AI Readiness Checklist: 30 Questions Before You Start

Most AI readiness checklists ask "is your data ready" and call it a question. It isn't. Nobody can answer that honestly, because it isn't specific enough to have a wrong answer.

A usable checklist asks something narrower: can a model reach this data without a manual export step. Does someone get alerted automatically if a tracked metric drifts. Those have real answers, and the honest ones are often no.

This article is for the CIO or data leader who wants a specific question bank to work through before a first enterprise AI project, not another eight-category framework that sounds thorough and answers nothing.

Key Takeaways

• A checklist question only works if it has a specific, checkable answer. "Is your data ready" isn't one. "Can a model reach this data without a manual export" is.
• These 30 questions make up an AI readiness checklist grouped under the same four dimensions used across this cluster: data, infrastructure, strategy, and governance.
• BCG's research found successful AI adoption is roughly 70% people and process, 20% technology and data, and only 10% algorithms, which is why a checklist skewed entirely toward technical questions misses the largest bucket.
• Answering "no" to a question isn't a failure. It's the specific thing to fix before committing further budget.
• This checklist works as a diagnostic pass, not a pass/fail gate. Most organisations will answer no to several questions, and that's the expected starting point, not a disqualifying result.

Data questions

• Can you name who owns the accuracy of the dataset this system would use?

• Has this data been tested against production volume, not just a curated sample?

• Can a model reach this data without someone manually exporting it first?

• Do two systems ever report different numbers for what should be the same metric?

• Is there a documented answer for how a specific field gets calculated?

• Would anyone notice if this data's quality quietly degraded over a few months?

• Is the data specific to the one use case in question, or scattered across unrelated systems?

• If the data looked wrong tomorrow, would someone notice within a day?

Answering no to more than two or three of these points to data readiness for AI as the dimension to fix first, before any of the others.

Infrastructure questions

• Has inference cost been estimated at expected production volume, not just training cost?

• Is there a tested rollback path if a new model version underperforms the one it replaces?

• Has the data pipeline been tested against full production volume, not a development sample?

• Is a specific metric being tracked automatically once the system is live?

• Does someone get alerted automatically if that metric drifts?

• Is infrastructure scoped to the one proven use case in front of you, or built for an assumed future roadmap?

• Can the current setup scale without a full rebuild if this use case actually succeeds?

This group maps directly onto AI infrastructure requirements, and it's usually where enterprises either underbuild, skipping monitoring entirely, or overbuild, provisioning for use cases that don't exist yet.

Strategy questions

• Can you name the specific decision this AI system is meant to improve?

• Is there a measurable target attached to that decision, not just a general direction?

• Does the use case have a named owner accountable for its outcome?

• Has anyone estimated the value at stake if the use case actually works?

• Would this use case still be worth funding if it took twice as long as planned?

• Is this the first system in a sequence, or a one-off with no plan for what comes next?

• Has this been distinguished from a proof of concept that merely "shows promise" in a demo?

A no on the ownership or measurable-target questions here is usually the reason why AI pilots fail at the proof-of-concept stage rather than the production stage.

Governance questions

• Is there a named owner for this system's output once it's live, not just during the build?

• Has someone been assigned to govern, map, measure, and manage this system, even informally?

• Would leadership notice if this system quietly degraded over the course of a month?

• Has anyone assessed the compliance implications of this specific use case?

• Is there a documented decision-maker for what happens if the system produces a harmful or clearly wrong output?

• Have the people who'll actually use this system been consulted, not just its sponsors?

• Is there a training or change-management plan for the team using it day to day?

• Would someone be comfortable explaining this system's decision logic to a regulator or an auditor?

That second-to-last question matters more than it looks. BCG's research found that roughly 70% of what determines AI success is people and process, not the other 30% combined, which makes the training question the one most checklists skip entirely.

Why Specific Questions Work Better

Most published versions of an AI readiness checklist for enterprises stop at a handful of broad categories: data, people, process, technology. Those categories are correct. They're also too vague to actually answer.

"Is your data ready" invites a comfortable yes from almost anyone asked it, because there's no specific claim to check it against. "Can a model reach this data without a manual export step" doesn't have that problem. It's checkable, and checking it honestly is the entire point of running this exercise at all.

This version functions less like an AI readiness assessment checklist in the abstract and more like the actual question bank behind one. It's grouped under the same four dimensions this cluster scores elsewhere, just broken down into the specific, checkable version of each. The same structure holds whether you're running it as an AI readiness checklist for organizations already mid-rollout or as a first pass before a single pilot.

How to use this AI readiness checklist

Work through all four sections honestly, and count the no answers per section rather than overall. A section with three or more no answers is the dimension to fix first, not the whole list at once.

This checklist is the expanded version of the same four dimensions scored in an AI readiness assessment. Where the assessment gives a number, this gives the specific question behind each point of that number, which is useful when the score alone doesn't say what to actually go fix.

Run it before scoping a specific investment, not as a one-time exercise for the organisation as a whole. The honest answers to these questions change based on which use case and which dataset is actually in front of a team, so a checklist run for one proposed project won't necessarily hold for the next one.

Common mistakes when using this checklist

Three patterns undermine this exercise most often.

1. Treating It as Pass or Fail

What it looks like: teams stop at the first no answer and conclude the organisation isn't ready for AI at all.

Why it happens: a checklist format invites binary thinking, even though readiness is a spectrum, not a gate.

How to fix it: count no answers per section, and treat each one as a specific fix rather than a verdict on the whole initiative.

2. Skipping the Governance and Strategy Sections

What it looks like: teams work through the data and infrastructure questions carefully, then skim the strategy and governance sections because they feel less technical.

Why it happens: technical questions have obvious owners. Governance and strategy questions often don't, which makes them easy to defer.

How to fix it: assign an owner to each section before starting, including someone accountable for the governance and strategy questions specifically.

3. Scoring It Alone Instead of Cross-Checking

What it looks like: one person answers all 30 questions based on their own view of the organisation, without checking with the teams actually touching the data or the system.

Why it happens: it's faster, and the person running the checklist often assumes they already know the answers.

How to fix it: where the internal answer feels uncertain, bring in AI readiness services to check the honest answer against an outside view.

In our experience, the governance section gets the most inflated answers when someone scores it alone, since "someone would probably notice" is a very different claim from "someone is specifically assigned to notice."

Let's Sum Up!

Thirty specific questions will tell you more than eight broad categories ever could, because each one has an actual answer instead of a comfortable generalisation. Work through all four sections, count the no answers per section, and fix the one with the most before touching the others.

This is one piece of the broader picture of AI readiness for enterprises, the same four dimensions in question form rather than score form. Classic Informatics has run this exact question bank against real enterprise AI proposals more than once, usually surfacing one dimension nobody had actually checked. Happy to run it against yours if it would help.