ones. This is driven by three trends upending the playbook of direct-to-consumer patient acquisition:
- Internet distribution is shifting from search engines to AI apps (and the AI apps are leaning more and more into personal health).
- The popularity of GLP-1s is colliding with tight payor coverage, leading to an explosion in cash-pay volume and engagement with the D2C companies ready to catch this demand.
- A fascination with longevity has catapulted bio-hacking into the mainstream, with patients tracking biomarkers (via lab panels or wearables) proactively.
AI-native startups are responding to this by building the infrastructure to allow almost any consumer company to spin up clinical offerings and treat its users as patients. This could take the form of a clinical agent cleared to write routine prescriptions (from antibiotics to GLP-1s) standing in for a human prescriber; it could also be a clinical service whose patient acquisition is driven largely by referrals from (or integrations with) AI apps. Previous telemedicine staffing models relied on armies of 1099s and had to walk the tightrope of a race-to-the-bottom on volume-based fees and rising labor costs. AI-enabled care models on the other hand can scale more flexibly, with potentially fewer of the headaches associated with state-by-state licensure, that is if the FDA wins out as clinical AI’s primary regulator (a big “if”).
Infrastructure plays
8. EHRs as Agent Orchestrators
Nominally, an EHR is a clinical database. But in practice it’s the OS of every medical practice, home to some of the most intricate and high-stakes workflows in all of software, engendering disdain (often) and diehard loyalty (sometimes) from its users, the doctors, nurses, and administrative staff who make up our healthcare system.
Just as CRMs are increasingly going “headless,” it’s likely we’ll see EHRs chart a similar path. As AI-forward models like the “disease management sidekick” take root, the agents seeing patients will need a place to document their work and to fetch new instructions from their human dispatchers. While today many startups go about giving EHR access to agents through RPA, APIs, or computer use models that mimic the point and click interactions that humans are accustomed to, jerry-rigging user access, the end game will inevitably be a set of clinical platforms tailor-made for human-AI collaboration.
9. Clinical AI evaluation platforms
Before a doctor can see patients, they go through a multi-year training process that starts with pre-med, then med school, then board exams, residency, fellowship, and on-the-job training. Once a clinician has demonstrated they’re competent to practice medicine, barring any extreme mistakes, they’re allowed to continue doing so indefinitely, even as the practice of medicine evolves, as long as they keep paying their license registration fees.
When doctors encounter cases that were never covered in training, the assumption is they can extrapolate from what they’ve learned to solve these new cases. Increasingly, the FDA is coalescing around a vision of clinical AI approval that aligns with this, in what they call a “competency-based” approach. As they put it in their recent discussion paper: “CDRH recognizes that evaluation approaches developed for software with bounded inputs and fixed outputs may not be appropriate for GenAI-enabled devices. Such evaluation has traditionally relied on extensive testing across a representative sample of device inputs and outputs; for GenAI-enabled devices, the range of possible inputs and outputs may be too large for such testing to be practical.”
This begs the question: who or what will facilitate this competency assessment? There are a few possibilities:
- While the FDA appears to be favoring retroactive evaluation, they acknowledge that in some cases prospective evaluation (for example, a randomized clinical trial) will be necessary to fully establish efficacy and safety. In that case, a new crop of CROs (Contract Research Organizations) could emerge, specialized in patient recruitment through digital channels for online care.
- Healthcare-focused AI vendor governance platforms are likely to play a role here as they can benchmark the performance of clinical AI models across customers and audit models for drift after launch.
- An increasingly relevant set of startups are building reinforcement-learning environments for healthcare use cases. Today, they create sandboxed applications and datasets that the AI labs use to train long-horizon agents on workflows ranging from patient care to revenue cycle management, giving them access to replicas of the types of tools (EHRs, payor portals) that doctors and their staff use. In the future, they’re well positioned to partner with agentic care delivery companies both to help them improve their capabilities and to help them demonstrate competency according to regulatory standards.
Conclusion
We’re entering a highly disruptive new chapter for tech-enabled care delivery in which even the most fundamental responsibilities of a doctor have the potential to be peeled away by automation. If this process plays out quickly, it will transform the healthcare system in ways we can’t even begin to imagine. These nine models bend the boundaries of clinical AI and are a safe bet for where we’ll see disruption happen first. If you’re building in any of these categories or dreaming up entirely new ones, let’s talk.
