Foundation models are reshaping clinical workflows — the vertical thesis, revisited
Eighteen months after the first wave of hospital pilots, the case for verticalization is stronger than we predicted.
Where we were right
Verticalized systems have out-shipped general-purpose deployments. Ambient documentation, prior authorization, and radiology workflow are all now dominated by specialists.
Where we were incomplete
We underestimated how quickly foundation model providers would themselves verticalize with fine-tuned clinical variants. This has compressed some of the moat we assumed for pure application-layer teams.
The renewed thesis
The durable moat is not the model. It is the workflow integration, the evaluation harness, and the clinical trust. All of these compound over time in a way models do not.
Where we are investing
Clinical workflow, evaluation infrastructure, and the multimodal patient record substrate.
The prediction that held
The original thesis argued that generic foundation models would need substantial clinical scaffolding, domain-specific evaluation, and workflow integration to be useful in practice, and that the companies building that scaffolding would capture more value than the model providers themselves. Eighteen months on, that has largely proven true across documentation and administrative workflows, where integration depth predicts customer retention better than any underlying model benchmark.
What we underweighted
We did not anticipate how much clinical trust would depend on explainability at the point of care, not just accuracy in aggregate. Clinicians adopted tools faster when they could see the underlying evidence for a suggestion, even at some cost to raw model performance, which reshaped how the strongest vertical companies designed their interfaces.
The thesis restated
Vertical companies win not by fine-tuning a better model, but by owning the workflow surface area, the data feedback loop from real clinical use, and the trust relationship with clinicians, all of which a horizontal model provider structurally cannot replicate without becoming a vertical company itself. This distinction has only sharpened as base model capability has become increasingly commoditized across providers.
Where capital is concentrating
We are prioritizing teams with a clear plan to close the feedback loop between model output and clinical outcome, since that loop is what compounds into a durable data advantage over time. We are more cautious about teams whose primary asset is a thin interface layer on top of a general-purpose model with no proprietary workflow data underneath.
A pattern across our portfolio
The vertical companies performing best eighteen months in are not the ones that started with the most sophisticated model, but the ones that established a clinical feedback channel earliest, however manual it looked at first, and then automated it as volume justified the investment. That sequencing, workflow trust before technical sophistication, has proven more predictive of durable adoption than any benchmark score we tracked.
The open question for the next cycle
It remains unclear how much of today's vertical advantage survives if base models continue improving at their current pace and begin absorbing more of the clinical reasoning that vertical companies currently handle through custom scaffolding. Our working assumption is that the workflow integration and trust relationship remain defensible even as the underlying model commoditizes further, but this is the assumption we are testing most actively with each new investment.
