There is a specific kind of anxiety that comes from being told to trust a machine you cannot interrogate. In most software, this is annoying. In medicine, it is a liability, a lawsuit, and occasionally a tragedy. Evidium, a San Francisco company founded in 2023 by Carl Bate, has decided that the anxiety is the whole business - and that the way to sell healthcare AI is to build the kind that can be cross-examined.
The company's coinage for this is "Referenced AI," which is a small phrase doing a great deal of work. It means that when the software tells a clinician something, or tells an insurer that a patient is likely to get sicker and more expensive, it also hands over the receipts: the guideline, the chart, the paper, the chain of reasoning. You can, in theory, follow the answer back to where it came from. This is unglamorous and, in the current market for large language models that confidently invent citations, quietly radical.
The technical approach has a name that sounds academic until money is on the line: neurosymbolic AI. The idea is to marry two traditions that have historically sniffed at each other. On one side, the generative and neural models that are good at reading messy human text - clinical charts, medical guidelines, research papers - and turning it into structured data. On the other, the older, stubborn machinery of ontologies, knowledge graphs, and symbolic logic, which is good at reasoning in ways you can trace and control. Evidium's bet is that the combination gives you both the fluency of the new stuff and the auditability of the old.
What Evidium says it is really building, underneath the products, is something closer to a "world model" for healthcare. The company describes projecting a patient across a computable map of clinical states, the transitions between those states, and what causes a patient to move from one to another. If that sounds abstract, the commercial version is not: it is the difference between "the model thinks this patient is high-risk" and "here is the state this patient is in, here is where patients like them tend to go next, and here is why."
The products that sit on top of this shared foundation split neatly along the fault line that runs through American healthcare - the one between the people who deliver care and the people who pay for it. For clinicians and care teams, Evidium sells quality insights with reasoning they can audit. For the risk-bearing organizations - insurers, health plans, self-insured employers, and the integrated delivery networks that the industry has taken to calling "payviders" - it sells probabilistic forecasting of disease states, care, and cost. One knowledge base, two audiences who rarely agree on anything, connected by a model that is meant to be explainable enough that both sides can argue with it on the same terms.
Bate, the founder and CEO, came to this from the advisory side. Before Evidium he spent years in consulting on artificial intelligence and digital health, which is a useful place to notice a problem repeatedly without being able to fix it. His framing of the core issue is admirably plain: "Healthcare runs on knowledge, but until now, much of that knowledge has been locked in text, statistics, and silos." The company he built is, essentially, an argument that the unlocking is the opportunity - and that doing it without losing the thread of why any given fact is true is the hard part everyone else skips.
It is a small company - roughly 33 people - but the composition is the tell. Alongside the machine-learning research engineers sit practicing physicians and clinical scientists: a chief medical officer, a chief scientific advisor, clinical knowledge engineers. In an industry where "we hired a doctor as an advisor" is often a marketing line, Evidium treats domain expertise as core infrastructure. Models are built next to the people who will have to read them.
