AI Readiness Checklist: 25 Questions Before You Start an AI Project

AI projects rarely fail because the model was not clever enough. They fail because the problem was vague, the data was not accessible, nobody owned the outcome, or the risk was discovered too late. Each of those is checkable up front. Work through the questions below; every "no" is a risk to resolve before you build, not a reason to give up.

1. The problem

• Can you describe the task in one sentence, including who does it today?

• Do you know how often it happens and how long it takes (volume × time)?

• Is there a measurable success metric — time, error rate, cost, revenue — with a current baseline?

• Would a 70–80% automated solution with human review still be valuable?

• Have you ruled out a simpler fix (a better form, a rule, an off-the-shelf tool)?

2. The data

• Do you know where the inputs live, and can an engineer get programmatic access to them?

• Do you have at least 50–100 real, representative examples, including the messy ones?

• For knowledge tasks: is the source content current, and does someone own keeping it current?

• Do you know which data is personal, confidential or regulated?

• Are there residency rules (for example, data must stay in the EU or India)?

3. The workflow and people

• Is there a named business owner who will make decisions and accept the result?

• Have you spoken to the people who do the work today, not only their managers?

• Is it clear where a human reviews, approves or overrides the AI?

• Is there time budgeted for the team to test the system and give feedback?

• Do you know which system the output has to land in (CRM, ERP, ticketing)?

4. Risk and governance

• What is the worst realistic outcome of a wrong answer, and who would notice?

• Are there actions the system must never take without human approval?

• Does legal or compliance need to review the use case, and have they been told?

• Do you need an audit trail of inputs, outputs and approvals?

• Do your vendor agreements cover sending this data to a model provider?

5. Delivery and operations

• Is the first release scoped to one use case that can ship in weeks, not quarters?

• Will you build an evaluation set from real cases before choosing a model?

• Who will monitor quality, cost and failures after launch?

• Is there a budget line for running costs (model usage, hosting, maintenance)?

• Do you know what "done" looks like for the pilot, and what happens next if it succeeds?

How to read your score

If you answered yes to 20 or more, you are ready to pilot. Between 12 and 19, run a short discovery sprint to close the gaps first — usually data access and success metrics. Below 12, start with the problem definition: the technology is not the bottleneck yet.