Owkin

The model is the product. Owkin sells an agentic research environment whose core is a reasoning model, Owkin Zero, with an orchestrator that selects and combines specialized biological AI skills and calls modality specific models underneath: K Pro for biopharma decision making and K Navigator for academic researchers.

The answers it produces come from the models querying a patient data network and a literature corpus; remove them and what is left is a database a researcher would have to search by hand. The company's stated direction, an autonomous AI scientist, points further the same way rather than toward a platform whose value stands without the models.

Oversight is described through its mechanisms, while the threshold is missing. What is published is specific: K Pro's reasoning is made visible to the user, explainability features log every data source, model decision and rationale so the chain is auditable, results are grounded in queried patient data and in named literature sources with provenance a user can follow back, retrieval is used to verify that cited literature identifiers exist, and the calls the orchestrator makes to underlying tools are monitored. Access is controlled by role, usage is monitored, partner institutions must accept an ethical usage policy, and the company states it will suspend access where misuse is detected.

The one universal claim in its published material, that every result is grounded in evidence a user can trace, is about the basis of outputs rather than their correctness, and the mechanism that makes it checkable is published. What is not published is any threshold: no confidence measure with a point at which K Pro declines to answer, no error rate, and no statement of what the autopilot mode may do without a person.

The approach and the data foundations are described, without the version and update discipline behind them. Owkin names its own reasoning model, Owkin Zero, and describes the architecture concretely: a single orchestrator selects and combines specialized skills and tools, with modality specific models called underneath and their calls monitored. The corpus is enumerated by source, spanning 19 biomedical databases, 26.5 million articles and spatial multiomic patient data from the MOSAIC initiative, and the related entity structure including the Bioptimus foundation model is disclosed rather than obscured.

What is not published is version discipline: no model version a customer can cite, no release notes, and no statement of how a customer learns that the orchestrator or an underlying model changed. The FAQ acknowledges that outcomes depend on which large language model is in use and that switching requires prompt and context changes assessed through internal evaluations, which makes the version question live rather than academic, and the third party models themselves are not named.

Substantial partial disclosure, with the outside model layer unnamed. Owkin names its own reasoning model, describes the orchestrator and the specialized skills it calls, enumerates the data foundations by source (19 biomedical databases, 26.5 million articles, MOSAIC spatial multiomic data), and discloses the related entities, including the foundation model company it incubated. On terms, the FAQ states that vendors are vetted through security assessments and contractual requirements aligned with the company's own commitments, that internal and external audits are run, and that where large language models or other third party components are integrated the company chooses hosting options it says guarantee privacy and confidentiality for all data and outputs.

The party that answer turns on is not named: which large language models are used, and who hosts them. No subprocessor list is published.

Named customers and case studies, with the headline figure unevidenced and no clinical candidate from the platform. Collaborations are named and substantial: AstraZeneca, including a three year licensing agreement to build biopharma agents, Bristol Myers Squibb and Sanofi, with case studies published on multiomic reports, new biology, new molecules, autopilot mode and clinical research support. A separate validation program is described, covering laboratory validation and what the company calls patient validation, and the group publishes open access preprints and datasets.

The claim that sets the bar is the company's own: target identification compressed from more than 12 months to roughly 3. No method, denominator or baseline comparison is published for it, and the case studies are the company's accounts of its own engagements rather than customer reported outcomes with figures. No molecule attributed to this platform is reported in a clinical study; the affiliate's trial is a separate program.

Categorical commitments are published, without the retention schedule behind them. The company states that de-identification happens at the source institution before data reaches it, and it describes federated learning as the approach it pioneered so that models travel to data rather than data to models. On the product, the FAQ says that for each K Pro user only the chat history is stored, always bound to that user and their organization, in a managed database on Owkin's own account reachable only by authorized staff, and that where outside language models or other third party components are integrated the company selects hosting options it says guarantee privacy and confidentiality for all data and outputs. A patient information page addresses the people whose data underlies the platform directly.

No retention period is published for chat history, prompts or outputs, no statement says whether customer prompts or uploaded material are used to improve the models, and the de-identification standard applied at source is not named. Safety engineering is not addressed.

A status is claimed and the instrument behind it is not published. The K Pro FAQ states that the company's privacy first approach ensures full compliance with GDPR and HIPAA requirements, and points to the controls around it: ISO 27001:2022 certification, vetted vendors under contractual requirements, internal and external audits, and hosting choices for third party components said to guarantee confidentiality of data and outputs. Separately, a patient information page sets out the European, United Kingdom and Swiss data protection regimes the company holds itself to for the patient data behind the platform, with de-identification performed at the source institution.

There is no published statement of business associate status, no agreement or its tier scope, and nothing on the circumstances in which Owkin would act as a business associate at all, which matters for a platform built on hospital sourced patient data and sold to biopharma.

Recognized certifications are named with their type and version in the company's own material, without the artifact. The K Pro FAQ states that Owkin is certified to ISO 27001:2022 for information security and ISO 13485:2016 for medical device quality, and describes the program around them: organizational and technical measures, regular internal and external audits, security risk assessments with every major change, and vendor due diligence with security assessments and contractual requirements before a partner is used. A trust center exists on a hosted compliance portal and may hold the underlying documents.

The site carries no certificate, certifying body, scope statement or expiry date, and the scope question matters for a group spanning a software platform, a diagnostics spin out and a foundation model company.

The software carries no device claim, and the product is scoped accordingly. K Pro and K Navigator are research and decision support tools for biopharma R&D and academic researchers; nothing presents them as regulated devices. The company states it is certified to ISO 13485:2016 for medical device quality, which governs how a manufacturer runs its quality system rather than what this product is. Its CE marked in vitro diagnostics sit with the spun out diagnostics company. An affiliate sponsors a clinical trial of a drug candidate, whose regulatory position belongs to that program.

The tension sits in the marketing rather than the paperwork: the company describes an autonomous AI scientist and a future biological artificial superintelligence, while what is sold today informs decisions that people make.

A governance framework with named process behind it, and no evaluation results. Owkin publishes an AI ethics page describing a dedicated AI safety and ethics team and what it works on: mindful deployment into healthcare systems, emerging threats to patient privacy, decisions during model building about what is and is not included and how bias is minimized, validation that accounts for heterogeneity and brings people back into the loop, and algorithmically enhanced decision making. The stated approach is to co-design with researchers and academics, include patients and clinicians in development, anticipate harms and re-evaluate products regularly, with ethical review and risk assessment described as part of the development process and an open science program publishing datasets, models and preprints.

None of that is a checkable evaluation. There is no bias or fairness study with a stated method, no results by patient subgroup for the models the platform runs, and no model card, even though the data foundation is drawn from particular hospitals and cohorts where representativeness is plainly at issue.

Mechanisms let a user check an output, with nothing standing behind it. The company publishes what it does to make results traceable: the model's reasoning is exposed to the user, explainability features log every data source, model decision and rationale, every result is presented with its provenance back to the source, literature answers are checked by retrieval so that cited identifiers exist, and tool calls are monitored. A dedicated patient information page sets out the data protection regimes the company holds itself to and how the people whose data underlies the platform can exercise their rights, and de-identification happens at the source institution.

Those rights reach the personal data the company processes, not the output of the system: a researcher acting on a wrong target hypothesis has no route that asks anything of Owkin, and no accuracy figure, warranty, indemnity or remediation commitment is published.

Integration is described at the infrastructure layer with no record system named. The platform does reach patient data: K Pro is built on multimodal patient data assembled with hospital and research partners, including the MOSAIC spatial multiomic oncology dataset, and Owkin runs a data partnership program. How that data arrives is described as a research pipeline with de-identification at the source institution rather than as a connection to a record system.

On the customer side the published integration is a deployment rather than an interface: K Pro is installed into a dedicated account inside the customer's own AWS organization, with a least privilege role and network egress to named dependencies. No electronic health record, clinical data warehouse, laboratory system or interoperability standard is named, and no production integration is cited.

Options and isolation are documented, with the processing path left open. The enterprise deployment is described step by step: K Pro is installed into a dedicated AWS account that forms part of the customer's own AWS organization, the Owkin team assumes a role the customer provisions in advance which follows least privilege and is reviewed by the customer's security team before deployment begins, and the customer allows egress to named dependency types. That puts the workload in the customer's own cloud, which answers both the tenancy and the location question for the deployed component. The academic environment is a hosted service instead.

What is not stated is the rest of the path: chat history is kept in a managed database on Owkin's own AWS account rather than the customer's, the outside language models the orchestrator calls are unnamed along with where they run, and no statement says which processing crosses a border for a company operating from Paris, New York and Geneva.

No price is published, and the commercial structure is discoverable in detail. Two tiers are described plainly: K Navigator is free to academic researchers, in beta with a waitlist and a stated evaluation period, and K Pro is an enterprise product sold through a demo and a license, with a three year AstraZeneca agreement named as an example and a distribution partner program published for organizations that want to resell it. The company also publishes what an engagement looks like before contact, including strategic working sessions run in house with pharma teams and trial periods of stated length.

No figure, unit of charge or tier structure is published for the enterprise product, so a buyer can size the shape of the relationship and not its cost.

Coverage is named across the research workflow, with validation behind part of it. The published use cases are specific: early portfolio decisions, patient and population decisions, and clinical trial decisions, supported by skills the company lists for data exploration, hypothesis generation, biomarker discovery and trial design, with two environments serving different users, academic researchers and biopharma teams. The data foundation is deepest in oncology, where the MOSAIC spatial multiomic dataset sits, and the case studies published are drawn from that work.

Beyond oncology the coverage is stated rather than demonstrated, and no performance is published by disease area or by data type.