Causaly

Causaly's central product is learned scientific research. Agents retrieve biomedical evidence, interpret relationships and produce research assessments using a knowledge graph and specialized retrieval system. Scientific Workflows adds repeatable methods and deliverables around that reasoning, while Bio Graph and Pipeline Graph support biological and competitive analysis. The software can also apply methods to data and preserve code and parameters for review.

Human scientists help customers configure the work, but the commercial offer is a software platform with its own inference and evidence extraction capabilities. It is more than access to a database or a service that merely hosts another organization's AI tools.

Deep Research uses a Research Agent to collect evidence and a Principal Investigator Agent to critique it. Missing sample sizes, inconsistent intervals or inadequate coverage can trigger another research pass before a report reaches the user. Scientists can inspect the resulting evidence and assumptions, but the final human approval rules remain the customer's responsibility. The copilot feature comparison also asserts validity of all results without a qualification in that statement.

That is a claim about output correctness, separate from the limited accuracy figures elsewhere. It conflicts with the service terms' warning that generated findings can contain material errors despite appearing detailed and credible.

The Scientific Information Retrieval System combines semantic similarity with graph search and ranking. Biomedical ontologies help connect queries to directional relationships, including evidence that refutes a proposed association. Generative components then synthesize retrieved material, while separate research and reviewer agents organize longer investigations. This provides an identifiable technical approach and a description of the evidence base.

The copilot architecture permits language models to be exchanged, but the current product material does not identify the deployed model suppliers or versions. No release schedule ties a reported benchmark to the precise configuration a customer receives, so retrieval design is more transparent than the generative engine.

The public processing addendum names Google Cloud and AWS as hosts, plus authentication, analytics, communications and customer support suppliers. It binds subprocessors and provides a process for changes and objections. The acceptable use policy separately acknowledges upstream AI models, and the copilot architecture allows language models to be swapped. Neither disclosure identifies the companies providing that generative layer.

Microsoft's role as an integration partner does not establish which model Causaly itself runs. The hosting chain is partly visible, but the recipients, retention terms and training conditions for each model request are not mapped to the individual and enterprise product offers.

Eisai's June 2026 case reports roughly 53% to 59% less time on literature review, although the results table starts at 52%. The account describes oncology and neurology workflows without a sample size, timing protocol or controlled comparator. ProQR's February 2025 case reports meeting its 2024 target identification goal by the third quarter and greater productivity than PubMed, again without enough measurement detail to reproduce the result.

Agentic Research advertises 88% precision on LitQA2 and 93% accuracy on biomedical claims. The named benchmark helps identify the task, but the product page does not specify the evaluated model version or test execution.

The enterprise data fabric carries a no training commitment for client data. Causaly Research's newer terms distinguish uploaded customer content from prompts, usage and personalization information; protection for performance data is broader on the Team plan than on individual subscriptions. The processing addendum sets a termination return or deletion process within 90 days, subject to legal retention. This does not provide a complete routine schedule for prompts or model provider copies.

Research agents use citations and review cycles to reduce unsupported conclusions. A full incident procedure for propagating corrections after a wrong scientific answer remains unspecified, even though general security incident obligations are described.

The UK privacy framework reaches Causaly's platform and identifies its controller and processor roles. A public data processing addendum binds subprocessors, addresses international transfers and assigns responsibility for incidents and customer instructions. The current acceptable use policy excludes sensitive categories of personal data unless a separate written amendment permits them.

This is a defined research data boundary rather than an assumption that all scientific content is anonymous. A US BAA is not part of the public contracting set. The older enterprise terms exclude personal data beyond account details, while the newer Research terms allow personal data under a valid processing agreement.

Causaly announced ISO 27001 certification in June 2022, describing an information security management system covering customer information. The current processing addendum permits customer audit requests to be satisfied through relevant independent reports or certifications. Its references to SOC 2 and ISO are examples of acceptable artifacts, rather than proof of a separate current SOC 2 attestation.

A retrievable certificate with edition, scope, issuer and expiry is not attached to the announcement. The historical certification is a meaningful outside assurance claim, but the material does not establish the current renewal period or map every newer Research and agent feature into the audited scope.

The current acceptable use policy states that the service is not a medical device and excludes diagnosis, prescribing and treatment selection for individual patients. Research outputs concern scientific evidence, candidate drugs and program decisions rather than authorized instructions for care. This is consistent with the enterprise workflows for target selection, clinical study comparison and regulatory evidence preparation.

A case about supporting an FDA submission describes research work for a customer, not clearance of Causaly itself. Clinicians can use the individual research offer, but that audience does not remove the explicit restriction on patient specific clinical decisions or turn literature synthesis into regulated medical advice.

The five dimension evaluation framework assesses factual accuracy, analytical depth, argument structure, explicit assumptions and limitations. It gives scientific reviewers a defined way to examine reports beyond checking whether an answer retrieves a fact. The agent architecture adds a critique and revision process, and the current home page retains uncertainty, failed searches and contradictory evidence in its examples. These are concrete governance mechanisms.

The available results do not compare performance across disease families, demographic groups or poorly represented evidence domains. Claims of unbiased information therefore exceed what those disclosed evaluations establish, and organizational security certification does not answer that scientific performance question.

Cited evidence, retained methods and visible agent critiques let researchers challenge conclusions. The August 2026 Research terms also provide a service complaint channel, acknowledgement within five working days and a target resolution period of fourteen days. A performance remedy covers nonconforming service elements, while separate clauses disclaim responsibility for decisions based on AI outputs.

It is therefore unclear whether an incorrect scientific answer itself qualifies for repair under that warranty. The older enterprise indemnity expressly excludes generated content. There is no explicit public promise to correct every faulty conclusion or reimburse research losses caused by an answer, despite the general service support route.

The Microsoft partnership describes an integration in which Microsoft Discovery calls Causaly APIs to retrieve biomedical knowledge and inserts cited results into its reasoning workflow. Causaly also demonstrates access through Microsoft 365 Copilot. Its enterprise data fabric supports APIs and connectors for customer stores in Azure, Google Cloud and AWS, with custom connections available for other material. These named research pathways are more specific than a generic integration claim.

They do not establish a bidirectional hospital record connection or a named production EHR deployment. The relevant flow is research evidence entering analysis and decision workflows, rather than clinical documentation being written back to patient charts.

Enterprise internal data uses virtual private cloud architecture and a secure customer instance. The newer Microsoft solution describes Azure governance and tenant isolation, while the processing addendum identifies Google Cloud as a production host and AWS for hosting services. The privacy notice allows storage and processing in the UK, Greece, the US or other service locations.

Those disclosures identify infrastructure and possible countries without committing a particular customer dataset or inference task to a region. They also leave the relationship between the different deployment offers partly open. A complete map of resting data, model processing and cross border access is not available.

Causaly Research offers Pro at $99 per month for one user with 400 credits, and Intelligence at $149 per month with 500 credits. Both have a 14 day trial route without a credit card. Intelligence adds the biological and pipeline graph agents and additional journal access. A Team plan is advertised at $250 per seat monthly for three to fifteen seats, with pooled credits, but remains marked as coming soon in October 2026. Larger enterprise deployments are negotiated separately. These individual prices establish a usable entry budget, while custom workflow design, private data integration and organization wide access have separate commercial scope.

Research coverage includes target assessment, biomarkers, disease mechanisms, clinical evidence synthesis and medical affairs. Named customer accounts support parts of that range: Eisai uses the platform across oncology and neurology, ProQR applies it to target identification for RNA editing, and Manas AI uses it to assess target evidence before committing resources. Those deployments establish research use rather than validation in clinical care.

The broader catalog also covers indication expansion and competitive intelligence, without equivalent evaluation for every workflow. Individual researchers and smaller teams now have a separate access route, but their subscription availability does not itself establish performance across every scientific specialty.