Deployment Checklist¶
Last updated: 2026-07-27
List of references
Use this checklist before a generative AI application, model, prompt, grounding index, or significant configuration change reaches production.
Foundation¶
- Intended use, limitations, success measures, and human escalation path are approved.
- Data classification, privacy requirements, and retention rules are documented.
- Model, deployment region, grounding data, and content safety configuration are approved.
- Prompts, application code, evaluation data, and infrastructure definitions are versioned.
Evaluation and release¶
- Representative quality, safety, adversarial, and failure cases are included in evaluation.
- Candidate results are compared with the production baseline where one exists.
- Latency, throughput, and cost expectations are verified.
- A staging test and rollback path are tested before production routing.
- Required business, security, and responsible AI approvals are recorded.
Operations¶
- Azure Monitor dashboards and alerts cover reliability, quality, safety, and cost signals.
- Logging minimizes sensitive prompt and response content and follows retention controls.
- Endpoint access uses least privilege, managed identities, and approved network boundaries.
- Incident owner, escalation process, and user communication plan are current.
- Feedback and continuous-improvement reviews have a defined owner and cadence.