Checklist · AI

AI Project Readiness Checklist

Run before committing to an AI project — whether the problem needs AI at all, whether the data exists, who owns the decision when the system is wrong, and the measurements that decide success.

Download the checklist Read the full guide

Markdown. No sign-up, no email.

Project: _______________ Owner: _______________ Date: _______

1. Does this need AI?#

[ ] The task involves judgement, language, or pattern-finding — not arithmetic on structured data
[ ] Rules would need constant updating, or nobody can enumerate them
[ ] A deterministic solution has been considered and ruled out, with a reason
[ ] Being occasionally wrong is acceptable, with a human able to catch it

🔴 The most expensive AI projects are the ones that should have been a query, a rule, or a form. If someone can write the rules down, write the rules down — they are cheaper, testable, and they explain themselves.

Deterministic alternative considered: _______________ Why rejected: _______________

2. The problem#

[ ] Stated in one sentence, in the user's words
[ ] The person who has this problem today has been asked about it
[ ] What they do instead today is documented, with how long it takes
[ ] The value of solving it is estimated — time, cost, or risk avoided
[ ] Failure is survivable, or a human reviews before anything acts

What happens when the system is confidently wrong? _______________

3. Data#

[ ] The knowledge needed exists in writing, not only in people's heads
[ ] Someone owns each source and is accountable for it being current
[ ] Access is legally and contractually permitted for this use
[ ] Personal data identified, with a lawful basis
[ ] Volume and quality checked by looking at it, not by asking about it

If the answer lives in people's heads, this is a documentation project wearing an AI project's clothes. Do that first.

4. Measurement#

[ ] Success is defined as a number, before building
[ ] A baseline exists — what does the current process achieve?
[ ] Test set written from real cases, with known answers
[ ] Someone has decided what "good enough to ship" means
[ ] Cost per operation will be measured
[ ] Latency requirement stated

Metric	Baseline	Target

Without a baseline you cannot demonstrate benefit, and the project will be judged on impressions.

5. Accountability#

[ ] A named person owns the output, not a team
[ ] It is clear who is responsible when it is wrong
[ ] Users are told they are interacting with AI where required
[ ] Generated content is labelled where required
[ ] There is a route to a human
[ ] Decisions with legal or significant effect on people have a human in the loop

Automation does not transfer accountability. Whoever owned the outcome before still owns it.

6. Operations#

[ ] Where it runs, and who is paged when it stops
[ ] Cost ceiling set and enforced
[ ] Model and version pinned
[ ] Behaviour when the model provider is unavailable
[ ] Prompts and configuration in version control
[ ] Bad outputs can be reported, and the report reaches someone

7. Before go-live#

[ ] Measured against the test set, result recorded
[ ] Tested with real users on real work
[ ] Failure paths exercised deliberately
[ ] Cost per operation known
[ ] Rollback or disable path tested
[ ] Documentation says what it cannot do

What this system must never be used for: _______________

Sign-off#

	Name	Date
Built by
Business owner
Approved for users

Back to AI