AI readiness: the honest checklist before you commit budget
Most organisations are more ready than they fear on the technical side and less ready than they think on the decision side. Readiness is mostly about clarity, not infrastructure.
By Quality AboveAll · · 8 min read
- A specific use case with a measurable outcome matters more than any technical prerequisite.
- You need the data the first feature touches to be usable, not an enterprise data platform.
- If nobody can say what success looks like in numbers, that is the gap to close before writing code.
Use case clarity comes first
The question that predicts success is not "do we have the right stack" but "what decision or task will be different, and how will we know". Teams that can answer that in one sentence tend to ship; teams exploring AI as a category tend to produce demos and stall.
Test your own clarity by writing down the specific task, who does it today, how long it takes, and what an improvement would be worth. If that paragraph is hard to write, the readiness problem is right there, and it is a cheaper problem to fix now than after a quarter of engineering.
Data readiness, scoped properly
Assess the data your first feature needs, not your organisation's data estate. If you want an assistant over your help centre, the question is whether that help centre is current and consistent, and nothing about your warehouse strategy is relevant yet.
The disqualifying problems are contradiction and missing permission metadata, as covered in preparing your data for AI. Messy formatting and mixed file types are tolerable. Distinguishing the two saves months of unnecessary preparation.
Technical foundations that actually matter
You need somewhere to run a service, a way to call an external API, the ability to deploy changes with reasonable frequency, and basic observability. That is a low bar, and most organisations clear it.
The genuine blockers are contractual and architectural: a policy that forbids sending data to third parties with no on-premise alternative budgeted, or a monolith where no new service can be deployed without a quarterly release train. Both are solvable and both need to be known before commitments are made.
Nobody has ever been blocked from shipping an AI feature by their choice of framework. Plenty have been blocked by a release process measured in quarters.
Skills and ownership
You do not need a research team. You need engineers comfortable integrating APIs and thinking probabilistically, and one person who owns the outcome and can decide what "good enough" means when the answer is 92 percent right.
The commonly missing role is that owner. Projects with an engaged product owner who reviews outputs weekly progress; projects where AI is a side task for three people converge on nothing. Team shapes are covered in structuring an AI product team.
Governance before you need it
Decide in advance what the organisation's position is on data leaving your control, on disclosing AI use to customers, on human review for consequential outputs, and on who signs off a release. These do not need to be elaborate, they need to exist.
Frameworks like the NIST AI Risk Management Framework are a reasonable scaffold to adapt rather than adopt wholesale. The practical version fits on a page, and having that page is what stops a launch stalling in legal review a week before it was due.
Frequently asked questions
What is the single biggest readiness gap you see?
Use case vagueness. Technical prerequisites are usually met; the ability to state what success looks like numerically is usually not.
Do we need a data platform first?
No. Scope data readiness to your first feature's corpus. Broader platform work should be justified by what the first features reveal.
How long does a readiness assessment take?
A focused one takes days, not weeks: interview the people doing the work, look at the relevant data, and check the deployment and policy constraints.
Want an outside read on whether you are ready and where to start? A free 30-minute consultation will give you a straight answer, including if the answer is not yet.
