Healthcare Never Built Middleware, So It Can Leap Straight to AI-Native
US healthcare spends the smallest share of revenue on technology of any industry and never laid down a SaaS middle layer, so there's no sunk cost to rip out — this is its first organic adoption wave, with doctors adopting technology on their own.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Before LLMs, healthcare's scarcity came from years of training
Julie's causal chain: healthcare is esoteric because it depends heavily on scarce training — to become a specialist provider you need at least seven years of schooling plus training. Before LLMs brought abundant intelligence, the bar to build technology approaching that level of expertise was extremely high. So AI isn't just another efficiency tool; it's the first time ‘replicating expert judgment’ becomes economically feasible. She also points to the payment-side push: the third-party payer system keeps consumers from feeling the pain of each dollar, so the system accumulated far more bloat than free-market dynamics would allow, and both government and payers are now squeezing that cost.
— Julie YooDeductibles turned consumers into actual payers
Employers used to cover most of the premium, so employees felt no financial pain. Then cost shifting began, introducing deductibles — for example, insurance doesn't kick in below $1,000, and you pay it all yourself. Julie says this is exactly the step that made consumers realize for the first time ‘this is the lousy service I'm paying so much for’, and they started voting with their feet and their wallets. The number she cites: there are now companies where you pay $10 a month out of pocket and get excellent service, and the same money buys something 1,000 times worse in the traditional system. This is the direct source of the demand-side adoption wave.
— Julie YooHealthcare never built middleware, so it can skip a grade
Julie says healthcare is the industry with ‘the lowest tech spend as a share of revenue’. Other industries spent decades and tens of billions of dollars laying down a workflow SaaS middle layer, and the moment AI arrives they have to rip and replace and retrain an entire generation of employees. Healthcare only has EHR as its one ERP layer, with everything else propped up by human labor, so it has no sunk-cost bias and can jump straight to agentic form. She calls this the leapfrog dynamic — past technology adoption in healthcare was unnatural (paying doctors to use it, the pandemic forcing telehealth out), and this is the first organic adoption: doctors use AI scribes because they're genuinely useful and because they change the nature of their work.
— Julie YooAI solved ‘knowing’, not ‘then what’
Julie breaks AI's role in healthcare into three stages: the first is lowering the barrier to accessing intelligence, which general-purpose tools have already solved; the second is that once you have the judgment you need to verify it, run clinical-grade tests, prescribe or operate, and have someone actually show up at the door; the third is longitudinal — staying with you until the problem is solved and until the next problem appears. She says today's medical AI is highly transactional, and the real opportunity is in the second and third stages, the ‘last mile’ that health tech builders have to chew through. Her example is portfolio company Council Health: an AI-native doctor's office, 24×7 asynchronous chat, with licensed MDs in the chat room who can prescribe, refer, and diagnose.
— Julie YooSeven years ago, raising consumer health got you thrown out of the room
Julie says seven years ago, mentioning consumer health in front of any investor got you kicked out of the room, because there was no viable business model, consumers had no agency, and no payment model existed. That's changed: the fastest-growing companies in health tech are cash pay — direct out-of-pocket, with low price disruption. She thinks the biggest lever is cost structure, and AI makes delivering a medically credible service about 100 times cheaper than before. The product design order has flipped too: it used to be design for insurers and doctors first, with consumers bolted on afterward; now the consumer is at the top of the stack. She wrote an article about this called consumers are the new payer.
— Julie YooThe best companies have to be both AI-native and AI-proof
Julie's article last week proposed AI native and AI proof: LLMs can answer questions you used to have to wait six months to see a specialist to ask, but the rest — making house calls for people who can't get to a hospital, drawing blood and actually diagnosing what you have, doing the things that in the US only regulated entities can legally do — AI itself can't solve. So she points the opportunity at full-stack challenger companies: they look like a retailer or a service provider on the outside, are highly AI-native on the inside, and therefore have surface area and a disruptive cost structure that were historically unattainable, with high margins that can feed back into innovation. Hardware and robotics are in scope too; she thinks robots have already found product-market fit in high-acuity healthcare settings.
— Julie YooThe healthcare payment system isn't broken, it runs as designed
Faced with the complaint ‘how did healthcare get so bad’, Julie's answer is: it's actually exactly as planned — if you study how incentives are aligned under the payment system, today's outcome is its inevitable product. So the most ambitious companies are asking: can we blow up the existing payment model and build a new one from scratch, essentially building a mini healthcare system inside one company. She names Devoted Health as doing this, and says it's possible now because the cost of entry to build a company has dropped (the infrastructure already exists), AI can do more, and appetite on both the consumer side and the traditional ecosystem side is unprecedented.
— Julie YooThe dataset to train medical-grade AI doesn't actually exist
Julie's judgment: we don't even have the dataset needed to train medical-grade AI. Because today it depends heavily on EHR data, and EHR is sporadic — you see a traditional doctor at most once a year over a lifetime, so the complete narrative arc of you as a patient is almost entirely absent from every existing dataset. Her inference is N of one health: lots of companies will run ultra-personalized experiments on you alone, and those experiments generate new data tracks that then drive the next generation of models. She also gives a ten-year prediction: every one of us will have a lifelong AI doctor in our pocket.
— Julie YooIn their own words · checked verbatim
Healthcare for better, for worse, did not spend that amount of money, right? So we only had really like the ERP layer with EHRs, and then labor, like we were just throwing bodies at every problem.
Julie Yoo13:22
we have less of a sunk cost bias as an industry to say, okay, this new technology paradigm has come along, and we can just leap right to it.
Julie Yoo14:25
This is currently is the only like the first real organic adoption wave that we've seen in health tech where doctors are just using AI scribes because they freaking work and they're so good and they really change the nature of their job for the positive.
Julie Yoo15:29
I'm like, it's actually designing exactly as planned. Like 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 experience today.
Julie Yoo23:46
I would argue that we actually don't even have the datasets necessary to truly train like a medical grade AI because we rely heavily today on electronic health record data, which is really sporadic, right?
Julie Yoo24:52
So I think a lot of companies will start doing these kind of N of one experiments that generate the new data rails that then drive the next generation of models that are built in our space.
Julie Yoo25:52
Figures
| Julie Yoo's years in healthcare | about 18 years | 1:00 |
| Years she worked as a software engineer | 7 years | 1:00 |
| Time for full EHR deployment across US doctors | about 5-7 years | 2:03 |
| Minimum years of schooling plus training to become a specialist provider | at least 7 years | 11:20 |
| Deductible example: out-of-pocket ceiling before insurance kicks in | you pay it all below $1,000 | 11:20 |
| Monthly fee for consumer subscription healthcare | $10/month | 12:22 |
| Factor by which AI cuts the cost of delivering a medically credible service | about 100 times | 20:37 |
| Gap in service bought for the same spend under the traditional system | 1,000 times | 12:22 |
| Out-of-pocket cost of going to the emergency room | $2,000 | 16:31 |
Glossary
- leapfrog dynamic
- Never having laid down the previous generation of middleware, so you can skip it and adopt the new technology paradigm directly.
- AI native and AI proof
- Julie's framework: use AI to restructure the cost structure while also doing the offline and compliance work that AI itself can't do.
- N of one
- Running ultra-personalized analysis on a single individual, generating new data tracks that feed back into models.
- AI scribes
- Tools that automatically record doctor-patient conversations and generate medical records; the entry-point scenario for doctors adopting AI on their own.
- third-party payer system
- Medical costs are paid mainly by insurers or employers rather than patients, causing a lack of price sensitivity and system bloat.
How to listen
Founders building healthcare AI or digital health, investors looking at the healthcare sector, and engineers who want to know which vertical AI cracks the business model first.
The personal history and industry-history setup from 0:00-8:16; you can start listening at 8:16.