Home / Blog

Startups and health care

Who pays whom, how much, when and why

Nigam Shah tells people building AI for health care to answer four questions about money before they write any code. Public prices show what the questions catch. The person who uses a product is often not the one who pays, and the money arrives on the buyer’s calendar.

Main takeaways
  • The user is often not the payer. Americans paid 11% of their $5.3 trillion health bill in 2024 directly. Insurers and governments sent most of the money, and households funded 28% once premiums and payroll taxes are counted.
  • Buyers pay most for what they cannot skip. A first-year HITRUST security certification costs $35,000 to more than $500,000, and vendors pay because health care customers ask for proof of security. Optional testing tools for AI agents start at $100 a month.
  • Price follows what the buyer cannot copy. Drugs aimed at targets with human genetic support were 2.6 times as likely to reach approval, yet the evidence itself can be downloaded free. Lilly agreed to pay about $1.0 billion upfront for a company with a gene-editing drug in trials.
  • Most of a headline deal is a promise. AI drug-discovery deals announced from January to mid-October 2025 paid about 2.7% of their headline value upfront.
  • Money moves on the buyer’s calendar. Health systems took 6.6 months on average to buy an AI tool, and health plans took 11.3 months. One community college approved its contract in June for a contract year starting on 1 July.

In August 2025 the venture firm Define Ventures announced a new adviser on artificial intelligence, Nigam Shah. Shah is the chief data scientist at Stanford Health Care and a professor of medicine and biomedical data science at Stanford University. In the announcement he gave founders a test. “To succeed, you must first understand who pays whom, how much, when and why,” he said. “This is the single most important thing to learn before you even write a single line of Python code for AI in healthcare.”

Many technical people start from the other end. A student who trains a model that reads chest X-rays well has built something useful, and it is natural to assume that useful things sell. Whether anyone will pay is a separate question with its own evidence. Much of that evidence is public. Companies report prices to stock market regulators, and colleges vote on contracts in open board meetings.

What a product is worth to the people it helps and what someone will pay for it are two different numbers, and Shah’s four questions measure the second. Each question fills one line of a payment slip. Figure 1 fills in five slips, from a dating app to a billion-dollar takeover. Three come from company filings and board minutes, and two from published cost estimates.

Four questions, one slipWho pays whom, how much, when and why?

Payment slip · No. 1Paid monthly
Who pays whomA Hinge subscriber pays Match Group, through Apple or Google
How muchAn average of $31.97 a month per paying user in 2025
WhenEach month the plan renews. Apple passes the money on within 45 days after the end of its fiscal month.
WhyTo meet people
Figure 1. Five payments, each written as Shah’s four questions. Three come from company filings and board minutes; the security audit and the training data come from published cost estimates. Sources: Match Group’s 2025 annual report and Apple’s payment terms (dating app); Carroll Community College board minutes, June 2023 (college counselling); Secureframe’s cost guide (security audit); Epoch AI’s interviews with people in the industry (training data); Lilly’s announcement of its deal for Verve Therapeutics, June 2025 (biotech). Links are in the text below.

The person who uses it is often not the one who pays

Health care shows the gap most clearly. The United States spent $5.3 trillion on health in 2024, or $15,474 per person, according to the Centers for Medicare & Medicaid Services (CMS), the federal agency that keeps the national tally. Patients paid 11% of it directly, at the pharmacy counter or in a bill from a clinic. Private insurance companies paid 31%. Medicare, the federal program mainly for people 65 and older, paid 21%. Medicaid, which states and the federal government run for people with low incomes, paid 18%.

Those are the organizations that send money to hospitals and doctors. CMS also asks whose money it was in the first place, and calls that party the sponsor. Households sponsored 28% of spending once their insurance premiums and payroll taxes are added to what they paid at the counter. The federal government sponsored 31%, and private businesses 18%. Figure 2 switches between the two views.

The US health dollar, 2024Who pays for 100 cents of American health care?

  • Patients, out of pocket11%
  • Private health insurance31%
  • Medicare21%
  • Medicaid18%
  • Other programs and public health11%
  • Other insurance programs and investment8%

Patients paid 11 cents of each dollar themselves. Insurers and government programs sent most of the rest.

Figure 2. US national health expenditure in 2024, $5.3 trillion, one square per percentage point. “Who sends the money” is CMS’s split by source of funds. Its palest 8 squares come from CMS’s detailed tables: children’s, military and veterans’ health programs (3.8%) and investment in research, buildings and equipment (4.5%). “Whose money it is” is CMS’s split by sponsor, the party that finally finances the bill, through premiums, out-of-pocket payments or taxes. CMS rounds the sponsor shares, so they add to 99. Gold marks households in both views. Sources: CMS National Health Expenditure fact sheet, 2024 highlights and tables.

A company that sells into health care meets this chain at every step. The doctor who uses a tool is rarely the one who buys it. The hospital manager who buys it spends a budget that someone else set.

User, chooser and payer

Three roles decide whether a product earns money, and they often belong to different people.

  • The user works with the product every day.
  • The chooser decides which product an organization buys.
  • The payer is the budget the money comes from. The person who controls that budget is called the budget owner, and that person’s signature releases the money.

The split shows up well outside hospitals. Figure 3 is a short quiz on six products.

A quick quizDoes the person using it pay for it?

Answered 0 of 6 · right 0

Hinge

A dating app with a paid subscription.

The user paysHinge took in $31.97 a month from each paying user in 2025.

TimelyCare

Virtual counselling and support for college students.

The college paysStudents pay nothing when they use it. At Finger Lakes Community College, an $18-a-semester student health fee helps cover the bill.

Google Search

The search engine most people use every day.

Advertisers paySearch ads brought in $224.5 billion of Alphabet’s $402.8 billion in revenue in 2025. All advertising brought in 73%.

Seismometer

Epic’s tool for checking how well an AI model works on a hospital’s own patients.

It is freeEpic gives it away as open-source software. Epic earns its money from the hospitals that license its health record system.

Open Targets

A database that ranks which genes make promising drug targets.

It is freeFive drug companies and two public research institutes run it as partners. Anyone can download the data free.

KLAS reports

Ratings of health care software that hospital buyers read before they choose.

Software vendors payHospitals can get the research by giving feedback on the software they use, or by paying for a membership. Software vendors pay for memberships and reports.

Figure 3. Six products and who pays for them. Sources: Match Group’s 2025 annual report (Hinge); Finger Lakes Community College board minutes, June 2025 (TimelyCare); Alphabet’s 2025 annual report (Google); Epic’s Seismometer repository on GitHub; the Open Targets partnership page; KLAS Research’s data use policy.

A free product still has a payer. Someone has decided that giving it away helps them sell something else. Epic, whose electronic health record software runs in many US hospitals, gives away Seismometer. More than 85% of Epic’s customers now use Epic’s own AI, Epic said in March 2026. A startup that sells a similar check competes with a price of zero, set by the company that already runs the hospital’s records.

Every hand on the money keeps some

The “whom” in Shah’s question is often more than one company. Money passes through middlemen, and each one keeps a share.

An app sold through Apple’s App Store is the simple case. Apple keeps 15% of each sale for developers with less than $1 million a year in proceeds, under its Small Business Program. On a $14.99 monthly plan, the developer receives about $12.74.

AI labs buy expert work the same way. Mercor recruits people with advanced degrees to write and grade the examples that AI models train on, and sells that work to the labs. The experts keep 60% to 70% of what Mercor’s customers pay, according to Dealroom, and Mercor keeps the rest.

Mercor’s page for biology experts says past rates ran from $70 to $105 an hour. Most of the money comes from a few buyers. About 91% of Mercor’s gross revenue in the first half of 2026 came from OpenAI and Anthropic, The Information reported, citing internal documents. AI Weekly, which summarized the story, notes that the exact split by customer has not been published. When a handful of customers supply nearly all the money, losing one of them takes a large bite out of revenue.

Follow the moneyWhere does one payment end up?

$14.99A subscriberpays each month
$2.25Applekeeps 15%
$12.74The developerreceives 85%
DeveloperApple

On a $14.99 plan, Apple keeps $2.25 and the developer receives $12.74.

Figure 4. How a payment splits between a middleman and the company or person doing the work. App subscription: Apple’s 15% rate for developers earning under $1 million a year. Expert data: Dealroom’s report that Mercor’s contractors keep 60% to 70% of billings; the striped band is the range.

Buyers pay most for what they cannot skip

The “why” sets the price more than quality does. A software company that sells to hospitals or insurers has to pass each customer’s security review, and security teams ask for proof. A common proof is a SOC 2 report, an outside auditor’s check of how a company protects its customers’ data. The audit alone costs from about $5,000 to more than $150,000. Secureframe, which sells compliance software, puts the whole effort, from preparation to a year of upkeep, at $80,000 to $350,000.

HITRUST certification, a stricter standard common in health care, costs from $35,000 to more than $500,000 in its first year, AccountableHQ estimates. Some customers make it a condition of doing business. In June 2015 three large health insurers, Anthem among them, said they would require the vendors that handle their members’ data to become HITRUST certified, GovInfoSecurity reported.

Vendors pay these sums because a missing report can stop a sale. Spending on evidence that customers do not ask for is a choice, and its price starts low. Tools that test AI agents, such as phone bots that book appointments, start at $100 a month at Coval and $500 a month at Cekura for small teams. Coval’s plan for regulated industries such as health care starts at $4,500 a month. A vendor can spend as little or as much on optional testing as it likes. It cannot buy a certification its customers ask for at less than the floor. Figure 5 sets the two side by side.

Asked for or optionalWhat does the cheapest year of AI-agent testing cost?

$30k
AI-agent testing, a year of an entry plan$1.2k–$6k
AI-agent testing, Coval’s plan for regulated industriesfrom $54k
HITRUST e1 certification, first year$35k–$90k
HITRUST i1 certification, first year$80k–$220k
SOC 2 report, preparation to a year of upkeep$80k–$350k
HITRUST r2 certification, first year$200k–$500k+
$100$1k$10k$100k$1M
Buyers ask for itOptional

Entry plans for agent testing list at $100 to $500 a month, about $1,200 to $6,000 a year. Plans for regulated industries start at $54,000. The cheapest certification that customers ask for starts at $35,000.

Figure 5. Costs of security reports that health care customers ask for (red) against a year of optional AI-agent testing at list price (blue), on a logarithmic scale where each step is ten times the last. HITRUST figures are first-year budgets including staff time; the SOC 2 figure is Secureframe’s estimate from preparation through a year of upkeep. The regulated-industries plan is Coval’s published starting price, $4,500 a month. Sources: Secureframe; AccountableHQ; Coval and Cekura pricing pages.

Some evidence sits between the two because it helps win deals. KLAS Research rates health care software. Hospitals can get its research by paying for a membership or by giving feedback on the software they use, while software vendors pay for memberships and reports. Anterior, which sells AI that reviews requests for prior authorization, an insurer’s approval that some treatments need in advance, advertises “99.24% clinical accuracy, independently validated by KLAS Research.”

A strong “why” is tied to a cost the buyer already counts. The hospital purchasing group Premier surveyed 516 hospitals about their 2022 claims. Nearly 15% of claims sent to private insurers were denied at first, and fighting each denial cost a provider $43.84 on average. A product that prevents denials can point to that number on the hospital’s own books.

Buyers pay for what they cannot make or download

Some of the widest gaps between worth and price are in drug development. Drugs aimed at targets with human genetic evidence for the disease being treated were 2.6 times as likely to reach approval, Eric Minikel and colleagues reported in Nature in 2024. Most drugs that enter trials fail, so evidence that raises the odds is worth a great deal to a drug company.

The evidence itself is cheap or free. Open Targets, a partnership of drug companies and public research institutes, releases its genetic evidence on drug targets under a free licence. Software for interpreting genetic results costs more, and the University of Iowa pays about $32,000 a year for each seat of QIAGEN’s Ingenuity Pathway Analysis. A drug company can also hire statistical geneticists and run the analysis itself.

The price rises as the buyer gets something it cannot copy. In 2020 Ionis invested $10 million in Empirico, with up to $30 million more tied to early research milestones, to develop drugs against up to ten genetically supported targets. In 2025 GSK agreed to pay Empirico $85 million upfront for worldwide rights to a drug already in its first human trial. Eli Lilly agreed that year to buy Verve Therapeutics, which was testing a one-time gene-editing treatment for high cholesterol, for about $1.0 billion in cash upfront. Figure 6 lines up the prices.

The price ladderHow does the price change as the buyer gets more?

$0Data$32kSeat$10MTargets$50MDeal$85MDrug$1.0BCompany
InformationComes with rights to targets or drugs

At the top, Lilly agreed to pay about $1.0 billion upfront for Verve Therapeutics, a company with a gene-editing drug in human trials.

Figure 6. Prices paid in different forms, from genetic data anyone can download to a company that owns a drug aimed at a genetically supported target. Step heights follow a logarithmic scale; green steps are purchases that came with rights to targets or drugs. The $10 million is Ionis’s equity investment in Empirico, so Ionis received shares in return; up to $30 million more depended on early research milestones. Sources: Open Targets licence; University of Iowa (QIAGEN seat price); Empirico (Ionis deal); insitro (Bristol Myers Squibb deal); GSK (Empirico deal); Lilly (Verve deal).

Selling target ideas has been a hard business even for well-funded companies. BenevolentAI, a London company that used AI to propose drug targets, agreed a deal in 2021 that valued it at up to €1.5 billion. The deal took it onto the stock market in 2022. Its revenue in 2022 was £10.6 million, mostly from its partnership with AstraZeneca. In March 2025 its shares were delisted.

A billion-dollar deal is mostly a promise

The headline number on a drug-discovery deal adds up every payment the partner might ever make. Only the first part is certain.

Upfront payments and milestones

The upfront payment is cash paid when a deal is signed. Milestones are payments due only if the drug or target reaches a later stage, such as a trial or an approval. People in the industry call the milestone total “biobucks,” because much of it never turns into dollars.

From January to mid-October 2025, AI drug-discovery deals carried about $29.7 billion in headline value and about $800 million upfront, or 2.7%, according to Nature Biopharma Dealmakers. Chris Dokomajilar of the deal tracker DealForma compared that with about 10% upfront for deals to license a specific biologic drug and about 6% for small-molecule drugs. Figure 7 asks for a guess first.

Headline value against cashOf a $1 billion AI drug-discovery deal, how much arrives on signing day?

25%
$1,000 million headline
AI drug discovery, Jan to Oct 2025
2.7% Small-molecule deals
~6% Biologic drug licences
~10%

One deal over time: insitro and Bristol Myers Squibb

2020202120222023202420252026
$50M upfront, Oct 2020
$25M in milestones, Dec 2024

Across AI drug-discovery deals from January to mid-October 2025, 2.7% of headline value was paid upfront, about $27 million per $1 billion. The rest depends on milestones years away.

Figure 7. Upfront payments as a share of headline deal value. AI drug discovery: about $800 million upfront on $29.7 billion of headline value, for deals from January to 17 October 2025. Comparisons for biologic and small-molecule deals come from DealForma, as quoted in Nature Biopharma Dealmakers. The insitro timeline shows the upfront payment and the milestone payments tied to the first ALS target, in a deal with more than $2 billion in possible milestones. It leaves out up to $20 million in near-term operational milestones whose payment dates insitro has not published. Sources: Nature Biopharma Dealmakers; insitro.

Milestones also take years. Insitro signed with Bristol Myers Squibb in October 2020 for $50 million upfront and more than $2 billion in possible milestones. In December 2024, four years later, it received $25 million in milestone payments when the partners selected their first target for ALS. Milestones in company sales behave the same way. By 2016, roughly a third of the milestone payments that had come due in life-science acquisitions had been paid. The biotech investor Bruce Booth reported the figure from data collected by the deal firm SRS Acquiom.

The buyer’s calendar sets the clock

The “when” decides how long a young company must survive before money arrives. Buying takes time. Health systems took 6.6 months on average to buy an AI tool in 2025, faster than the 8.0 months they took for ordinary IT, according to a Menlo Ventures survey. Health plans took 11.3 months, slower than their 9.4 months for ordinary IT. Payment terms come on top. Apple pays developers within 45 days after the end of each of its fiscal months, which run four or five weeks. A US card payment through Stripe settles into the seller’s balance in two business days. Figure 8 starts the clock on all of them at once.

The race to cashYou start selling today. When does each buyer’s money arrive?

Card payment settles with Stripe2 business days
App Store sale, paid by Apple45 to about 80 days
Health system buys an AI tool6.6 months
Health plan buys an AI tool11.3 months
insitro’s first-target milestones from Bristol Myers SquibbAbout 50 months
1 day1 week1 month1 year4 years

A card payment settles in two business days. The milestone payments for insitro’s first ALS target took about 50 months.

Figure 8. Time from the start of selling to money in hand, on a logarithmic time scale. Green dots are cash arriving. Blue dots are the end of the average buying cycle for AI tools, before payment terms add more time. The App Store range runs from 45 days to about 80 days, depending on where in Apple’s four- or five-week fiscal month the sale falls. The insitro dot is measured from the day the deal was signed, which also brought $50 million upfront. Sources: Stripe; Apple; Menlo Ventures, 2025: The State of AI in Healthcare; insitro.

Some buyers can only say yes in certain months. Drug companies plan their marketing budgets once a year. Phreesia, which sells patient messaging to drug brands, says in its annual report that these contracts are “typically negotiated during the fourth quarter of the calendar year.” Most run for a year and must be sold again the next year. The spending follows a calendar too. OptimizeRx sells messaging services to the same brands. Its annual report says many of them spend the largest share of their brand marketing in the fourth quarter, so its first quarter is usually slower.

Colleges decide in public. At Carroll Community College in Maryland and Finger Lakes Community College in New York, trustees approved TimelyCare contracts at June board meetings. At Finger Lakes, each contract year runs from 1 July to 30 June. Figure 9 shows what a missed window costs.

Budget seasonsIn which month do you first meet the buyer?

A collegeFirst money on 1 July next year, 10 months later
A drug company’s brand teamFirst money in January next year, 4 months later
This yearNext yearThe year after
First meetingDecision windowWaitingMoney starts

A first meeting in September misses the college’s spring vote, so the money starts the following July.

Figure 9. How a buyer’s budget calendar turns the month of a first meeting into a wait. College: at Finger Lakes Community College, trustees approved the contract in June for a contract year starting on 1 July. Drug company: contracts for the next calendar year are typically negotiated in the fourth quarter. The cut-offs (a first meeting by April for the college, by October for the drug company) are illustrative. Sources: Carroll Community College and Finger Lakes Community College board minutes; Phreesia annual report.

When the user does pay

Consumer apps answer the first question simply, because the user pays. Hinge took in an average of $31.97 a month from each paying user in 2025, its owner Match Group says in its annual report. The hard parts are what it costs to find each paying user and how long that user stays.

Most people who install an app never pay. RevenueCat, a company that runs subscription billing for apps, found in its 2026 report that the median health and fitness app turned 2.9% of downloads into paying users within 35 days. At $5.30 per install, the North American average in an AppsFlyer guide, each paying user costs about $183. That figure is the customer acquisition cost, the marketing spent to win one paying customer.

Paying users also leave. Among people who subscribed monthly to AI apps, 6.1% were still paying after 12 months, RevenueCat found in data that TechCrunch reported in March 2026. That is the same as losing about 21% each month at a steady rate, though in practice more people cancel in the first months. A $14.99 plan that nets $12.74 after Apple’s fee then brings in about $58 per paying user in the first year, against the $183 it cost to win them. Figure 10 lets you change the inputs.

The leaky bucketDoes a paying user earn back what it cost to find them?

$14.99
$5.30
2.9%
6.1%
Users still paying after each month, out of 100100012345678910116.112
Cost to win one paying user
$183
What they pay in year one
$58

Each paying user costs $183 to win and brings in $58 in the first year. The app loses about $125 on each one before paying for anything else.

Figure 10. A cohort of 100 paying users, counted after each month, and the year-one revenue from each against the cost of acquiring them. Revenue is after Apple’s 15% fee and before any other cost. The monthly loss rate is the constant rate that leaves the chosen share still paying at 12 months. Default inputs: median download-to-paid conversion for health and fitness apps (RevenueCat, 2026); 12-month retention of monthly AI-app subscriptions (RevenueCat, via TechCrunch); North American average cost per install across all apps (AppsFlyer). The hard-paywall preset uses RevenueCat’s 10.7% median conversion, across all categories, for apps that require a purchase or a trial sign-up before any use.
Averages hide the spread

The inputs in Figure 10 come from different groups of apps. The conversion rate is the median for health and fitness apps, the best category in RevenueCat’s data. The retention rate is for AI apps, and the install price is an average across all apps in North America. Most health and fitness subscribers buy yearly plans. Any one app can do far better or worse. The arithmetic holds for any inputs, and a paying user has to bring in more than they cost before they leave.

Selling to an organization changes the arithmetic. A college that buys TimelyCare pays one price for a whole group of students. Carroll Community College pays $22,500 a year for an estimated 700 full-time-equivalent students, about $32 each. There is no advertising cost per student and no monthly cancellation. The contract is priced by the year and approved on the college’s calendar.

The whole slip in one set of minutes

The four questions can often be answered before a product exists, from public records. On 4 June 2025 the trustees of Finger Lakes Community College renewed the college’s contract with TimelyCare, according to the minutes of their meeting. The payer is the college and the payee is TimelyCare. The amount is at most $156,325 for the first year, rising to $165,845 in the third. The contract runs from 1 July 2025 to 30 June 2028. TimelyCare had already served the college for four years.

The minutes also record where part of the money comes from. To offset the cost, the college moved TimelyCare into its student health fee, at $18 a semester.

Further reading

startups health care economics AI