Induced from Robert (Bob) Wherry’s ten-step framework lecture at the KENX AI in GxP East 2026 conference (2026-09-16-1107-gxp-ai-ml-implementation-framework-lecture), 2026-09-16.
The spectrum replaces the binary
Traditional validation assumes you can prove a system does what it claims. Non-deterministic AI breaks the assumption, so the framework splits the response: validate what you can (deterministic components, data pipelines, guardrails) and verify the rest against statistics-based performance metrics. Underneath sits the motto “Never Trust! Always Verify!” and the four variability sources any AI system carries: model, infrastructure, data environment, and human. The motto is the same demand as Human verification of AI output must be explicit because document creation no longer carries implicit verification, and the statistics-based remainder is why AI validation evidence has a shelf life.
The framework around the spectrum
His ten steps include an authorized-use policy, honest token-cost budgeting, quality and regulatory SMEs on development teams from the start, a formal AI lifecycle with its own SOPs, pinned public models with bounded inputs, procurement controls with an AI inventory, QMS quality gates, documented human verification, and a mock audit by an SME before the formal one. The dedicated SOPs are AI development needs its own SOPs, separate from the software development lifecycle; pinning and bounding are Bounded inputs make AI validation tractable; documented human verification is Human-in-the-loop must be a named hard gate, not rubber-stamp oversight. His pushback from the floor: force-fitting AI into the software development lifecycle is a square peg in a round hole.
Related
- Governance Risk and Human Factors MOC
- AI development needs its own SOPs, separate from the software development lifecycle
- AI validation evidence has a shelf life
- Human verification of AI output must be explicit because document creation no longer carries implicit verification
- Bounded inputs make AI validation tractable
- Human-in-the-loop must be a named hard gate, not rubber-stamp oversight