AI Governance

PrimusReview uses Claude, developed by Anthropic PBC, accessed via the Anthropic API. Model: Claude Sonnet (claude-sonnet-4-5) Provider: Anthropic PBC, San Francisco, California, USA Access method: Anthropic Messages API API version: 2023-06-01 For our Video Review feature, uploaded audio/video content is first transcribed into a timestamped transcript by AssemblyAI Inc. (United States) before being reviewed by Claude in the same manner as document reviews. Anthropic is an AI safety company whose mission is the responsible development and maintenance of advanced AI for the long-term benefit of humanity. Further information about Anthropic's safety practices is available at anthropic.com.

When a user submits a document for review, PrimusReview: 1. Extracts text content from the submitted file (PDF, PPTX, image, or text) 2. Constructs a structured prompt containing the extracted text, the material type selected by the user, applicable regulatory standards, and any source documents from the user's library 3. Sends this prompt to the Anthropic API 4. Receives a structured JSON response containing the compliance analysis 5. Parses and displays the analysis to the user The AI performs a first-pass review checking for: - Compliance with the ABPI Code of Practice 2024 - Mandatory inclusions (AE reporting statement, job bag reference, date of preparation etc.) - Potentially misleading or unsubstantiated claims - Reference and balance issues - Material type-specific standards (MRS Code for market research, NICE methodology for market access etc.)

The system prompt is calibrated per material type to apply the appropriate regulatory standards: - Promotional (HCP): Full ABPI Code 2024, maximum rigour - Promotional (consumer/patient): ABPI Code + PAGB guidelines - Non-promotional (medical/MSL): ABPI Code Chapter 22, disguised promotion check - Non-promotional (education): ABPI Code Chapter 22, funding disclosure - Market access/HEOR: ABPI Code + NICE reference case methodology - Market research: MRS Code of Conduct, ISO 20252 - Internal: ABPI Code factual accuracy requirements Each material type triggers specific mandatory inclusion checks and clause references appropriate to that category.

PrimusReview maintains version control on its compliance review prompts. Each review result is stored with the prompt version used to generate it, enabling audit trails that identify which version of the compliance logic was applied. Current prompt version: v1.0 When significant updates are made to the system prompt (e.g. following a new edition of the ABPI Code), the prompt version is incremented. Historical reviews retain their original prompt version reference.

Users must be aware of the following limitations: Hallucination risk: Large language models can generate plausible-sounding but incorrect output. Clause references, case citations, and regulatory interpretations should always be verified by a qualified professional. Context window limitations: Very large documents may be truncated before submission to the API. The system processes up to 50MB files but extracted text is subject to token limits. Abstract-level analysis only: The AI reviews extracted text and cannot assess visual design elements, prominence of safety information in the context of the overall layout, or the relative size and positioning of text. Current knowledge only: The AI's knowledge of regulatory standards reflects its training data and the source documents provided. It may not reflect very recent PMCPA rulings or interim ABPI guidance issued after its training cutoff. Not a qualified signatory: PrimusReview does not constitute the review of a qualified ABPI Code signatory, medical reviewer, legal adviser, or regulatory affairs professional. All outputs require human review before being relied upon.

PrimusReview is designed as a first-pass tool that supports, not replaces, human review. The following safeguards are built into the product: Explicit AI disclaimer: Every analysis panel displays a prominent disclaimer stating that outputs are AI-assisted and require human validation before reliance. Status tracking: Users can annotate each flag with a status (Open, Actioned, Disagree, Escalated) to document their human review of each issue. Commentary field: Users can add notes and comments to each flag documenting the rationale for their decisions. Overall review status: Users can mark reviews as Draft, In Progress, Actioned, or Signed Off to track the human review workflow. Export with annotations: The export function includes all human annotations alongside the AI output, creating an audit trail of both AI findings and human decisions. We actively discourage users from submitting materials to formal MLR review based solely on PrimusReview output without human validation.

Anthropic does not use inputs submitted via the API to train its models by default. Document content submitted to PrimusReview for review is not used to train the underlying AI model. PrimusReview does not use review outputs or user interactions to train any AI model. For further information on Anthropic's data practices, see anthropic.com/privacy.

Risk: Confidential pharmaceutical materials are processed by third-party AI and transcription providers. Mitigation: Data is processed under Anthropic and AssemblyAI's standard commercial and data processing terms. Users are advised not to submit materials containing identifiable patient data. Data is transmitted over TLS and not retained for training by either provider.

The following governance processes are in place: ABPI Code updates: When a new edition of the ABPI Code is published, the system prompt will be reviewed and updated accordingly. The prompt version will be incremented and existing users notified. Model updates: When Anthropic releases significant model updates, PrimusReview will evaluate the impact on review quality before updating the model version used in production. Incident response: Any reports of significant AI errors or misleading outputs will be investigated and used to improve the system prompt and associated guidance. This AI Governance document is reviewed at least annually and updated to reflect material changes to our AI use.