Level 4: Optimized

Last updated: 2026-07-27

List of references

• GenAIOps for MLOps practitioners
• Well-Architected guidance for MLOps and GenAIOps
• Progressive exposure deployment
• Evaluation of generative AI applications

Level 4 uses automation and measured feedback to improve quality, safety, reliability, and cost without bypassing governance. Optimization means making better decisions quickly, not automatically deploying every change.

Source: Evaluate generative AI applications.

Optimization practices

Practice	Outcome
Automated evaluation	Candidate prompts, models, and retrieval configurations are measured before release.
Progressive exposure	Canary, blue-green, or A/B releases reduce the risk of broad negative impact.
Feedback loops	User feedback and production telemetry produce prioritized, reproducible improvements.
Advanced analytics	Usage, quality, safety, and cost trends are visible to product and engineering teams.
Automation	Workflows trigger evaluation, approval requests, remediation, or retraining with appropriate controls.

Safe continuous improvement

1. Identify a measurable improvement hypothesis.
2. Build a candidate prompt, model, or retrieval configuration in version control.
3. Evaluate against the approved dataset and safety cases.
4. Route limited traffic to the candidate and monitor the agreed signals.
5. Promote, roll back, or iterate based on evidence and documented approval.

Use automation and continuous improvement for implementation patterns.