Why the next generation of medical AI companies will look nothing like the last
A synthesis of what we have learned over two years of building, investing in, and diligencing clinical AI companies.
The old shape
The 2021 vintage of medical AI companies looked like traditional software — SaaS pricing, feature roadmaps, standalone applications. That shape is quietly being retired.
The new shape
The 2026 vintage looks more like an AI-native operating layer — deeply integrated with the EHR, priced by outcome or transaction, and shipped alongside a clinical service. The line between software vendor and clinical operator is blurring.
Implications for founders
You will need a clinical operating capability, not just an engineering capability. Plan for it from day one — the retrofits are painful.
Implications for investors
The next generation of durable healthcare franchises will look more like hybrid software-and-service businesses than pure SaaS. Underwriting frameworks need to evolve accordingly.
Why the model layer stopped being the moat
Two years ago a company could raise a seed round on a fine-tuned model and a promising benchmark. That is no longer sufficient, because the underlying foundation models improved faster than any single team's fine-tuning effort, and the benchmark gains got arbitraged away within a quarter. The durable value moved to workflow integration, data access, and the operational trust built with clinical staff.
We have watched several teams rebuild their pitch mid-diligence once they realized this. The ones who adapted stopped describing themselves as model companies and started describing themselves as the system of record for a specific clinical decision. That reframing changed everything downstream, from hiring to go-to-market sequencing.
An illustrative pattern from diligence
One company we evaluated had genuinely superior model accuracy on a narrow diagnostic task, verified against a modest internal validation set. It struggled anyway, because the customer's actual bottleneck was not accuracy but the number of clicks required to act on the output inside an existing worklist. A competitor with a mediocre model but a seamless embed won the deployment.
That outcome is common enough that we now weight workflow fit above model performance in early screens, reversing our own priors from a few years back.
The counterargument worth taking seriously
Some operators argue this framing understates the value of genuine clinical accuracy gaps, particularly in specialties where error costs are severe and margins for workflow shortcuts are thin. In those pockets, a materially better model still commands premium pricing and faster adoption, and workflow elegance is a secondary concern to the buying committee.
We think both things are true simultaneously, and the skill is in knowing which regime a given clinical problem sits in before building the go-to-market plan around it.
What we tell founders raising right now
Do not lead a pitch with model performance unless the clinical error cost is unambiguously catastrophic and your accuracy edge is durable against a fast-moving frontier. Lead instead with the specific workflow you own end to end, the data feedback loop that compounds, and the operational relationships that took real time to build. Those are harder to displace and easier for investors to underwrite.
The diligence checklist we now run
We now spend the first diligence call asking a founder to walk through the exact clicks a clinician takes from alert to action, rather than asking about model architecture. Teams that cannot answer in concrete detail usually have not spent enough time in the actual workflow, and that gap tends to show up later as a stalled pilot.
