A pragmatic validation pattern for AI-assisted CAPA under 21 CFR 820 and ISO 13485

Short summary

The author prototyped an AI assistant for CAPA triage in a medical-device QA context and confronted the validation challenge under 21 CFR 820 and ISO 13485. They describe a pragmatic pattern: shadow-mode testing against historical cases, a golden-test corpus with measurable acceptance criteria, continuous drift monitoring, human-in-the-loop sign-off, and full traceability via source-controlled prompts and test suites. The approach treats the LLM as controlled assistance rather than a deterministic SOP, producing validation artifacts regulators can accept.

• •Prototyped LLM-assisted CAPA triage but faced validation challenges under 21 CFR 820 / ISO 13485
• •Used shadow-mode testing, golden-test corpus, and measurable acceptance criteria to validate AI suggestions
• •Human-in-the-loop sign-off and full traceability make the controlled-assistance process auditable

Generated with AI, which can make mistakes.