QR-05

Gen AI Project Checklist

Idea to production in six phases. Tick every box before you ship, and again before every review.
Last verified 2026-07-09
Read this once, then work top to bottom. Skipping a phase does not save time, it defers the failure to a more expensive stage. Most Gen AI projects that stall in production skipped Evaluate; most that get killed in review skipped Govern.

FrameStep 1 of 6

ItemWhy it matters
Problem stated as a user outcomeNot "let's use AI". Name the user, the job, and what better looks like
Success metric defined and measurableDecided before any build. "Accuracy above 85%" beats "improved experience"
Baseline capturedWhat does the current process deliver? Without this you cannot prove a lift
Build vs buy vs API decidedRough cost per request, per month, per year. Even a bad estimate reveals the shape
Sponsor and users namedOne accountable owner. Three named users who will test it

Ground The DataStep 2 of 6

ItemWhy it matters
Source data identified and reachableIf retrieval is needed, it must exist somewhere queryable
Rights and permissions confirmedWho owns it, who can see it, what happens if it changes
Sensitive data classifiedPII, PHI, financial, IP. Different handling for each
Residency and cross-border rules checkedRegulated industries stall here. Confirm before you build
Chunking or indexing strategy chosenThe unit of retrieval decides what the model can ever know. See QR-03
Freshness plan definedHow stale is too stale? Who re-indexes, and when?

BuildStep 3 of 6

ItemWhy it matters
Model selected by fit and costRoute by workload, not by hype. See QR-02 and QR-15
Prompt written to a specRole, task, context, constraints, output, escape hatch. See QR-01
RAG or fine-tune decided, not both by defaultFacts belong in retrieval, behaviour in fine-tuning. See QR-06
Guardrails in placeInput filters, output checks, refusal patterns. Not optional
Prompt injection defence testedUntrusted content fenced. Instructions inside data are ignored
Structured outputs where downstream code depends on themJSON schemas, validation, retry on malformed
QR-05

Gen AI Project Checklist

Evaluate, Ship, Govern — and the skip-steps that will bite you.
Last verified 2026-07-09

EvaluateStep 4 of 6

ItemWhy it matters
Golden dataset built from real inputs50 to 200 hand-labelled examples. Not clean cases, real ones
Metric matches the goalRAGAS for retrieval, exact match for extraction, LLM-as-judge for prose
Automated evals run in CIEvery prompt change and model change re-runs the set. See QR-11
Adversarial and edge cases testedEmpty inputs, hostile inputs, injection attempts, off-topic requests
Fairness or bias checks where relevantEspecially for hiring, credit, health, legal. Document the results
Regression gate on the pipelineShips only when the golden set clears the bar

ShipStep 5 of 6

ItemWhy it matters
Cost caps and rate limits setPer user, per feature, per day. A runaway loop is expensive fast
Tracing and monitoring live from day oneEvery prompt, every response, every tool call. Debug without guesswork
Prompt caching enabled where possibleAround 90% off repeated input if the front of the prompt is stable
Human-in-the-loop for destructive actionsWrites, deletes, sends. Friction is a feature, not a bug
Rollback path definedEvery deploy needs an undo. Prompts are code
Launch communications draftedWhat it does, what it does not do, who to contact. Sets expectations

GovernStep 6 of 6

ItemWhy it matters
Risk tier assessedAgainst EU AI Act, NIST AI RMF, or your internal framework. See QR-13
Decisions and limitations documentedA model card. What it does well, what it fails at, what data it saw
Data handling documentedWhat is logged, how long, who can access it, when it is deleted
Owner named for the monthly reviewSomeone is accountable. Not "the team", a person
Incident response runbook existsWhat to do if it hallucinates in production, leaks data, or breaks
Retirement plan sketchedEvery project ends. Decide up front how

Common Skip-Steps these are the ones that will bite you

The skipWhy it hurts later
Shipping without evalsEvery regression is a surprise. Every rollback is a firefight
Chasing a demo, not an outcomeImpressive prototype, no real user problem, no path to production
Assuming the model can hold your knowledgeIt cannot. Give it the documents (RAG). See QR-03 and QR-06
One heroic prompt doing three jobsDebug it once and you will split it. Split it up front instead
No cost cap, no rate limitThe first runaway loop teaches this the expensive way
"We will do governance later"Later is after the incident. Do it now, or a regulator will
Testing on the happy pathThe failure mode is what you did not test. Look for it deliberately
Picking the biggest model availableRoute by workload. A 25x price gap makes routing worth engineering
The 60-second pre-launch review. Walk this list before every release. Success metric measurable? Golden set green in CI? Cost cap set? Human-in-the-loop wired for destructive actions? Owner and runbook named? Risk tier documented? If any answer is "we will do that later", stop. Later is what the outage retrospective will call "the root cause".