PhaseV

Causal machine learning is central to identifying treatment responders and refining trial eligibility. Response Optimizer processes trial data, selects features, estimates differences in treatment effect, and identifies subgroups for further validation. Trial Optimizer combines these insights with statistical simulations to compare alternative designs. These analyses determine the platform's recommendations rather than serving only as an optional writing assistant.

AI Conductor adds generated clinical development documents and automated coordination of dependent changes. Conventional statistical methods, expert services, and team approvals remain part of the workflow, but removing the learned analyses would materially change the core product offered to sponsors.

AI Conductor prepares forms, propagates protocol amendments, checks captured data, and assembles downstream study outputs. Team approval is an explicit part of the workflow, including approval of dependent document changes before release. However, the assertion that propagation leaves nothing behind makes an unqualified completeness promise about reaching every affected document. Traceability and approval do not establish that omitted dependencies will always be found.

The boundary between deterministic propagation and model interpretation is not sufficiently specified to resolve that tension. Confidence thresholds, conditions that suspend automated changes, and recovery procedures for missed amendments remain unclear despite the defined human review role.

The analytical approach combines causal inference, heterogeneous treatment effect estimation, feature selection, and simulation. CARD describes a conformal prediction method for identifying responders while controlling false discoveries, including an adjustment for nonrandom treatment assignment. Trial design optimization also uses Monte Carlo simulation to compare statistical and operational scenarios. These descriptions give technical substance to the analytical products.

However, deployed algorithm versions, complete training data provenance, and update procedures are not mapped consistently across the portfolio. AI Conductor's underlying language models and their versions are also unnamed, so the transparency of the older statistical methods does not fully explain its generated documents.

AWS is identified as the platform host, and the analytical modules use methods developed by PhaseV. The company also combines sponsor-provided data with a proprietary data lake, external trial information, and research sources. These details identify part of the infrastructure and analytical ownership. They do not identify every downstream model processor.

In particular, AI Conductor's foundation model providers, inference services, and applicable prompt retention or training terms remain unspecified. General privacy commitments to contracts with service providers do not substitute for a current named subprocessor list that maps those providers to the clinical data each module sends.

A March 2025 Oramed case describes reanalysis of a failed Phase III oral insulin trial. The analysis combined Phase III and earlier Phase II data, checked findings against external Phase III studies, and included sensitivity and false discovery tests. Responding subgroups differed by age, body mass index, and baseline HbA1c; the case connects the analysis to an amended trial protocol.

Separately, the vendor-authored CARD preprint explains responder detection methods and simulation experiments. This supports a defined retrospective analytical use. It does not establish independent prospective improvement in clinical outcomes or validate the current AI Conductor document workflow.

The privacy policy retains personal information during service and afterward for legal and business needs, without a fixed deletion deadline for trial datasets. Nonidentifying usage statistics can persist separately. Marketplace permission to improve services using User Data cannot be extended to patient or clinical trial data, which that definition excludes. Patient-data training permissions, deidentification methods, and model-specific retention therefore remain unclear.

Document traceability and human approval help reviewers examine generated changes, but do not define hallucination detection thresholds or recovery after a model error. The privacy policy's breach notification commitment concerns data incidents rather than a complete clinical AI safety process.

HIPAA and GDPR compliance are claimed for the cloud platform. The service privacy policy covers account information, uploaded content, communications, international access, and individual privacy rights, giving customers more context than a security badge alone. It also describes contractual restrictions on service providers handling personal information. A readable BAA identifying eligible products, agreement terms, and corresponding subcontractor obligations remains unavailable.

The privacy policy is not itself that agreement. Trial data can have a different contractual status from account and website data, particularly because the marketplace terms define User Data separately from clinical trial and patient information.

ISO 27001 certification is claimed for PhaseV, alongside HIPAA and GDPR compliance. The privacy policy describes restricted access through secured networks and confidentiality obligations for personnel with special access rights. It also requires contractual safeguards for third parties processing personal information. These statements identify a recognized information security standard and some operational controls.

The actual certificate, standard edition, certified entity and service scope, issuing body, and validity dates remain unavailable for direct inspection. As a result, the certification claim cannot establish precisely which current modules and infrastructure are covered, particularly the newer AI Conductor functions.

The product supports research design, statistical analysis, trial operations, and preparation of sponsor submissions. Response Optimizer identifies groups for study planning, while AI Conductor produces development documents and standardized datasets for review by the sponsor's professionals. This is a clinical development workflow rather than an advertised autonomous diagnostic or treatment device.

References to regulatory readiness, statistical guarantees, or a sponsor's amended protocol concern the work produced with the software. They do not establish clearance of the software itself. No medical device authorization is claimed for the current portfolio, and regulatory acceptance of a particular trial remains specific to that sponsor and submission.

Analytical validation includes separate training and test data, external datasets, sensitivity analysis, and false discovery controls. The CARD research also addresses bias from nonrandom treatment allocation through propensity score adjustment. These are relevant statistical safeguards for responder analysis, but they answer different questions from fairness across demographic groups or governance of a deployed document model.

Finding groups with different treatment responses is not evidence that error rates are equitable across those groups. A current portfolio-wide governance process, comparative fairness results, and customer review artifacts for AI Conductor remain unspecified. Older methodological research does not establish controls for every current module.

Sponsor teams can review generated changes and trace study outputs to their underlying data, which provides a practical way to identify and correct work before release. That review process is different from a contractual promise to compensate for an erroneous output. The marketplace EULA disclaims output warranties and generally limits liability to $10,000.

Standard technical troubleshooting is available, but neither an output accuracy guarantee nor a defined remedy for affected trial participants is established. Privacy access and correction rights concern personal information; they do not create a dedicated appeal or compensation mechanism for an incorrect model recommendation or missed document change.

AI Conductor creates case report forms for electronic data capture systems and compares incoming records with the expected form structure. It also transforms raw study data into SDTM datasets, derives ADaM analysis datasets, and produces tables, listings, and figures. These are meaningful trial data workflows, distinct from a verified connection to a hospital's patient record.

The platform is described as compatible with most EDC systems, but named connectors, supported versions, and production exchange details remain unclear. External APIs and containerized custom analytics are additional direct-contract offerings rather than standard capabilities included in the marketplace subscription.

The standard platform is a cloud service hosted on AWS, with subscriptions available for self-service use and separate project services. Sponsor data is described as isolated from the broader data resources used by the platform. Privacy terms allow access from the United States and Israel and contemplate transfers beyond a user's home jurisdiction, with contractual safeguards for relevant transfers.

Those terms do not commit patient-level trial data to a particular storage or inference region. The isolation design, backup locations, support access boundaries, and model processing route are not sufficiently detailed to establish a complete residency commitment for each module.

Trial Optimizer has public 12-month prices through AWS Marketplace: $50,000 for Exploration Access, $70,000 for one Phase 1 or Phase 1/2 optimization, $120,000 for one Phase 2 or Phase 3 optimization, and $500,000 for a five-trial bundle. Exploration covers fixed trial scenarios; the optimization plans add more advanced simulation and design functions. A purchase route is available before a sales conversation, although customized offers are also offered.

Transactions are nonrefundable and additional infrastructure charges may apply. Directly contracted services, external integrations, and the Causal-to-Design addition are outside this standard pricing scope, as is a complete AI Conductor subscription price.

The intended setting is sponsor and CRO clinical development, with users across biostatistics, medical writing, clinical operations, and regulatory affairs. Applications span retrospective analysis, prospective trial design, recruitment planning, site selection, and submission preparation. Diabetes has a concrete named example in the Oramed oral insulin work. The broader portfolio also targets oncology, immunology, neurological disease, metabolic disease, and rare disease research.

Support for several therapeutic areas is plausible because the software operates on trial data and statistical designs, but that breadth is not equivalent to measured performance for every indication. Module-specific experience matters when moving from responder analysis to new document automation.