AI Readiness Checklist.

What the model is going to learn from or retrieve over.

• At least one workflow has labeled examples (≥200 input/output pairs).
• Single source of truth for each metric you'd show an AI agent.
• Data retention and deletion policies match the model vendor's terms.
• PII boundaries documented — what crosses to the model, what doesn't.
• A named owner per dataset — not "the data team".
• Domain-specific terminology documented for prompt context.

What you need before any LLM call hits prod.

• Vector store or retrieval layer live (not just "we'll add it").
• Eval harness runs in CI on every prompt change.
• Per-call logging captures input, output, model, latency, cost.
• Distributed tracing extends through LLM calls (not a black box).
• Cost ceiling alerts wired to Slack/PagerDuty.
• Rate-limit and timeout strategy per upstream provider.

Who owns this surface when it breaks at 2am.

• At least one engineer with hands-on LLM application experience.
• Prompt versioning lives in code — not in a Notion doc.
• On-call rotation includes AI features (not just core services).
• A documented "how to debug a bad output" runbook.
• Engineers can describe the difference between fine-tuning and RAG.
• A budget owner separate from the engineer building the feature.

Where the AI actually plugs into the business.

• At least one named workflow with documented inputs and outputs.
• Human-in-the-loop checkpoint for any mission-critical surface.
• Documented fallback path when the model fails or is rate-limited.
• Success metrics that are not "the demo works".
• A defined off-ramp — what triggers killing the AI feature.
• User-facing copy when the AI is uncertain or refuses.

The boring layer that keeps you out of court and headlines.

• Model selection rubric documented (capability, cost, latency, vendor risk).
• Vendor lock-in mitigation — abstraction over at least 2 providers.
• Audit trail for high-stakes outputs (who saw what, when).
• Leadership accountability for AI KPIs — named exec.
• Legal/compliance review path defined before launching to users.
• Disclosure policy for AI-generated user-facing content.