From Demo to Dependable System: An AI Readiness Checklist

A prototype proves that a workflow can work. Readiness proves that a team can operate it when inputs are messy, dependencies fail, usage grows, and the model changes.

Seven Gates Before Launch

1. Ownership: A named team owns quality, policy, incidents, cost, and retirement.
2. Evaluation: Representative cases, difficult slices, and unacceptable failures have release thresholds.
3. Security: Data scope, tool permissions, untrusted content, secrets, and consequential actions have enforced boundaries.
4. Fallback: The product can degrade, retry safely, hand off, or stop without corrupting work.
5. Observability: Teams can see latency, cost, quality signals, tool errors, user corrections, and external outcomes.
6. Change control: Model, prompt, retrieval, tools, and policy changes are versioned and tested before rollout.
7. Economics: Expected volume, slow paths, review labor, and dependency costs fit a defined budget.

Rehearse the Bad Day

Run an exercise with a provider outage, stale retrieval index, revoked credential, sudden traffic increase, or harmful output report. Verify who receives the alert, what the user sees, how work is preserved, and how the system is restored.

Launch as a Controlled Expansion

Begin with a bounded audience and reversible actions. Compare real behavior against the evaluation set. Review user corrections and near misses, not only visible incidents. Expand only when evidence supports the next level of autonomy.

Production readiness is not a document signed once. It is the operating discipline that keeps a useful demo useful after launch.