What business models are working in healthcare AI?

Why is healthcare AI suddenly becoming a real business?

Healthcare AI has crossed a line: physicians are already using it routinely, and several categories now have enough volume to support very large businesses.

The latest American Medical Association survey makes the change unusually clear. In 2023, 38% of physicians said their practices used at least one AI use case. That figure is now 72%, while 81% report at least some AI use or awareness in their practice. The average number of use cases has more than doubled from 1.1 to 2.3. Documentation, chart summaries, medical research summaries and care-plan generation are among the areas spreading fastest.

Commercial deployment is following the same curve. Abridge says its platform will support more than 100 million patient-clinician conversations this year across more than 300 health systems. Nabla is deployed in more than 130 healthcare organizations, with more than 85,000 clinicians and 20 million encounters a year. Microsoft says more than 100,000 clinicians now rely on Dragon Copilot in their daily practice.

Healthcare AI is no longer just a collection of hospital pilots. The harder question today is whether vendors can turn all that usage into good economics.

The first commercial winners point to a fairly simple answer. AI works best when it removes something healthcare already spends heavily on: physician time, administrative labor, missed reimbursement, diagnostic work, pharmaceutical research or access to doctors.

That is why documentation, payments, prior authorization and reimbursed diagnostics are moving faster commercially than many more ambitious attempts to automate medicine itself.

When can we say a healthcare AI business model is actually working?

A healthcare AI business model is working when customers repeatedly pay because the product creates measurable economic value, not because the technology looks impressive.

For this article, we need a stricter definition than adoption. A hospital can run a successful pilot and still cancel it. A pharmaceutical company can sign a large research partnership while the AI company continues burning hundreds of millions of dollars. A startup can process 100 million interactions without telling us what those interactions are worth.

The evidence becomes much stronger when several things happen together. Customers expand deployments. Usage keeps rising after the pilot. Revenue grows with usage or customer expansion. The buyer can point to money saved, money recovered or a reimbursable service. Margins improve as the system gets larger.

That changes how we rank the market. A reimbursed cardiovascular analysis growing nearly 50% with an 83% gross margin is much more commercially proven than an AI drug-discovery platform whose quarterly R&D spending is more than ten times its revenue, even if the latter has more technological upside.

It also keeps funding rounds out of the argument. Healthcare AI companies have raised billions lately, but venture capital tells us what investors expect. We want to know what customers are already paying for.

What we look for	Strong evidence	Weak evidence
Customer demand	Renewals, expansion, sustained usage	Pilots and announcements
Economic value	Revenue recovered, labor saved, reimbursed care	Generic efficiency claims
Monetization	Recurring, transactional or usage revenue	Funding and valuation
Scalability	Margins hold or improve as volume grows	More services work for every new customer
Durability	Workflow, data, reimbursement or distribution advantage	Slightly better model performance

Is ambient clinical AI the first big healthcare AI winner?

Ambient clinical AI is already one of healthcare AI's clearest winners because it gives clinicians back something extremely expensive: their time.

The scale is now hard to dismiss. Abridge says it is live across more than 300 health systems and expects to support more than 100 million patient-clinician conversations this year. Nabla reports more than 85,000 clinicians across 130-plus healthcare organizations. Microsoft says Dragon Copilot is used by more than 100,000 clinicians in daily practice.

Those are three separate platforms reaching tens of thousands of healthcare professionals, which tells us more than another isolated hospital announcement would.

The productivity gains can also be large enough to justify real budgets. Nabla reports that 55% of clinicians in one Carle Health deployment saved at least an hour of documentation time per day. At Codman Square Health Center, the organization attributed roughly $250,000 of additional first-year revenue to its deployment. Other systems have reported faster note completion, higher visit capacity and lower burnout.

Even Epic's own customer data point in the same direction. John Muir Health found that physicians using AI charting saved about 34 minutes a day on notes, while physician turnover fell 44%. UPMC reported close to two hours less daily "pajama time" for participating clinicians.

We should be careful about treating every customer case study as an independent clinical trial, but the pattern now appears across too many health systems and vendors to ignore.

Ambient AI solves a repetitive problem that happens during almost every clinical day. That repetition is why real budgets are showing up.

Can ambient AI stay valuable now that Epic and Microsoft are bundling it?

Standalone AI note-taking is becoming vulnerable fast, so ambient companies now need to own much more of the clinical workflow.

Epic released built-in AI Charting this year. The system listens during visits, drafts notes and queues orders from the conversation. Epic has already begun extending ambient charting beyond physicians to nurses. Microsoft is making the same move with Dragon Copilot, expanding from documentation into broader clinical assistance.

Hospitals already have deep contracts with Epic and Microsoft. A standalone vendor that does little more than produce a good note can eventually find itself competing against a feature included inside software the hospital already uses.

The leading ambient companies clearly see the danger. Abridge now talks about one clinical intelligence layer spanning preparation before the visit, documentation during the encounter, orders, clinical decision support and downstream work. Its latest clinical intelligence product can be made available to every clinician across a partner health system, including clinicians who do not use Abridge for documentation.

Nabla is also widening the product beyond ambient notes. Ambience has pushed into coding, CDI and revenue-cycle work.

Ambient AI remains a strong business. Basic note generation alone looks much less defensible than it did two years ago.

The companies with the best chance of keeping premium pricing will use the clinical conversation as an entry point into several valuable actions. Notes got them into the hospital. Orders, coding, clinical intelligence and financial workflows can keep them there.

If you want more recent data on this point, please see our latest healthcare AI market report.

Is revenue-cycle AI the strongest healthcare AI business model right now?

Revenue-cycle AI currently has some of the best economics we can find in healthcare because the customer can see exactly where the money comes from.

Waystar is a useful example even though it was already a large healthcare payments company before the current AI boom. That actually makes the evidence more interesting. AI is being inserted into a transaction network that was economically valuable long before generative AI arrived.

In its latest quarter, Waystar generated $319.7 million of revenue, up 18% year over year. Adjusted EBITDA reached $136.7 million, giving it a 43% margin. Subscription revenue grew 34%. The number of customers generating more than $100,000 of trailing annual revenue reached 1,453, up 15%.

Those are mature software economics while the company is still growing at a high-teens rate.

Waystar uses AI across areas such as denial prevention, payment workflows, charge capture, recoupment and work prioritization. When one of those systems prevents a claim from being denied or helps collect money sooner, the buyer does not need a philosophical discussion about the value of AI. The dollars show up in the revenue cycle.

Healthcare payments also produce huge amounts of proprietary transaction data. Every claim, denial, payment and appeal creates another example the system can learn from.

Revenue-cycle automation belongs near the top of the healthcare AI ranking. Some AI-native startups may grow faster, but very few models currently combine such clear customer ROI with recurring revenue, transaction volume and 40%-plus adjusted EBITDA margins.

Can prior-authorization AI become a big payer business?

Prior-authorization AI already looks like a serious payer business because one health-plan contract can put the software in front of millions of patients and hundreds of thousands of providers.

Cohere Health now supports more than 600,000 providers and processes more than 12 million prior-authorization requests annually. Its current platform data show that 85% of approvals can happen in real time in some specialties. Cohere also says 9 million authorizations used FHIR APIs during the last year.

The economics can be substantial because prior authorization still consumes large amounts of clinical and administrative labor. Cohere reports up to 47% administrative savings and 50% faster medical-necessity reviews. Those are vendor-reported figures, so we would not treat them like audited financial results, but the operational scale itself is difficult to fake.

Healthcare policy is also pushing prior authorization toward electronic, faster and more standardized workflows. That creates pressure on payers to upgrade systems that were historically built around portals, faxes, phone calls and manual review.

The latest product direction shows where the category is heading. Cohere has expanded from authorization into payment accuracy, appeals, quality and care management. Once the system already understands the patient's record, the insurer's policy and the authorization decision, that same intelligence can be reused elsewhere.

A pure "AI that approves requests" product would be narrow. A clinical decision infrastructure layer for health plans has a much bigger path.

Does reimbursed diagnostic AI beat hospital SaaS?

Reimbursed diagnostic AI is one of the strongest healthcare AI models today because revenue rises every time a doctor orders another analysis.

HeartFlow gives us unusually clean evidence. In its latest quarter, revenue reached $64.1 million, up 48% year over year. Gross margin reached 83%. The company raised its full-year revenue outlook again and now expects roughly $246 million to $250 million.

That growth is being driven largely by case volume rather than by adding thousands of software seats.

The reimbursement structure changes everything. HeartFlow's FFRCT analysis has broad U.S. payer coverage, while its newer plaque analysis now has a Category I CPT code and coverage from major insurers. Current Medicare rates illustrate the size of the economic unit: roughly $877 for FFRCT and $950 for plaque analysis in the hospital outpatient setting, with physician-office rates around $887 and $1,021 respectively.

The exact amount HeartFlow receives is different from the reimbursement paid to providers, but the payment system gives hospitals and physicians a clear mechanism for getting paid when they use the technology.

That is much easier to scale than asking every hospital CFO to create another discretionary AI software budget.

Reimbursement is hard to earn. Companies need clinical evidence, coding, payer coverage and physician adoption, sometimes over many years. Once those pieces fall into place, though, reimbursement creates a powerful commercial moat.

HeartFlow's latest results show the payoff. A healthcare AI company can now grow close to 50% while posting gross margins above 80% from AI-supported clinical analyses that are ordered one patient at a time.

If you want more recent data on this point, please see our latest healthcare AI market report.

Can a free AI tool for doctors make more money than paid healthcare SaaS?

OpenEvidence shows that a free AI product for doctors can become a huge business when pharmaceutical companies are willing to pay for access to the audience.

The reported numbers are remarkable. OpenEvidence is now generating close to $300 million of annualized revenue, roughly twice its run rate seven months earlier. More than 860,000 licensed U.S. clinicians have registered, and gross margin has reportedly reached about 90%.

Using those figures, we get roughly $350 of annualized revenue per registered clinician even though clinicians can use the core product for free.

Advertising explains the apparent contradiction. A physician searching for information about a disease or treatment is an extremely valuable audience for pharmaceutical companies. The advertiser can subsidize the product while the doctor removes the friction of paying for another subscription.

The business may still be early in its monetization curve. Reporting from The Information indicates that OpenEvidence is selling less than 5% of its available advertising inventory even at the current revenue run rate.

This may be the most surprising healthcare AI business model working today. OpenEvidence has taken an activity doctors already perform constantly, searching medical evidence, made it free and much easier, then monetized the professional attention around it.

The model obviously depends on keeping doctors' trust. Advertising that starts influencing clinical answers would damage the product quickly. But if the separation holds, free clinical AI funded by life-sciences marketing could end up being more lucrative than many enterprise subscriptions.

Is healthcare data licensing becoming a better business than selling AI software?

Healthcare data licensing is already a large and attractive AI business when the company generates proprietary data through its own clinical operations.

Tempus is the best example because we can see both sides of the machine. In its latest quarter, Tempus produced $382.5 million of total revenue. Diagnostics contributed $289.3 million, up 20%, while Data and Applications contributed $93.2 million, up 28%.

Inside the data business, licensing and modeling grew 36%, and Tempus signed roughly $200 million of new Data and Applications licenses during the quarter.

The margins show why the model is attractive. Data and Applications produced a gross margin of about 70%, compared with a lower margin in the diagnostics business. Tempus can perform a clinical test once, add the resulting molecular and clinical information to its dataset, then use the accumulated data again in pharmaceutical research, trial design, patient identification and AI model development.

That is a much better structure than buying generic datasets from someone else and trying to resell access.

Tempus now stores more than 500 petabytes of data and says roughly 55% of U.S. oncologists and 65% of academic medical centers are connected to its network. The company's diagnostics operation continuously refreshes the asset rather than leaving it as a static database.

The model also seems to be reaching financial scale. Tempus reported positive adjusted EBITDA in the latest quarter while raising its full-year revenue guidance to around $1.6 billion.

Proprietary healthcare data ranks highly, especially when data creation is already paid for by another business. The clinical operation generates the raw material, and AI makes that material increasingly valuable to another set of customers.

If you want more recent data on this point, please see our latest healthcare AI market report.

Can AI drug discovery make money before it creates a successful drug?

AI drug discovery can generate revenue before a drug succeeds, but the business still looks much more like biotech than software.

Schrödinger gives us a useful side-by-side comparison. In its latest quarter, software generated $32.5 million of revenue with a 71% gross margin. Annual contract value grew 27%, even though recognized software revenue fell because the company is shifting customers toward hosted subscriptions.

Its drug-discovery business generated $23 million during the same quarter, but $10 million came from a collaboration milestone. Cost of revenue for drug discovery was almost $16 million, leaving a much thinner gross margin than software.

Recursion makes the risk even clearer. Its latest quarterly revenue was only $7.7 million, mostly from collaborations, while R&D spending reached $89.6 million. The company still expects to spend hundreds of millions of dollars in cash operating expenses this year.

There is nothing inherently wrong with those economics. Drug development has always involved years of spending before a successful asset reaches the market, and one successful medicine can eventually generate billions.

Calling this a proven AI software model would be misleading, though. Partnership upfront payments and milestones help finance the research, but revenue jumps around depending on scientific progress and contract events.

For a company such as Schrödinger, selling computational tools to pharma is currently the more predictable business. Owning drug assets offers much larger upside if one succeeds, together with much larger risk.

Are patient-facing healthcare AI agents a real business yet?

Patient-facing healthcare AI agents have clearly proven they can operate at huge scale, but we still cannot say with the same confidence that the economics are proven.

Hippocratic AI has now passed 250 million patient interactions. The company says more than 300 clinical use cases are live, it has more than 50 EHR integrations and seven of the world's top 20 pharmaceutical companies use its technology.

Those numbers are far beyond demo territory.

The use cases are also commercially sensible. Healthcare organizations already pay people to make post-discharge calls, chase screening appointments, recruit trial participants, follow chronic-care patients, check medication adherence and contact patients who have disappeared from care.

Hippocratic AI is now moving from individual agents toward "orchestrators" that coordinate several agents around outcomes such as lower readmissions, better HEDIS scores and improved chronic-care management. That shift is worth watching because selling an outcome can support much higher pricing than selling automated phone calls.

The missing piece is financial transparency. We still do not have public revenue, renewal and margin data comparable with Waystar, HeartFlow or Tempus.

Our confidence should stay lower here. Patient-facing agents have passed the scale test. They have not yet passed the economic test publicly.

If the revenue eventually matches the interaction volume, this category could move up the ranking very quickly.

Who should actually pay for healthcare AI?

The best healthcare AI companies charge whoever captures the financial benefit, and that is often someone other than the person actually using the product.

Ambient AI is used by doctors, while the health system usually pays because it benefits from more physician capacity, faster documentation and potentially lower turnover.

HeartFlow is used in clinical care, while reimbursement from insurers and government programs makes the economics possible.

Cohere's prior-authorization technology affects physicians and patients, but health plans pay because they control the authorization process and its administrative costs.

Tempus' data is ultimately produced through patient testing, while pharmaceutical companies pay to use the resulting datasets and models for research.

OpenEvidence takes the separation furthest. Doctors get the product free, and life-sciences advertisers pay for access to their attention.

Healthcare often separates the user, buyer and financial beneficiary. That makes pricing less obvious, but it also creates more routes to monetization.

Consumer AI subscriptions often struggle precisely here. A patient may find an AI navigator useful but still refuse another $20 monthly subscription. The insurer may save thousands of dollars if that same navigator keeps the patient on treatment or prevents unnecessary care.

Whenever we look at a new healthcare AI company, one of the fastest tests is simply to ask who makes or saves money when the product works.

If the vendor cannot answer that clearly, monetization usually becomes difficult.

What actually gives a healthcare AI company a moat?

The strongest healthcare AI moats today come from workflow, proprietary data, reimbursement and distribution because model quality alone is becoming easier to copy.

Epic's rapid move into ambient charting is the warning. A startup can spend years building excellent note generation, then suddenly compete with a similar feature built directly into the dominant EHR.

General AI models are also getting better while inference costs keep falling. A temporary lead in summarization quality therefore gives us much less confidence than it did a few years ago.

The harder assets live around the model.

HeartFlow has accumulated reimbursement coverage, clinical evidence and integration into cardiology workflows. Tempus continuously produces proprietary molecular and clinical data through its testing network. Waystar sits inside healthcare payment flows and sees huge numbers of transactions. OpenEvidence has direct distribution to hundreds of thousands of verified clinicians. Cohere has trained its systems around millions of authorization decisions and health-plan policies.

Those assets survive even if everyone gains access to a better foundation model next year.

Healthcare's complexity can actually help the winners here. Getting payer coverage, integrating deeply with Epic, producing regulatory evidence or building longitudinal clinical datasets takes years. A new competitor cannot recreate those advantages by calling a newer API.

The healthcare AI companies we trust most are building assets that become painful for customers to replace.

If you want more recent data on this point, please see our latest healthcare AI market report.

Which healthcare AI business models get stronger as they scale?

The best healthcare AI models get more valuable as they handle more claims, tests, clinical encounters, searches or authorization decisions.

Waystar gains more payment and denial data as transaction volume increases. Tempus gets more clinical and molecular information every time its diagnostic network processes another patient. OpenEvidence becomes more valuable to advertisers as more physicians use it. HeartFlow spreads its commercial, regulatory and infrastructure costs across a growing number of reimbursed cases.

Cohere has a similar effect in prior authorization. More requests create more examples of clinical documentation, health-plan policies and decision patterns that can be reused across its platform.

These businesses gain something beyond revenue when they grow.

The contrast with implementation-heavy AI software is important. A vendor that has to hire another team of engineers and consultants every time it wins a hospital can double revenue without improving the underlying economics very much.

Healthcare contains an enormous number of repetitive events: visits, claims, prescriptions, tests, calls, authorizations and evidence searches. The strongest AI companies attach themselves to one of those events and collect value every time it happens.

Volume is especially powerful when every event also produces proprietary data. More usage should improve both the revenue base and the underlying asset.

So which healthcare AI business models are actually working now?

Healthcare AI is already producing several strong business models, and the clearest winners are revenue-cycle automation, clinical workflow AI, reimbursed diagnostics, proprietary data licensing and physician-attention advertising.

Revenue-cycle AI sits near the top because customers can connect the product directly to money collected or administrative work removed. Waystar shows that this market can already support billion-dollar revenue with adjusted EBITDA margins above 40%.

Clinical workflow AI has also broken through. Ambient documentation created the opening, while the leading vendors are now moving into coding, orders, clinical intelligence and revenue workflows. That broader model has a much better future than standalone note generation.

Reimbursed diagnostic AI deserves the same level of confidence. HeartFlow's latest 48% revenue growth and 83% gross margin show how powerful AI can become once reimbursement turns every clinical use into a potential revenue event.

Tempus shows another route. Diagnostics create proprietary clinical data, and the company can then license that data and AI-driven analysis to pharmaceutical companies at higher margins.

OpenEvidence has uncovered an entirely different model. Free clinical AI can generate hundreds of millions of dollars when the users are verified physicians and advertisers already spend heavily to reach them.

Prior-authorization AI also looks strong. Cohere has enough real transaction volume to show that automation works operationally, although private-company financial disclosure still gives us less evidence than we have for Waystar or HeartFlow.

Patient-facing agents sit one level lower for now. Hippocratic AI has reached extraordinary interaction volume, yet the revenue and margin evidence remains too limited for us to rank the model among the proven leaders.

AI drug discovery is further down if our question is about a reliable business today. The scientific upside remains enormous, but Recursion's current spending and Schrödinger's much better software margins show how different drug economics remain from software economics.

Across all these categories, one rule explains most of the winners. Healthcare organizations do not pay much simply because AI has become smarter. They pay when AI frees a doctor, gets a claim paid, performs a reimbursable analysis, makes proprietary clinical data more valuable, removes administrative work or captures a scarce professional audience.

The sharpest final answer is that healthcare AI works best when the AI is attached to a transaction or workflow that already has a buyer, a budget and measurable financial value. Companies still trying to sell "AI" as the product itself have a much harder road.

Healthcare AI business model	Evidence today	Why customers pay	Our judgment
Revenue-cycle and payment AI	Billion-dollar-scale revenue, strong growth, 40%+ adjusted EBITDA margins at Waystar	More revenue collected and less administrative work	Working very well
Clinical workflow and ambient AI	100,000+ clinician deployments and 100M+ annual conversations at leading platforms	Physician time, capacity, coding and workflow gains	Working very well
Reimbursed diagnostic AI	HeartFlow revenue +48%, 83% gross margin in latest quarter	Each clinical use can be reimbursed	Working very well
Proprietary healthcare data and AI	Tempus data revenue +28%, licensing/modeling +36%	Pharma pays for better data, models and trial decisions	Working
Free physician AI + advertising	OpenEvidence near $300M annualized revenue and ~90% gross margin	Pharma pays for verified physician attention	Working very well
Prior-authorization and payer AI	12M+ annual authorizations at Cohere, high real-time approval rates	Lower administrative costs and faster decisions	Working
Life-sciences software	Schrödinger software gross margin around 70%	Pharma already has large research software budgets	Working
Patient-facing AI agents	250M+ interactions at Hippocratic AI	Cheaper and broader patient outreach	Promising, economics still unclear
AI drug discovery partnerships	Large collaborations but very high R&D spending	Milestones and eventual drug value	Commercially mixed
Standalone ambient note generation	Strong usage, but Epic and Microsoft now bundle it	Documentation savings	Vulnerable on its own
Generic consumer healthcare AI subscriptions	Much weaker large-scale proof	Patients pay directly	Still relatively weak

If you want more recent data on this point, please see our latest healthcare AI market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.