Physix Frontier
Owkin's AI scientist isn't as magical as it seems
I compared Owkin K Pro with general-purpose LLMs for breaking down biomedical problems and actually ran it through the paces. K Pro is part of Owkin's so-called "AI Scientist" platform; simply put, it's an agentic R&D assistant: you give it a goal, and it breaks the task down into literature review, target identification, patient subgroup design, and evidence chain listing. Another selling point behind it is a multimodal patient data network—basically connecting different data sources like imaging, pathology, genetics, and medical records. Sanofi gave K Pro a five-year license, with Owkin handling end-to-end development. AstraZeneca was already on board, and Boehringer Ingelheim signed this week, showing that big pharma isn't just watching from the sidelines.
On day one, I didn't dare throw sensitive data at it, so I only input a few public target questions and a de-identified patient cohort description. The general model's answers were smooth, reading like popular science articles; K Pro felt more like a ticketing system, separating "finding evidence," "listing hypotheses," and "marking gaps." The surprise was that it proactively pointed out which steps lacked experimental validation instead of forcing a conclusion. The bottleneck was speed—it took about twenty minutes per task on my end, with pauses waiting for tool calls. From a deployment perspective, this doesn't feel like chatting; it feels like having a junior researcher write a plan first.
Day 3 to One Week Later
By day three, I switched to a narrower task: given a tumor indication, have it organize patient subgroup hypotheses. The path it provided was clear, but the terminology density was high, making it basically unreadable for non-medical backgrounds. My testing showed its real value isn't in "writing well" but in "knowing where to break things down." However, the downsides are obvious: the display of evidence sources isn't detailed enough, and some conclusions I can only treat as hypotheses since I can't verify them on the spot. A week later, I threw the same task at a general model; it was faster but tended to conflate correlation with causation. K Pro is more conservative, suitable for initial screening but not for making final decisions.
Sanofi and Owkin started with a €90 million collaboration in 2021, later granting K Pro a five-year license, aiming to integrate AI agents into drug discovery.
I think this thing suits two types of people: pharma R&D teams with professional validation capabilities and investment due diligence experts; it's not for novices who treat AI as an answer machine. It's like an early demo of Robotaxi—it can run the process, but it's far from being fully autonomous without human takeover.
Real-world data suggests that Owkin K Pro is worth piloting on a small scale for professional teams but isn't suitable for directly replacing researchers.
📌 This article is compiled from Bloomberg Tech. Original text: https://www.bloomberg.com/news/articles/2026-09-02/boehringer-signs-deal-for-owkin-ai-joining-astra-and-sanofi
All rights reserved by the original authors. This is a compilation and independent analysis based on public reports.
