Checklist

The first AI incident arrives before the process is ready unless the process is written first. This checklist covers detection, containment, reconstruction, remediation, disclosure, root cause, and prevention for AI systems and agents.

A model that passed evaluation can still fail in deployment through skew, missing fallbacks, or an untested rollback. This checklist covers registry and approval, parity, progressive rollout, monitoring, rollback, and documentation.

AI systems can be up and fast while being wrong. This checklist covers what observability must include: traces, operational metrics, quality signals, cost telemetry, drift indicators, alerting with runbooks, redaction, and retention.

Pilots fail by proving nothing: no hypothesis, no baseline, no exit criteria. This checklist covers what a pilot must define before it starts and deliver before it ends, so the go or no-go decision is made on evidence.

AI systems create new personal-data flows through prompts, retrieval, logs, and model providers. This checklist structures a privacy impact assessment: data mapping, legal basis, minimization, transfers, rights, retention, risks, and mitigations.

A handoff is complete when the receiving team can operate, change, and extend the system without the builders. This checklist covers documentation, runbooks, evaluation assets, access, trained owners, support transition, and backlog.

Before an AI agent touches production systems, every item on this checklist should be true and evidenced. Specification, permissions, gates, evaluation, safety, reliability, observability, and operations.

APIs that expose AI capabilities carry latency, cost, non-determinism, and safety concerns ordinary APIs do not. This checklist covers contracts, streaming, idempotency, errors, versioning, authentication, limits, observability, and cost.

AI budgets that count only licenses and a build fee are wrong before they are approved. This checklist covers the full cost categories, the baseline and value case, platform amortization, run and oversight costs, contingencies, and governance.

AI systems are abandoned for human reasons more often than technical ones. This checklist covers communication, role redesign, involvement, training, trust calibration, adoption measurement, incentives, and resistance, for each deployment.

Chatbots fail publicly when they answer outside their knowledge, make promises, or strand users. This checklist covers scope, grounded knowledge, guardrails, escalation, evaluation, privacy, analytics, staged rollout, and operations.

AI spend is optimized by making it visible, then attacking the largest lines in order. This checklist covers attribution, caching, routing, context discipline, batching, budgets, review efficiency, and unit economics.

Data readiness is the best predictor of whether an AI project reaches production. This checklist scores the sources a use case needs across access, structure, metadata, permissions, quality, governance, and tacit knowledge.

AI systems need documentation that regulators, auditors, operators, and successors can use. This checklist covers the specification, model or system card, data documentation, evaluation reports, runbooks, decision records, and change history.

Accepting an AI system means verifying it meets the specification under real conditions, not watching a demo. This checklist covers acceptance criteria, golden set results, business scenarios, safety, performance, operability, and sign-off.