AI inside
Assistance you can show an inspector.
The agent works inside the record, not around it. Everything it proposes is evidence; everything that counts is a human decision.
- What the agent does
- Finds similar past deviations, checks a triage or root cause against the evidence, proposes the next why, runs the SMART check on a CAPA, drafts citations and the final report text.
- What it never does
- It never sets a severity, closes a record, signs, or advances a stage. Its output is stored as evidence next to the decision a named person took.
- How a model is admitted
- Each model is qualified inside the product: reviewed evaluation cases, a candidate run, a replay the QA reviewer approves, and a registry entry. No approval, no generation.
- What is recorded
- Every run keeps the input hash, the prompt version, the model, the raw output, token usage and the human decision, append-only.
This follows the direction of EU GMP Annex 22 on artificial intelligence: generative models may assist, and a qualified person takes every GMP decision. Probity is AI-assisted, never AI-decided.
Evaluate Probity with your own synthetic case.
Evaluation sites run Probity on synthetic data with a local or cloud model. The model is qualified inside the product before it is used on any record.