GenAI choices and the pre-launch checklist
Generative deployments add their own questions - fine-tuning, retrieval-augmented generation, vector/graph databases and agentic architectures. Engineering and governance then share a pre-launch checklist: determine applicable laws, document appropriate uses, assess risk tolerance, and build in sufficient TEVV cycles.
Generative deployments add their own questions → then engineering and governance share a final checklist.
- Is the model used as-is, or was it (or can it be) fine-tuned?
- Was retrieval-augmented generation used → optimising LLM output by referencing a knowledge base beyond the training data?
- What vector and/or graph databases are involved?
- Are agentic architectures appropriate → e.g., customer support agents, personal AI assistants, AI research assistants, workflow automation bots?
- Maintain best practices on performance fit and governance requirements → accuracy, transparency, fairness.
- Keep contingency plans for model or vendor issues.
- Build the timeline with sufficient TEVV cycles → test, evaluation, verification, validation.
Determine the applicable laws and policies (AI-specific, sector-specific, privacy → e.g., HIPAA may cover training data in healthcare) · consider system options including redress · document the appropriate uses of the AI to prevent unintended-purpose use, since the AI's appropriateness does not transfer to new purposes · assess risk tolerance · perform or review a risk assessment · evaluate the vendor or licensing terms. And the recurring rule → with competing values like accuracy vs privacy, prioritise with stakeholder consensus and document the decision.