Subscribe | Advertise | PRO Member Hub

TOGETHER WITH LEGION HEALTH

Perspectives from the people building the future of health AI…

We sat down with Arthur MacWaters, President and Co-Founder of Legion Health, an AI-powered, full-stack psychiatry clinic. Legion’s providers work in concert with clinical and back-office AI to provide care directly to patients. It hires its own clinicians, contracts with payers, acquires patients directly, and uses LLM-powered tools they’ve built in-house throughout the stack to make care cheaper, more personalized, and easier to navigate.

Arthur shared why the team pivoted from B2B staffing to tech-enabled care delivery, how they are using LLMs across patient support, billing, and clinical workflows, and why they believe the right way to get to the future “AI doctor” care is to start as a real clinic and automate from there.

Let’s start from the top. What is the origin story behind Legion?
My co-founders and I were roommates at Princeton. We were close friends, but we did not leave school saying, "Let’s start a company together." I went to McKinsey, Danny became a technical PM at Microsoft, and Yash joined the Congressional Budget Office working on healthcare policy and economics.

A few years later, Yash pulled us back together with a simple argument. If we were ever going to build something, it should be in healthcare. All of us had seen people we love struggle to get care, especially for mental health, where access is terrible and prices are high. For me it was my younger brother. For Yash it was his dad. For Danny it was close friends. We quit our jobs, got into YC in the Summer 2021 batch, and started with a B2B model that helped health orgs load balance behavioral health providers.

When ChatGPT 3.5 arrived, it was obvious that AI would transform care itself. We wanted to own the full journey, not just be a staffing layer, so we pivoted into a vertically integrated telepsychiatry clinic. To support this approach, we brought on seasoned builders, including founding engineer Omar McAdam, a multi-time founder and 25-year healthcare and AI veteran known for building mission-critical clinical systems and products

Give us the quick elevator pitch. What is Legion today, and who do you serve?
Legion is a modern telepsychiatry clinic. We are full stack, which means we hire our own providers, hold our own insurance contracts, and acquire patients directly. So it is a D2C model, but in network for most patients.

Right now, we operate in Texas and focus on medication management. If you are a Texas patient with commercial insurance who needs psychiatric care, you can come to us for diagnosis, ongoing medication management, and follow-up. We’ll expand to more states quickly, but the core idea stays the same: high-quality psychiatry, powered by AI, but delivered by licensed clinicians.

You describe Legion as both a tech company and a care delivery company. How do you balance those two identities?
Most care delivery companies are not AI native. Their personalization is essentially limited to whatever the individual clinician can remember or document, or they hire large patient support teams to fill in the gaps.

We started from the opposite direction. We are a technology company first and a care delivery company second. Our internal rule is that anything a human can do, an LLM has to be able to read and eventually do as well. That drives how we architect every part of the business: from scheduling, to billing, to clinical support.

Because we own the stack, we can use everything we know about a patient. Their medications, their history, what they told us about their schedule, the exact language in past notes. That lets us personalize the care journey in a way that would be nearly impossible with manual workflows.

How do you think about the path to the “AI doctor,” and how will Legion build it?
A lot of companies are trying to build the AI doctor by starting with AI-only care. An agent joins a Zoom visit, handles everything, and hopefully gets good enough for patients and regulators to accept it. We think of that as the Waymo model: map everything up front, then try to run fully autonomous from day one. The problem is that healthcare has real constraints. AI cannot prescribe meds, insurers do not reimburse AI-only care, and regulatory change will move slower than many expect.

We chose the Tesla model instead. Start with a world-class clinic staffed by real providers, instrument everything, and automate the workflow step by step in real conditions. As AI systems learn from providers and our API-first stack, we can safely hand off scheduling, triage, follow up tasks, and documentation. Providers remain in the loop, but need to intervene less over time.

This approach lets us deliver high-quality care today while getting asymptotically closer to an AI-first experience, limited only by what regulators permit and what patients want. It is the practical path to the AI doctor: build real care, automate in the background, and scale until the human touch becomes minimal and supervision is all that remains.

How do LLMs and agents actually show up in the product today?
We think in four pillars.

First is AI patient support. Scheduling, messaging, and churn prevention are heavily automated. For example, every follow up visit is scheduled by an LLM that understands when meds will run out, when the provider is free, and what times the patient usually prefers.

Second is an AI CFO or RCM layer. We built tools that help us code and bill correctly every time. That reduces surprises for patients and has driven roughly a 12% margin uplift just from getting billing right.

Third is a clinical co pilot for providers. Before a visit, AI summarizes the chart, surfaces likely diagnoses, and suggests questions so that the provider can get to the right medication and dose faster. After the visit, agents help execute the plan, whether that is sending labs, routing an ROI, or pushing a prescription.

Fourth is direct patient personalization. After a medication change, for example, patients get tailored education on what to expect, what side effects to watch for, and how to adjust habits. Because we own the full stack, we can tie that to their history and prior visits.

What does the technical stack look like under the hood? Are you training your own models?
We are very API first. For the clinical record we use Healthie. For billing we use Candid. Everything else we build assumes that if a human can click a button in a UI, an LLM can read the same data and act on it through APIs.

Our stack is TypeScript and React, hosted on Vercel. On the AI side we use the main frontier models from OpenAI, Google, and Anthropic and switch between them as needed. Right now most of our performance comes from how we prompt and structure the workflows, not from custom model training.

We are spending far less on models than a traditional telehealth company spends on human operations. There is still always a clinician in the loop today. Over time we think human involvement in the routine parts of care will asymptotically go to zero, but within the constraints of what regulators and patients are comfortable with.

How do patients actually feel the impact of all this AI work?
We see three big differences. It shows up in cost, personalization, and reliability.

Most patients pay under $30 out of pocket, and cash pay is roughly half the price of a typical psychiatrist. Because providers come into visits fully briefed, patients feel more listened to and less rushed. And because we use LLMs to anticipate things like when a prescription will run out and when a patient is likely to show up, we can schedule and execute before problems arise.

Those little admin failures matter a lot in mental health. Avoiding a gap in meds can add up to dozens of days where someone is not left hanging because a fax did not go through or a callback was missed.

What metrics do you watch most closely, and what are you seeing so far?
At the business level we care about patients in care and revenue, but we also track some more operational and clinical metrics.

One key metric is the human cost of operations per patient per month. We literally have a dashboard that shows it trending toward zero as we scale. In the last 6 to 9 months we have roughly tripled our patient count while adding only one operations hire. That is the power of designing the company so LLMs can do most of the coordination.

We track patient retention and see an average of about 5.3 visits, which is significantly higher than the 2.9 to 3 visits that are typical in mental health. Some patients have stayed for 20 or more visits, which tells us that the combination of provider relationship and product experience is working.

On the outcomes side, we target that 70% of patients achieve a clinically meaningful reduction in PHQ-9 and GAD-7 scores, about 5 points, within 90 days. Today that is true for our moderate and high severity patients, which is encouraging.

From a growth perspective we have more than tripled run rate revenue in the past year, and are focused on scaling within Texas and then into states like California, Florida, New York, New Jersey, and beyond. Our revenue path is mostly about disciplined execution at this point.

You talk a lot about using the fee-for-service dollar as an innovation budget. How does that work?
Around 80% of our patients pay with insurance. If you give people the option between cash and "covered by your plan," they usually pick the latter. So we start from that fee for service payment and ask how much value we can deliver inside it.

Because AI lets us run operations with a very small human team, we do not need to squeeze every cent of margin. Instead, we use the room we create to offer "free" AI features to patients. Things like rich visit summaries or between visit AI tools are products that a pure consumer AI company would charge for. We bundle them into the same copay someone would pay at another in-network psychiatry practice.

It is a bit like buying a car for a normal price and getting advanced driver assist, parking, and security features included. Same ticket price, more capability.

What is your long-term vision for Legion and for AI in mental health?
Our internal goal is to be the last mental health company. By that I mean we want to build a model that is so good operationally and clinically that there is no obvious better way to deliver psychiatric care end to end.

In the near term that means continuing to be a full-stack psychiatry clinic that uses AI everywhere it can, while staying inside the existing regulatory and licensing rules. Over time, as rules evolve and as we collect more outcomes data, we will move closer and closer to AI-native care where the human provider focuses only on the parts that truly require judgment and connection.

More broadly, I think a lot of point solutions in healthcare AI will struggle. If you sell into health systems and depend on 12 to 18 month sales cycles and complex integrations, you are also competing with incumbents like Epic who can ship “good enough” AI features with instant distribution. Owning the full-stack care delivery lets us innovate faster and avoid that trap.

Long term, the companies that win will be the ones that can show, with data, that AI makes care better and safer. Our bet is that building a real clinic now is the best way to get there.

What stood out, what’s tricky, and why it matters…

Legion Health is building an AI-native, full-stack telepsychiatry clinic rather than a point solution or tooling layer. They hire their own clinicians, contract with insurers, and go direct to patients, then let LLMs quietly run much of the operational and personalization layer behind the scenes. The result is care that’s cheaper, more consistent, and more tailored, without removing humans from the core clinical loop.

Instead of chasing the “AI doctor” from day one, Legion’s bet is that you start with high-quality human care and progressively automate everything around it: scheduling, billing, documentation, triage, follow-up. Over time, the human share of the work shrinks toward zero where regulations allow, but patients still experience continuity with their chosen clinician and a smoother journey in between visits.

They’re also thinking smartly about the fee-for-service dollar. With ~80% of patients using insurance and most paying under $30 out of pocket, Legion uses AI-driven efficiency to pack more value into the same benefit, including between visit AI tools, rich summaries, and proactive outreach, without charging extra fees. It’s a Tesla-style strategy in psychiatry where the copay stays the same but the capability keeps increasing.

With a $6M seed round from investors like Y Combinator and Alumni Ventures, a super strong founding team, seeing thousands of patients today, and a run rate that has more than tripled over the past year, Legion’s real-world path toward the AI clinic and clinician puts them on a credible trajectory to emerge as one of the leaders in tech-enabled psychiatry.

What stood out

What’s tricky

Final thoughts

Legion Health is a compelling counterpoint to the wave of AI tooling and AI doctor concepts. By owning the full stack of psychiatry and using LLMs as behind-the-scenes infrastructure rather than a marketing layer, they are converting AI into lower costs, smoother operations, and more personalized care. If they continue to scale efficiently, show durable clinical improvement, and navigate the regulatory barriers ahead, Legion has a real chance to become one of the defining models and companies for AI-powered care delivery.