Will Hu

The through-line

Will Hu's consistent argument is that healthcare AI is not a new arrival and that the binding constraint on it has never been the models. It has been workflow, economics and policy. He dates the current wave back roughly two decades to the government push that digitised records, and describes everything since as one continuous build-up rather than a sudden break: "the chat GPT aha moment kind of like hits all of us but it has been building I would say over the past two two and a half decades" [1]. What has actually changed, in his telling, is the business shape of the thing. The industry has moved off point solutions sold into IT budgets and toward integrated, service-oriented offerings that sit inside how providers already work [1][3][4][5].

The second half of the through-line is that the technology is now ahead of the rules that pay for it. He builds on the assumption that AI can already augment or take over meaningful tasks, and that it will do more over five to ten years, but insists the reimbursement side has to move before any of that reaches patients at scale: "policy has to change and they're so far behind" [1]. Throughout, his framing is augmentation rather than displacement, with AI expanding what providers can deliver rather than removing them [1][3][4][5].

On the two decades that got us here

Hu traces the arc from government mandates, encouragement and penalties around electronic health record adoption, which turned visits, vitals and reports into "a very data Rich environment" that analytical and AI systems could in principle draw on, using Epic and its patient-facing my chart as the familiar example [1]. The early 2000s attempts failed, he argues, because the underlying technology "really hasn't changed since the 1960s 1970s" in terms of available models, so decision support systems were tried and abandoned, leaving investors with a lasting bad taste [1]. The adoption failure had a workflow cause as much as a technical one: "doctors don't like another portal to be open," and when too much information was pushed back into the EHR through advisory mechanisms, "people would just ignore it" [1].

Maturing deep learning in the 2010s produced ambient AI scribing, capturing the conversation in the room to relieve pressure on doctors and improve billing, with generative AI powered scribing arriving toward the end of that period [1]. The present phase, as he describes it, is AI-enabled services and agents, to the point that a patient may get a phone call to schedule an appointment "and that would be Ai and then you wouldn't even know that that's there" [1][3][4][5].

On why healthcare is a service business, not a SaaS market

The mistake Hu keeps naming is applying SaaS instincts to healthcare: build one solution, sell the identical thing everywhere, scale. That fails because "Health Care is first and foremost a service business" and because every setting and every provider is unique [1]. His prescription is to dogfood your own services first, connect the dots inside a practice, learn how they actually do things, and then work out which point solutions can be combined into a single package [1]. Two decades of trying to sell AI to doctors have shown how hard it is to make money that way: sales cycles into hospitals and provider groups are long and willingness to adopt is low [1].

The bundling move is also an economic one. Because agents can now hold empathetic, interactive conversations with patients, a company can package healthcare AI with other tech-enabled services, which changes the buyer: "instead of just selling to an IT budget" you sell into the operational budget of a service business, which is a far larger addressable market [1]. The design principle that follows is stated plainly: "we want to provide the best solution for the providers but we don't want to change your workflow" [1]. On the technology side that means abstracting everything away; on the services side it means taking over work that is inherently hard for people because it is too laborious or too difficult to track, delivering the outcome as a service rather than "giving you a software and hope for the best" [1][3][4][5].

On timing, disease progression and the right patient at the right time

Hu's own research background sits in disease detection and disease progression analytics, and he is candid that the field's earlier ambition, an all-capable AI doctor able to diagnose anything or predict breast cancer five years out, was largely funded by pharma companies doing drug targeting and hunting for "Opportunity patients," with oncology drugs the big spenders [1]. What he took from that work is that the hard question is not diagnosis in the abstract but timing: "it really is about timing," because you want a system that "sees everything the doctor see and evaluate it the way the doctor do," asking what your current concern is, what happened last month, over six months, over the past year, and how you responded to different treatment regimens over time [1].

That produced a temporal recommendation approach to disease progression. It works, he argues, for conditions with mechanically or linearly predictable trends, ageing, injury, musculoskeletal, rheumatology, cardiology and diabetes, in contrast to something like sepsis where a patient does well for a long time and then suddenly trends downward [1]. The value is triage at population scale, because "doctors have too many patients and they can't see all of them" and patients get lost in the cycles and journeys; the system's job is to surface the most severe and "who's the most undertreated patient" and get them treated faster [1]. His worked example is asthma: around 25 million patients a year in the US, roughly 10% at a moderate level requiring steroids, and 10% of those progressing to severe asthma needing biologics. The typical path wastes years cycling through inhalers and then steroids across different doctors, when failure to respond to two different steroids should trigger immediate specialist referral for biologics [1].

On reimbursement as the real bottleneck

Hu's strongest position is that payment models, not algorithms, decide whether healthcare AI reaches patients. Existing codes for follow-up visits or remote telehealth visits are time based, designed in 20, 30 and 40 minute increments on the assumption that a human is manually reading records, writing reports, uploading them and deciding what to say. An AI-enabled platform does not need that time to deliver equal or better care, so the payer side has to adjust [1]. He extends this to a thought experiment about an AI therapist delivering care at scale with humans present largely as a safeguard or for hands-on needs: if the care is good and cheap but insurers do not reimburse its use and development, "then no one would do it" [1].

The one mechanism he can point to is new technology add-on payments, which require certification of novel technology and in practice suit radiology startups best, since an algorithm doing a specific thing fits the criteria [1]. He is scathing about the incentives it creates: a hospital with the right software installed gets paid around a thousand dollars every time an MRI image passes through, which "is kind of like it's robbing the bank," costs health systems a lot of money, favours inpatient hospital utilisation for imaging that could be done anywhere, and pays out even where the doctor disagreed with the AI read [1]. Alongside this, Medicare rules restricting time-based care to US-based personnel mean that even where AI could perform a task perfectly today, nobody would, because it would not be paid for [1]. His summary is that the current model either creates adverse incentives against adoption "or it just straight up prevents it" [1][3][4][5].

On policy as the shaper of healthcare business models

More broadly, Hu reads the biggest healthcare companies and trends of the past decade as policy artefacts. The US being the only country permitting direct-to-consumer advertising is, in his account, what created an entire industry of data aggregators, analytics firms, consultancies and privacy-adjacent companies working out who writes the most scripts for a drug and who is inclined to, extending into credit agencies supplying consumer data on what you read and the best channel to reach you, and into intake companies that give the product away and monetise eyeballs during intake [1]. Similarly, the Affordable Care Act produced the marketplaces and the crop of Medicare Advantage plans, with Medicaid HMO plans now moving in a similar direction [1].

In his own line of work the policy surface is integration and privacy: HIPAA, HITECH, TEFCA and care interoperability networks. He regards much of it as redundant compliance work rather than a genuinely high security bar, but work you must nonetheless do [1].

On integration economics and the business in a box

Hu sees the crowd of point solutions selling into the same practice as a straightforward waste, because they all depend on the same delivery mechanism and the same write-back integrations to put information into the system [1]. His alternative is to ingest all of a practice's data, train a centralised model on it collectively, and deploy it in a customised way per practice, spreading training cost across practices [1]. The same logic applies downstream: old-school HL7 integration runs roughly five to seven thousand dollars per interface, an application typically needs three or four interfaces, and there is a 20% annual maintenance charge on top, but if the same structure carries four or five solutions to the provider, the cost does not multiply [1].

That is the "business in a Box" model, where one vendor replaces what used to be several point solutions, integrates once, and bolts services on top [1]. The payoff he emphasises is capability that small practices could never build alone: a group with no in-house physical therapists cannot offer remote telehealth-based physical therapy guidance or keep track of the compliance requirements, but can access it through an enterprise-grade platform deployed to many small providers, "so they have access to what used to be only available to the hospital systems" [1].

On what comes next

What excites Hu is the convergence of two data streams that are currently separate: care delivery and data capture inside the visit, via ambient scribes recording everything said in the room, and the digital front door that interacts with patients while they are outside the practice [1]. Combining the two is the route to genuinely holistic care [3][4][5]. He holds this alongside his consistent view that the point of the technology is to strengthen providers' ability to deliver better patient care rather than to replace them, and that the net effect on employment in the field will be positive [1][3][4][5].

Takeaways

• Healthcare AI's failures of the 2000s were adoption failures as much as technical ones: doctors resist another portal, and information pushed back into the EHR in volume simply gets ignored [1].
• Build for the practice as it is: the operating principle is to give providers the best solution without changing their workflow, abstracting the technology away entirely [1].
• Selling AI as a bundled service moves the purchase from the IT budget to the operational budget, which widens the addressable market considerably [1].
• Progression modelling should be temporal, evaluating what happened to a patient over months and years and how they responded to prior regimens, in order to find the most severe and most undertreated patients [1].
• Time-based billing codes assume a human doing the work manually and therefore price AI-enabled care out of existence; new technology add-on payments largely reward radiology and pay out even when the clinician disagrees with the AI read [1].
• Integration is a fixed cost worth amortising: HL7 interfaces run roughly five to seven thousand dollars each, three or four are typically needed, plus 20% annual maintenance, and one integration can carry several solutions [1].
• The near-term prize is combining ambient capture inside the visit with patient interaction outside the clinic, with AI augmenting rather than replacing healthcare workers [1][3][4][5].