The AI is already in the building. Nobody approved it.
Clinicians use AI at work every day. Most of them have no tool their institution sanctioned. We are building the one that can be.
Two health systems. One pattern.
Most clinicians already use AI at work. Far fewer have a tool their institution sanctioned. The gap is not policy; it is which tool a clinician actually opens. Where a clinical-grade default exists, it wins from the bottom up. In Europe nobody has built one.
81%
United States
use AI at work, 40% have a sanctioned clinical tool
69%
United Kingdom
use AI at work, 15% have a sanctioned clinical tool
AMA Augmented Intelligence Research 2026 · Royal College of Physicians 2025
A grade has to show its working.
Medical AI tools have started putting a letter next to the sources they cite. A letter only means something when the method behind it is written down and every grade can be checked against its source. So we built an engine that grades what it reads, publishes how the grade is worked out and links every claim to its record: a controlled library, a grading layer and the right to refuse.
Why this is possible now.
The models
Since 2024, language models read a study about as well as a trained assistant. Two years ago this project was not possible.
The law
The EU AI Act puts clinical decision support under real obligations, and inspectable reasoning stops being a feature and becomes the licence to operate.
The flood
Two million papers a year, and retractions at a record. The cost of not grading grows every month.
How we decide what ships
Cite everything
Every claim links to the study or label it rests on. If we can't source it, we don't say it.
Refuse over guess
When the evidence can't support a safe answer, we say so. A clear no beats a confident maybe.
Say how strong it is
A weak study is graded weak. We'd rather lose a little shine than overstate what the evidence shows.
The doctor decides
The engine reads, grades and cites. It never acts, and it never signs anything off. The clinician stays responsible, and the product is built to keep it that way.
Privacy by default
Our logs hold request metadata, never the question or the answer, and patient data never lands in our access logs. The text of a checked claim is kept 45 days for review, then deleted.
Never for placement
No commercial interest belongs inside a prescribing decision. Nothing in the product is sponsored, promoted or placed.
Supported by the University of Potsdam.
Novo Lab is supported by the Startup Service of Potsdam Transfer, the University of Potsdam’s centre for start-ups, innovation and the transfer of knowledge and technology. It advises founding teams on the way from research to a company.
About the Startup ServiceWant to talk?
Access, partnerships or the science behind the engine. Write to us and a person answers.