Context

Julie Yoo, general partner at Andreessen Horowitz and an 18-year veteran of the healthcare industry, joins host Sophia Dew to explain why she believes healthcare stands to benefit more from AI than almost any other sector. Yoo previously spent seven years as a software engineer before switching into health tech during the federal push to digitize medical records, then spent nine years as co-founder and chief product officer of Kairos Health before selling the company and joining a16z to lead its healthcare investing team. She traces healthcare's slow, unnatural history of technology adoption (paper records, fax machines, government-subsidized electronic health records, pandemic-forced telehealth) and argues that this same slowness now means healthcare skipped the expensive middleware software layer other industries built, letting it leapfrog straight into agentic AI. For a PM, the episode is a case study in how a historically change-resistant, highly regulated industry can become a founder-friendly opportunity once cost pressure, consumer expectations, and AI capability all cross a threshold at the same time.

The Big Idea

Healthcare's historically slow adoption of enterprise software is now an advantage rather than a liability, because the industry never built the expensive legacy middleware layer other industries have to rip out, so it can leapfrog directly into agentic AI instead of migrating away from decades of sunk-cost software.

Yoo argues this is compounding with unsustainable cost pressure on hospitals, insurers, and employers, and a consumer base whose expectations have been reset by other on-demand industries, to produce what she calls the first genuinely organic technology adoption wave healthcare has ever had.

Key Insights

Healthcare's tech laggard status became a leapfrog advantage

Other industries spent roughly two decades and tens of billions of dollars building workflow software (SaaS) middleware on top of their core infrastructure, then had to train a generation of workers on it, and now must rip much of that out to adopt agentic AI. Healthcare, by contrast, mostly skipped that layer. It built electronic health record systems as a basic system of record and otherwise relied on human labor to fill gaps. Yoo argues this means healthcare has far less sunk-cost bias holding it back, so it can adopt agentic AI directly rather than migrating away from an expensive prior generation of software.

This is healthcare's first organic AI adoption wave

Every previous technology shift in healthcare was forced from outside the system: the federal government paid doctors tens of thousands of dollars each to implement electronic health records, and the COVID pandemic forced telehealth adoption by making in-person visits impossible. AI is different, Yoo says, because doctors are adopting AI scribes purely because the tools work and materially improve their day-to-day job, a bottom-up, product-led-growth pattern rather than a mandate. This distinction matters for anyone selling into healthcare: it signals that frontline clinician pull, not policy or crisis, is now driving adoption.

Three forces converged to break healthcare's status quo

Yoo names three simultaneous pressures that created the current moment: consumer expectations reset by on-demand services like Uber, Airbnb, and Instacart made healthcare's experience feel newly unacceptable by comparison; the incumbent system (hospitals, insurers, employers) faced unsustainable cost increases from labor shortages, a COVID-driven exodus of healthcare workers, and government pressure to cut bloat; and AI finally made it possible to replicate parts of an expertise that used to require seven-plus years of medical training, something no earlier technology could do. She calls AI "the straw that broke the back" of a system already under strain from the other two forces.

Deductibles taught consumers to shop for healthcare

A specific mechanism Yoo points to: employers shifted from fully covering insurance costs to introducing deductibles, meaning consumers now pay out of pocket for care below a certain threshold before insurance activates. That direct cost exposure, something consumers previously never felt, is what pushed people to notice how expensive and low-quality routine healthcare services were, and to start seeking out cash-pay alternatives that cost a fraction of the price for comparable or better service.

Medical AI today solves access, not the full care journey

Yoo breaks the patient journey into three parts: getting an initial answer or diagnosis, verifying and acting on it (testing, prescriptions, procedures, in-person care), and staying with a patient over time as new issues arise. She says general-purpose AI tools like ChatGPT and Claude have effectively solved the first part, lowering the barrier to a first opinion from a multi-month wait or a two-thousand-dollar emergency room visit down to opening an app. The real founder opportunity, in her view, sits in the second and third parts: what happens after the AI gives you an answer, and who stays with you for the long run.

Cash-pay models exploded because AI cut delivery costs

Yoo says that seven years ago, pitching a consumer-paid healthcare business would have gotten a founder turned away by investors, since consumers had no established willingness or ability to pay directly and the underlying cost of delivering a medically credible service was too high. AI has since cut that delivery cost by roughly 100 times, according to her estimate, which is what makes disruptively low, direct-to-consumer pricing viable. This flips the traditional healthcare product design logic, which optimized for insurers and doctors first and treated the consumer as an afterthought, into one where the consumer is the primary customer.

Mental Models & Frameworks

AI native and AI proof as a founder filter

Yoo evaluates healthcare startups on whether they are both AI native (built around agentic AI from the ground up, not bolted onto old workflows) and AI proof (defensible against a general-purpose AI model simply replicating what they do). The second half matters because a strong medical LLM can answer a lot of questions, but it cannot draw blood, deliver in-person care, or perform a regulated clinical act. Companies that look like an ordinary service or retail business from the outside, but are highly AI native and AI proof on the inside, mirror what she calls "full stack challenger" businesses from earlier tech cycles: a surface area competitors can't easily copy, plus a cost structure disruptive enough to fund continued innovation.

The three-part patient journey as an opportunity map

Yoo's framework for where founders should build: part one is getting an initial answer or diagnosis (largely solved by general LLMs today), part two is verifying and acting on that answer (testing, prescriptions, procedures, in-person care), and part three is staying with the patient over their whole life as new issues emerge. She points founders toward parts two and three as the areas with real unclaimed opportunity, since part one is increasingly commoditized by generalist AI tools.

Trade-offs & Nuance

Consumers accept unregulated AI advice for lower friction

Yoo acknowledges a real trade-off: medical advice from a general AI model is not regulated, licensed, or fully validated the way a doctor's advice is. But she argues consumers are rationally choosing it anyway, because the realistic alternative is a four-month wait for an appointment or a two-thousand-dollar emergency room visit. She frames this less as consumers being reckless and more as a system that has, for decades, offered no lower-friction legitimate option, so a lower-friction unregulated one wins by default.

Healthcare AI may lack the data it needs to fully mature

Yoo raises an underdiscussed limitation: training a genuinely medical-grade AI model may not be possible yet, because the primary data source, electronic health records, is sparse. A typical patient sees a doctor at most once a year, so the record captures almost none of a person's actual health narrative between visits. She argues this gap is itself a business opportunity: companies running personalized, "N of one" health experiments can generate the richer, continuous data that the next generation of medical AI models will need, rather than waiting for that data to already exist.

Practical Application

Target the after-the-answer part of the healthcare journey

If building or evaluating a healthcare AI product, avoid competing directly on giving a first diagnosis or answer, since general-purpose models like ChatGPT and Claude have already made that step fast and largely free. Instead, look at what happens after the answer: getting a diagnosis verified through actual testing, getting prescribed or referred, or getting ongoing longitudinal care. Yoo's framework suggests this is where defensible, monetizable products still need to be built.

Design healthcare products for the consumer as the primary buyer

Build the product experience around the person paying and using it directly, rather than optimizing first for the insurer and the doctor as historically was standard practice. Yoo points to a wave of cash-pay companies successfully doing this at prices roughly 100 times lower than legacy delivery costs allow, because AI reduced the underlying cost of delivering a credible service.

Combine an AI-native workflow with a real regulated clinical act

Structure the product so that AI drives efficiency and personalization internally, but the business still performs a legally regulated act (prescribing, diagnosing, drawing blood, delivering in-person care) that a pure chatbot cannot. Yoo cites Council Health, an a16z portfolio company, as an example: it pairs 24/7 asynchronous AI-native chat with real licensed doctors who can prescribe, refer, and diagnose, combining low-friction access with clinical authority a chatbot alone can't provide.

Treat sparse patient data as a build opportunity, not a blocker

If a healthcare AI product is data-constrained because electronic health records only capture a once-a-year snapshot of a patient, consider building a product that generates continuous, personalized data through ongoing use (an "N of one" approach) rather than waiting for better third-party data sets to appear. Yoo suggests this is exactly the gap that will produce the training data for the next generation of medical AI models.

Questions to Consider

• If a general-purpose AI model like ChatGPT or Claude can already answer most of your target users' first-pass questions for free, what part of the journey after that first answer is your product actually built to own?
• Does your healthcare product depend on a regulated clinical act (prescribing, diagnosing, in-person delivery) that an AI-native competitor without a licensed clinical layer could not easily replicate?
• If your industry, like healthcare, has under-invested in workflow software for decades, could that be a leapfrog advantage into agentic AI rather than a disadvantage, the way Julie Yoo argues it is for healthcare?
• What continuous, longitudinal data could your product generate through everyday use that the industry's existing records (like once-a-year doctor visits) simply don't capture today?

Bottom Line

Julie Yoo argues healthcare's decades of slow technology adoption removed the expensive legacy middleware layer other industries now have to rip out for AI, letting it leapfrog straight into agentic, AI-native workflows just as cost pressure and consumer expectations reach a breaking point. For builders, the real opportunity isn't answering a patient's first question, since general AI models already do that, it's owning what happens after: verification, regulated care, and staying with someone for life.

Case Studies Mentioned

Council Health's AI-native doctor practice

Council Health, an a16z portfolio company, runs a 24/7 asynchronous chat-based practice where patients interact with an AI-native interface backed by real, licensed doctors who can fully prescribe, refer, and diagnose. Yoo cites it as an example of combining the low-friction feel of chatting with a bot with the clinical authority that only a licensed physician can legally provide, a combination she says was not possible just a few years ago.

An Anthropic and AWS rare-disease hackathon

Host Sophia Dew describes a hackathon run by Anthropic and AWS Cloud called the Rare Disease Real Kid Hackathon, offering $50,000 in prizes. Organizers uploaded a child's genome and clinical data, with parental consent, and let the community generate insights. Yoo highlights it as an example of AI lowering the barrier to specialist-level intelligence for conditions where too few human specialists exist to serve the small population of patients who need them, especially for rare pediatric diseases where the economics never supported enough dedicated research.

People to Follow

Julie Yoo

General partner at Andreessen Horowitz, where she leads healthcare investing. Yoo spent seven years as a software engineer before moving into health tech, then co-founded and served nine years as chief product officer of Kairos Health, a company focused on solving patient access by pooling appointment inventory across the healthcare system, which she and her co-founder sold before she joined a16z. She writes regularly about healthcare investment theses, including pieces on consumers becoming the primary healthcare payer and on AI-native, AI-proof business models.

Notable Quotes

"It's actually designing exactly as planned. If you actually study the innards of how incentives are aligned based on the payment system that exists in healthcare, it results in exactly what we experienced today." (Julie Yoo)

"I do think there will be a lot of roll your own healthcare, N of one solutions that are super personalized to you but still get the benefit of the broader set of intelligence that everybody's contributing back to." (Julie Yoo)