Checklist

Engineers joining an AI team need more than repository access. This checklist covers the architecture, the evaluation suite, the cost model, and the failures already learned from.

Access to AI systems accumulates quietly — people move, agents gain tools, service accounts persist. This checklist covers the periodic review that removes what is stale.

Board reporting on AI works when the same numbers appear every quarter. This checklist covers what to include so a crisis conversation starts from shared understanding.

AI spending is moving onto operational budgets where it competes with everything else. This checklist covers what a defensible budget review contains.

AI logs contain the data most in need of protection and the records you cannot recreate. This checklist covers capturing enough, protecting it, and deleting it on time.

Deployed AI systems drift. This quarterly review covers quality, scope, permissions, cost, and corpus health, and identifies what should be retired.

Rollback decisions are made under pressure with incomplete information. This checklist covers deciding, reverting the right things, and handling what the bad version did.

AI vendors introduce risks ordinary software vendors do not: model providers as subprocessors, training use of your data, and decisions you must account for.

An agent's permissions are its blast radius. This checklist covers reviewing them as a set rather than tool by tool, which is where the real exposure hides.

AI capacity planning has unusual constraints — provider rate limits, per-call cost, and quality that degrades under load shedding. This checklist covers sizing properly.

Generated code is plausible, which is exactly why it needs review. This checklist covers what to check when the author cannot explain their reasoning.

AI systems fail in ways conventional disaster recovery plans do not cover. This checklist adds provider outage, model withdrawal, and quality collapse to the plan.

Availability monitoring stays green while an AI system produces wrong answers. This checklist covers the signals that actually indicate something is wrong.

Releasing an AI change is different because the change may be a prompt, a model, or a corpus. This checklist covers what to verify before any of them reaches production.

An agent that works in testing is not ready to act on real systems. This checklist covers the scope, permissions, evaluation, observability, and escalation work that go-live requires.

Migrating a corpus or vector store loses content quietly unless it is checked. This checklist covers inventory, permissions, re-embedding, validation, and a cutover you can reverse.

Model selection made on benchmarks and impressions ages badly. This checklist covers evaluating on your own cases, at your real volume, with the commercial terms that matter.

Leaving an AI vendor is harder than leaving ordinary software because the data includes history you cannot recreate. This checklist covers exiting without loss.

Most AI cost problems are a few specific patterns. This checklist covers attribution, routing, caching, prompt size, and the waste that nobody sees until it is measured.

A prompt change is a behaviour change. This checklist covers what a reviewer should look for before a prompt reaches production, and what evidence should accompany it.

Most RAG problems are corpus or retrieval problems that present as model problems. This checklist covers what to audit, in the order that finds faults fastest.

Before the first agent reaches a production system through MCP, work through this checklist. It covers the gateway, identity, server design, permissions, testing, operations, and governance, so adoption does not become a new layer of shadow IT.

Generated code inherits the vulnerabilities of its training data and the agent's misunderstandings of your internal APIs. This checklist gives the controls that treat it as untrusted input: pipeline checks, access boundaries, review, and audit.

Prompt injection cannot be fully solved at the model layer, so defense is layered. This checklist covers content separation, detection, least-privilege tools, output validation, approval gates, monitoring, and adversarial testing.