# AI production-readiness checklist

*This checklist helps you measure and document readiness. It does not certify safety, CE/MDR, or EU AI Act conformity.*

## 1. Purpose and ownership

- [ ] Named product owner and engineering owner for the LLM feature  
- [ ] Written intended use (what the system must and must not do)  
- [ ] Kill criteria agreed (when we stop the pilot)  

## 2. Data and retrieval

- [ ] Sources of truth listed (docs, APIs, tickets)  
- [ ] Access control on retrieval matches source systems  
- [ ] PII and secrets policy for prompts and logs  
- [ ] Health-tech only: no patient data in public evals; synthetic or client-vaulted sets  

## 3. Quality gates

- [ ] Golden and eval set exists (≥50 cases for a thin product; more for customer-facing)  
- [ ] Metrics defined: grounding, task success, refusal, latency, cost per task  
- [ ] Pass/fail thresholds written before the next model or prompt change  
- [ ] Regression suite runs in CI or on a fixed schedule  

## 4. Safety and abuse

- [ ] Prompt-injection and jailbreak cases in the suite  
- [ ] Tool and agent permissions inventoried (least privilege)  
- [ ] Human escalation path when the model abstains or fails  

## 5. Operations

- [ ] Logging of inputs/outputs with retention limits  
- [ ] Cost and latency dashboards  
- [ ] On-call or owner for quality regressions  
- [ ] Rollback plan for prompt/model/config  

## 6. Documentation for stakeholders

- [ ] One-page risk summary for leadership  
- [ ] Evidence folder counsel can reuse (tests, scores, gaps)  
- [ ] EU AI Act: inventory row and transparency notes if customer-facing — **with your lawyer**, not as “certified”  

## Score yourself

| Checks done | Signal |
| --- | --- |
| Under 10 | Pilot theatre — Readiness Check maps blockers in one week |
| 10–18 | Partial — LLM Quality Audit measures the live system |
| 19+ | Ready to harden — still re-test after every material change |

**Next step:** [Request audit scope](https://arviser.com/#scope) on arviser.com. Optional: [sample LLM Quality Audit](https://arviser.com/sample) (work email for the full file).
