1. Our approach to AI
Bayes Pharma.ai builds decision-intelligence software for regulated, high-stakes scientific work. We treat AI-assisted features the same way we treat any other model in the platform: as a source of structured evidence and hypotheses that must remain inspectable, bounded, and subject to expert review — never as an autonomous decision-maker.
2. Where AI is used
AI-assisted capabilities appear in a defined subset of workspaces, including:
- LLM Drug Design Chat — a structured conversational assistant for exploring target hypotheses and design questions.
- Generative Chemistry — candidate molecule generation and scaffold comparison against stated design objectives.
- Virtual Perturbation Lab — ranking of genes, compounds, and combinations by predicted disease-relevant cell-state effects.
- Active Learning Loop — prioritization of the next experiment using model uncertainty.
- ML Integration workspace — covariate selection, clustering, and hybrid pharmacometric modeling support.
- Evidence-grounded copilots in Evidence Hub–adjacent workspaces, which answer questions using retrieved source evidence and are designed to say when evidence is insufficient rather than fabricate an answer.
Workspaces outside this list may still use conventional statistical or mechanistic models (for example, classical PopPK or PBPK modeling); those are model-informed but not what we mean by "AI-assisted" in this statement.
3. Our principles
Where a feature is AI-assisted, the workspace says so, and shows the evidence, confidence, or source trail behind a suggestion wherever practical.
AI-assisted outputs are designed to sit inside a review workflow with a named, accountable human reviewer — not to bypass one.
Where feasible, AI features are designed to answer from retrieved, source-linked evidence and to say "not enough evidence" rather than invent an answer.
Confidence, limitations, and gaps are surfaced alongside AI-assisted outputs rather than hidden behind a single confident-sounding answer.
AI features are built so that one customer's confidential inputs and outputs are not exposed to another customer.
We monitor reported issues with AI-assisted features and update prompts, guardrails, or scope in response.
4. What AI does not do
- AI-assisted features do not make autonomous development, safety, regulatory, or business decisions.
- They do not replace qualified scientific, clinical, pharmacometric, or regulatory judgment.
- They are not guaranteed to be accurate, complete, or free of hallucinated content, and outputs may reflect limitations or biases in underlying data and models.
- They do not remove the need for validation, quality control, and accountable sign-off before an output is used in a consequential decision.
5. Data used for AI features
Where a workspace calls a third-party AI provider to generate a response, only the information necessary to answer the specific request (such as the evidence package relevant to the question asked) is sent to that provider, subject to that provider's own data-handling commitments. Bayes Pharma.ai does not use one customer's confidential Customer Content to train models that are shared with, or that improve outputs for, a different customer. Any use of aggregated, de-identified data to improve platform-wide AI features is designed to avoid re-identifying a customer or their confidential inputs.
6. Your responsibilities when using AI features
- Treat AI-assisted output as a starting hypothesis or draft, not a finished, validated conclusion.
- Verify material claims against primary source evidence before using them in a decision, report, or regulatory submission.
- Apply your organization's own review, validation, and sign-off requirements to any AI-assisted output used in a consequential decision.
- Avoid entering protected health information or other regulated patient-identifiable data into AI-assisted features unless your agreement with Bayes Pharma.ai expressly permits it with appropriate safeguards.
7. Alignment with AI governance
We design AI-assisted features with reference to emerging AI-governance frameworks and good-practice guidance, including risk-management approaches such as the NIST AI Risk Management Framework and the trajectory of regulation such as the EU AI Act, as well as general principles around bias, transparency, and human oversight increasingly expected in regulated industries. This is a statement of design intent and ongoing practice, not a claim of formal certification under any specific framework.
8. Feedback & reporting issues
If an AI-assisted feature produces an output you believe is inaccurate, biased, unsafe, or otherwise concerning, please tell us — this directly shapes how we tune and scope these features.