Where clinical AI actually wins, and why almost everyone is building for the wrong bottleneck.
I have spent nearly thirty years operating on children and young adults and more than a decade building and advising AI companies. The two halves of that sentence are usually kept separate. Surgeons are meant to be sceptical of technology and founders tend to be impatient with medicine.
Having lived on both sides, I want to set out where I now believe value will be created in clinical AI over the next decade, and I want to do it in a form specific enough to be wrong.
That last part matters a lot. Most writing about AI in healthcare is unfalsifiable. It gestures at scale, quotes a market size and commits to nothing. I would rather make claims you can check in the future. At the end of this essay, there are five predictions with dates attached. Screenshot them. If they fail, I look forward to hearing about it.
What Viz taught me
In 2016, I was one of the co-founders of Viz.ai. Viz.ai built software that detected large vessel occlusion strokes on CT angiograms and alerted the treating team before the formal radiology read. The company became a unicorn. The algorithms Viz.ai built were only part of the solution.
Detection was necessary and nowhere near sufficient. Plenty of groups had models that could find a blocked vessel. What made Viz.ai work was everything wrapped around the model. Viz.ai chose a single, narrow indication where minutes of delay destroyed brain tissue and protocolised treatment pathways already existed. Viz.ai measured outcomes that clinicians and payers already cared about, such as time to treatment and functional recovery, rather than inventing proxy metrics. And the reimbursement case was designed into the product from the start, which is why Viz.ai became the first AI software to receive a new technology add-on payment from CMS.
Viz.ai is a story about clinical integration, evidence generation and payment architecture. The model was the entry ticket. The company built everything else.
I did not fully understand this at the time. I understand it much better now. I have watched dozens of technically excellent clinical AI companies fail to sell anything, and a smaller number of technically ordinary ones grow more quickly than expected. There are consistent differences between them.
The constraint has moved
Every system has a binding constraint, the single factor that limits output regardless of how much you improve everything else. For clinical AI, the constraint used to be model performance. Getting a network to read a scan as well as a trainee was genuinely hard in 2015. It is much easier now. Foundation models, open weights and commodity tooling have made respectable diagnostic performance available to any competent team. Accuracy has become table stakes.
The constraint has moved downstream, but most of the industry has not moved there with it. The bottleneck today sits in two places.
The first is integration. A model that lives outside the clinical workflow does not exist. If the output arrives in a separate portal, requires a second login or adds a click to a pathway that already takes too long, it will be ignored by exhausted people who are ignoring multiple other alerts. Real integration means the insight arrives inside the systems clinicians already use, at the moment a decision is actually being made, with a clear action attached. This is unglamorous work. It involves HL7 interfaces, information governance committees and the politics of hospital IT departments. It is also where the moat is, because it cannot be replicated by fine-tuning a model over a weekend.
The second is evidence. Outcome evidence as opposed to accuracy studies. The NHS, CMS and every serious payer on earth are converging on the same question: show me the patients who did better and the money that was saved. A ROC curve answers neither of these. Prospective evidence tied to hard endpoints is expensive, slow and requires clinical partnerships that most companies do not know how to build. Which is precisely why the companies that can build them will win. When a capability is scarce and demand for it is structural, it becomes the business.
So here is the first plank of the thesis, stated plainly. Model performance no longer predicts commercial success in clinical AI. Integration depth and evidence quality do. Investors still tend to price the former. The mispricing is the opportunity.
The paediatric anomaly
The second plank concerns where this plays out, and it comes directly from my clinical life.
Paediatrics, and musculoskeletal medicine broadly, are close to invisible in health AI. The capital has gone to radiology triage, adult cardiology, oncology pathways and documentation. Children's MSK health, despite being one of the commonest reasons a child is seen by a doctor, has attracted very little. I have reviewed enough pitch decks and diligence requests over the past few years to say this with confidence.
The standard explanation is that the market is structurally hostile. Buyers are fragmented across community services, district hospitals and a small number of specialist centres. Trials are harder because consent involves families and endpoints may extend beyond growth. Regulators treat children as a special population. Every one of these facts is true. Every one of them is why the opportunity exists.
Structural difficulty is a filter. It keeps out tourists. The teams that can run a prospective study in a paediatric population, navigate consent, hold relationships across fragmented buyers and design for a growing skeleton are rare. Rarity at the point of a real clinical need is where durable companies get built. Conditions like developmental dysplasia of the hip are screening problems with lifelong consequences, the exact type of problem where early detection has provable economic value. Scoliosis surveillance, fracture pathways and gait analysis all share the same structure; high enough prevalence, delayed detection and measurable downstream cost.
There is a second-order effect too. Because so little has been built, the data infrastructure barely exists. Registries in paediatric orthopaedics are patchy, national efforts are underfunded and longitudinal outcome data is scattered across the literature. Whoever builds the data layer for this specialty will sit underneath every application that will be built. Infrastructure businesses are less glamorous than diagnostic ones and considerably more durable.
I am aware of the objection that specialist markets are small. It misunderstands how clinical AI companies grow. Viz didn't start as a platform. It started as one indication done properly, and expanded from a position of proof. Narrow is where you enter, not exit.
The pattern that repeats
Put the two planks together and a playbook emerges. I think it is repeatable, and I think it looks something like this.
Pick one narrow indication where delay or missed diagnosis carries measurable cost. Build for the existing clinical pathway rather than asking workflow to reorganise itself around your product. Generate prospective outcome evidence as a core company function, funded from the first cheque, run with clinical partners who have skin in the game. Design the reimbursement or commissioning case before writing production code, because a product without a payment mechanism is a nice-to-have research project. Then, and only then, expand.
Notice what is absent from that list. There is no requirement for a novel architecture, a proprietary model or a research breakthrough. The defensibility lives in the clinical relationships, the evidence base and the integration depth. These compound with time. Models depreciate.
I also want to name the failure mode I see most often. A strong technical team builds an impressive model, publishes a retrospective validation, wins an innovation award and then discovers that no hospital has a budget line for it, no clinical team has time to use it consistently and no payer has a code for it. Eighteen months later, the company pivots to selling tooling to other AI companies. The founders conclude that healthcare is broken.
What I am doing about it
Three things follow from this thought process.
I will write publicly whenever I can, applying this lens to live developments in clinical AI. Some of it will be wrong. The point of publishing is to be wrong in public, and be willing to learn and re-evaluate.
I will invest my time, effort and capital in companies that fit the pattern. Narrow indication, evidence-led, integration-deep, with a payment thesis the founders can articulate. If you are building in health tech AI and you recognise your company in this essay, my inbox is open.
And I will build in the space myself, on which more to follow.
Five predictions I'm willing to be wrong about
One. By the end of 2028, at least one AI screening product for a paediatric MSK condition will hold both FDA clearance and a defined NHS commissioning route.
Two. By 2029, outcome evidence requirements will be formalised in payer policy on both sides of the Atlantic, meaning retrospective accuracy studies alone will no longer secure reimbursement for new clinical AI. The direction is already visible in CMS's handling of algorithm payments.
Three. At least two of the ten largest clinical AI exits between now and 2030 will be companies whose core asset is a specialty data registry or evidence infrastructure. Neither will have been priced as such at seed.
Four. The majority of clinical AI companies funded at Series A in 2024 and 2025 on the strength of model performance will fail to reach meaningful revenue. Post-mortems will blame the market rather than the absence of an evidence strategy.
Five. Paediatric health AI investment will at least triple as a share of digital health funding by 2030, from its current position of roughly one percent. The firms that entered early will have bought their positions at a structural discount.
The point
Medicine adopted anaesthesia within a decade of its first demonstration because the benefit was undeniable at the bedside. It took the pulse oximeter thirty years to become standard because the benefit, though real, was statistical. Clinical AI thinks it is anaesthesia when it's more like a pulse oximeter. The companies that understand this will spend their capital on proof rather than parameters. They will be the ones still standing when the model layer has become a commodity nobody remembers paying for.
I'm proud to be part of one company that made that bet. The conditions for repeating it are better now than they were in 2016. The corner of medicine I'm clinically closest to is the emptiest part of the map.
That is the thesis. Hold me to it.