What do AI healthcare startups actually sell?

Most AI in healthcare startups that get attention sell clinical work: imaging, diagnosis support, drug discovery. The ones landing enterprise contracts inside large health systems increasingly sell operations instead, which is the administrative machinery around care rather than care itself. Referral intake, prior authorization, scheduling, billing, records transcription. Luminai is the current example: its Cleveland Clinic partnership starts with processing external referrals, many of which still arrive by fax, and Luminai's co-founder says most Cleveland Clinic referrals come in that way. Operational work has a measurable before and after, an existing budget line, and a buyer who does not need a clinician's sign-off to try you.

Why do hospitals still use fax machines?

Hospitals still run on fax because it is the one interface every provider in the country can already reach, regardless of which electronic health record they run. It is a lowest common denominator that survives because no shared alternative was ever adopted across competing systems, and because it carries a settled legal and privacy status that newer channels had to argue for. That is bad for patients and useful for founders: it means the integration surface into a hundred-year-old institution is an inbox of unstructured documents rather than a locked-down API, and turning those documents into structured, routed data is work a startup can do without asking permission to touch the system of record.

How much of US healthcare spending is administrative waste?

US administrative waste is smaller than the pitch decks claim and still enormous, at roughly 400 billion to 800 billion dollars a year. Founders routinely cite thirty percent of spending as administrative waste, over a trillion dollars, but the published research separates two things. Administrative spending is roughly fifteen to thirty percent of national health expenditure, and at least half of that administrative spending is wasteful, so roughly seven and a half to fifteen percent as a floor. Against 2024 US national health expenditure of 5.3 trillion dollars that is a range starting around 400 billion to 800 billion dollars. Large enough to build on, small enough that quoting the top of the range to a hospital CFO who knows the literature will cost you credibility.

Should a first-time founder start an AI startup in healthcare?

A first-time founder can start an AI startup in healthcare, but only with eyes open about the entry cost. Pylon's co-founders, on the same YC Root Access series, explain why they chose a horizontal product instead: going into a vertical like healthcare means conferences, relationship building, and a long wait before you get a foot in the door, while a horizontal product lets you start selling in whatever market you already have connections in. Luminai's route around that tax is worth copying, because it did not choose healthcare at the start. It sold horizontally, then found that north of eighty percent of its customers were already in healthcare, and narrowed onto evidence rather than onto a thesis.

AI in Healthcare Startups: Start at the Fax

The standard pitch for AI in healthcare startups is the part of medicine that looks like medicine. Diagnosis support, imaging, drug discovery, a model that reads a scan better than a tired radiologist at 2am. It is the demo that gets the room, and it is also the demo that has to survive clinical validation, regulatory review, liability questions, and a buyer who cannot approve it alone.

Meanwhile the company now partnered with a health system that recorded 15.9 million patient encounters last year is sorting faxes.

Luminai is an AI-native automation platform for health system operations. On April 9, 2026 it raised a 38 million dollar Series B led by Peak XV Partners, bringing it to about 60 million dollars raised in total. The Cleveland Clinic partnership went public the same day, and on September 9, 2026 Cleveland Clinic published its own announcement of the work, naming fax referral transcription as the first focus area.

On YC Root Access, co-founder and CEO Kesava Kirupa Dinakaran walks through how a general-purpose automation company ended up there. The funding is the least interesting part. What should hold a founder's attention is that almost every decision that made Luminai work reads as a downgrade in ambition at the moment it was made.

Why AI in healthcare startups should start at the fax

Dinakaran opens with Cleveland Clinic, which reported 15.9 million patient encounters in 2025 across its system. Patients come from everywhere to be treated there. And then:

This is the unfortunate reality about healthcare in America, but the way that most of those patients get referred into the Cleveland Clinic is through a fax.

A physician somewhere writes a note about a patient and sends it to a fax line. On the other end sit operational teams whose job is to read what arrives:

Their entire job is basically to look at these faxes, figure out is this sales spam that's coming to Cleveland? Is this like a thank you note from some random provider? Or is this a high critical cancer patient who needs immediate attention today?

Read that as a founder rather than as a patient. Every property you want in a first enterprise wedge is in that paragraph. The volume is high and growing. The work is unambiguously manual. The cost of a mistake is severe and legible to the buyer, which means somebody senior already worries about it. And critically, the task has a correct answer that a human can check in seconds, so you can prove your accuracy without a clinical trial.

The fax is not a quirk to be embarrassed about. The usual explanation for why it persists is that it is the one channel every provider in the country can reach no matter which electronic health record they run, and that its legal and privacy status is long settled where newer channels had to argue for theirs. For a startup that is close to ideal. The integration surface into a hundred-year-old institution turns out to be an inbox of unstructured documents rather than a vendor-controlled API you need a partnership agreement to touch.

Luminai's description of the product follows directly:

We're converting all of this unstructured data like faxes into structured data. And then we have a workflow engine on top where you can essentially build a set of verticalized agents to go solve very specialized and very important problems.

That is the two-layer shape a lot of the working enterprise AI companies share now: a transformation layer that turns the mess into records, and narrow agents on top that act on those records. The hard half is not the one founders assume. The agents are the part everyone can build. The transformation layer, tuned to one institution's documents and one institution's exceptions, is the part that takes months of embedded work. We wrote about the general version of that pattern in vertical AI agents.

The number in every healthcare deck, and what the research actually says

Dinakaran states the market the way most healthcare founders state it:

With the 30% of spend that's on the administrative waste within health care, which is over a trillion dollars, we can get rid of that.

That is the single most repeated number in healthcare startup pitches, and it does not survive contact with the literature in the form it is usually said.

US national health expenditure reached 5.3 trillion dollars in 2024, according to the Centers for Medicare and Medicaid Services. A Health Affairs review of the research puts administrative spending at roughly fifteen to thirty percent of national health spending, and wasteful administrative spending at at least half of that, so roughly seven and a half to fifteen percent as a floor. The 2019 JAMA review by Shrank and colleagues put administrative complexity specifically at 265.6 billion dollars a year inside a total waste estimate of 760 billion to 935 billion.

So the honest version is something like 400 billion to 800 billion dollars of administrative waste, rather than thirty percent of spending and over a trillion dollars. The thirty percent figure is close to the ceiling of estimates for administrative spending overall, which is a different quantity: not all administration is waste, and someone has to do the necessary part.

That is not a gotcha. It is a working rule. The gap between administrative spending and administrative waste is exactly the kind of distinction a hospital CFO knows cold, because it is their own budget. Quoting the top of the range at them prices your credibility down in the first meeting, and in a sector where the whole sale runs on being trusted with patient data, credibility is the asset you are actually accumulating. Take the smaller, defensible number. It is still one of the largest addressable pools of manual work in the economy.

How a horizontal platform found out it was a healthcare company

Luminai did not start in healthcare. Dinakaran describes the earlier company as "a much more generalized and horizontal automation platform" that had reached dozens of customers and many millions in revenue. The problem showed up in sales conversations rather than in the product:

The moment you start to say I can do anything or I work with someone else who doesn't look like you at all, the credibility you have reduces pretty dramatically.

This is the specific failure mode of horizontal positioning at enterprise scale, and it is counterintuitive because breadth reads as strength everywhere else. To a hundred-year-old institution deciding whether to route its referral intake through your software, a claim that you can automate anything translates as a confession that you have not done this before. The logos on your slide that come from unrelated industries are not proof of range, they are proof of dilution.

What made the decision tractable was that Luminai did not have to guess. It read its own customer list:

North of 80% of our customers were in healthcare.

And then made the call:

We can't try to be everything for everybody. Let's focus in and say let's go all in on healthcare.

The lesson in that sequence is not the vertical itself. Luminai discovered its vertical instead of picking one: a general product sold to whoever would buy, and the market voted with contracts before anyone wrote a thesis. Narrowing onto that evidence is a very different act from narrowing onto a conviction, and it carries far less risk, because the revenue that justifies the focus already exists.

The second narrowing was harder. Being horizontal within healthcare turned out to be nearly the same problem one level down:

These institutions are large and they don't have time to think about what they can do creatively with your platform. You have to almost tell the story of what's possible very very clearly and you have to understand the problem so deeply that you're able to talk to their problem like you've known them for 20 years.

That last clause is the bar. Not that you integrate with the incumbent record system. Not that you are HIPAA compliant. You have to describe the buyer's own workflow, including the parts that annoy them, better than they would describe it themselves. Everything Luminai does around wedges follows from it:

Have a very clear set of initial starting use cases and wedges that makes them realize value incredibly quickly.

Note that the platform did not go away. The architecture stayed extendable. What changed was the first thing you show someone, which is a solved instance of a problem they have, not a canvas.

How AI in healthcare startups actually deploy

The Cleveland Clinic release contains a detail worth more than the headline. Luminai's forward-deployed engineering team works embedded alongside the hospital's clinical and operational staff to adapt the workflows, rather than rolling out a standard configuration.

That is expensive, and it is the correct expense in this market. It is also the model showing up wherever AI companies sell into complex institutions, which we covered in forward deployed engineers. The reason it fits healthcare specifically is that the exceptions are the work. Every health system routes referrals slightly differently, names its departments differently, and has a set of rules that live in the heads of people who have worked there for twenty years. A configuration wizard cannot collect that. An engineer sitting next to the intake team can.

The release also describes the compounding effect that makes this economical: over time the platform incorporates organization-specific context, so subsequent use cases deploy faster. That is the defensibility. Not the model, which everyone rents from the same three labs. The accumulated, institution-specific context, which a competitor would have to earn the same slow way. Luminai's Series B release puts the current average time to value at 48 days, which is the number to beat rather than a number to assume.

There is a security dimension here that founders underrate, and it is the reason the embedded model is not just a sales tactic. You are handling protected health information from the first pilot. That means the confidence thresholds that decide when a document routes to a human rather than through the agent, the audit trail that records what the system did and on whose authority, and the access boundaries around the record are not later hardening work. They are the product, in the same way brakes are part of a car. Forbes reports that the system routes work to human staff when confidence levels fall below a threshold, which is the shape to copy: automate the volume, escalate the ambiguity, and log both.

The honest counterargument: why some founders walk past healthcare

The other side of this is well argued by a company on the same interview series.

Pylon, a B2B customer support platform founded by Marty Kausas, Robert Eng and Advith Chelikani, deliberately chose horizontal over vertical. On YC Root Access, talking through how they picked the space, one of the co-founders names healthcare as the example of what they were avoiding:

If you go into healthcare, you're going to have to go to conferences, you're going to have to build relationships, it's going to take a long time to even really get your foot in the door.

A horizontal product, by contrast, lets you start in whatever market you already have a connection in. For Pylon that meant starting in tech, an industry of early adopters they already knew, with an unfair advantage: they hit up about twenty founder friends for validation and several were already in market for a solution.

Both companies are right, and the disagreement is narrower than it looks. It is about when you pay the entry cost, not whether it exists. Pylon's argument is that a vertical charges you the relationship tax before you have a product, when you can least afford it. Luminai's history is the accidental answer: it also refused to pay that tax up front. It sold to whoever would buy, accumulated the relationships and the domain knowledge as a side effect of revenue, and only narrowed once the customer list had already made healthcare the obvious choice.

Where they genuinely disagree is on what happens next. Pylon's bet is that a horizontal product compounds through breadth, because every new industry is an expansion of the same product. The bet implied by Luminai's path, though Dinakaran does not put it this way himself, is that the tax you pay to enter healthcare is the same tax the next entrant has to pay, which converts it from a cost into a barrier. That argument about whether regulated complexity defends a company is worth taking seriously on its own terms, and we work through it in competitive moats in the AI era.

If you have no relationships in healthcare and no revenue to fund the wait, Pylon's read is the safer one. If you already have healthcare customers you acquired for other reasons, Luminai's is close to free.

Getting in the door

Dinakaran was closing large institutional customers at around twenty years old with no network. His method was not cold outbound:

Really major institutions don't really talk to you unless it's someone who they trust is like making the referral.

So the team wrote a script to scrape the contacts of everyone in their orbit, matched those contacts against a target customer list, and came back with two or three names per person, then asked each investor or advisor whether they actually knew any of them. It did not matter whether the person they reached was an executive. Anyone inside the company was enough to start, because the goal was to be passed two layers up toward the person who would eventually champion the deal.

The champion mechanic itself, selling to an individual inside an institution rather than to the institution, is the same motion that runs public sector deals, and we covered it in AI for government. The healthcare-specific twist is that the referral is what buys you the meeting. In a market where every vendor claims clinical-grade accuracy, a warm introduction is a trust signal that a landing page cannot manufacture.

What to do this week

Find your fax. In whatever regulated industry you are looking at, find the highest-volume task still done by a person reading an unstructured document. That is your wedge, not the impressive model demo. Write down the daily volume and the cost of getting one wrong.
Read your own customer list before you pick a vertical. Sort existing revenue by industry. If one segment is already north of half, the decision has been made for you by people who paid. If nothing dominates, you do not have the evidence to narrow yet.
Rebuild your market number from primary sources. Take the headline stat in your deck, find the underlying study, and check whether it measures spending or waste. Present the defensible range. In a trust-based sale, being the vendor who quotes conservatively is a positioning advantage.
Write the buyer's workflow in their words. One page, describing their process including the exceptions, with no mention of your product. Show it to someone who does that job. If they do not correct you, you are ready to sell.
Design the escalation path before the automation. Decide the confidence threshold at which a task goes to a human, what gets logged, and who can see the record. In any market handling sensitive data, that design is the product, and retrofitting it after a pilot is how deals die in security review.
List your target accounts and work the graph. Name the institutions, then ask your investors, advisors and customers which of their contacts sit inside any of them. Two names per person is a good return. Anyone inside is enough to start.

Choosing which work to automate first, then building the operating discipline around it, is the actual job, and it is what I teach in AI Operating System for Startups. The wider map of how these decisions connect sits in our pillar on AI for startups.

Sources

This Startup Is Automating America's Biggest Hospitals, YC Root Access, the interview this article distills, with Luminai co-founder and CEO Kesava Kirupa Dinakaran. Luminai's own releases style him "Founder and CEO"; TechCrunch's 2022 coverage of the company and his LinkedIn both record a co-founder, so the co-founder form is used here.
Cleveland Clinic Partners with Luminai to Transform Health System Operations, Cleveland Clinic newsroom, September 9, 2026, for the partnership, the fax referral use case, and the forward-deployed engineering model. The confidence-threshold routing to human staff, and the April 9, 2026 date on which the partnership first went public, come from the Forbes coverage below.
Luminai Raises 38 Million Series B for the round, the total funding and the 48-day average time to value, with additional coverage in Forbes and MedCity News. Company profiles: Luminai and its Y Combinator page.
Cleveland Clinic Facts and Figures for the 15.9 million patient encounters recorded in 2025. Dinakaran says "a little over 16 million" in the interview; the system's own published figure is used here.
On administrative cost: the CMS National Health Expenditure fact sheet for 2024 spending, Health Affairs on the role of administrative waste for the fifteen to thirty percent range, and Shrank, Rogstad and Parekh in JAMA for the administrative complexity estimate.
Rebuilding Customer Support for the AI Era, YC Root Access, for the horizontal counterargument from the Pylon co-founders. Profiles: Marty Kausas, Robert Eng and Advith Chelikani, and the company at Pylon.