Is your data ready for AI?

Almost every company we talk to has been asked some version of the same question by its board: what are we doing about AI? It is a fair question. The tools have become genuinely useful, and some of the gains are real.

But when we look under the bonnet, the problem is rarely which model to choose. It is that the sales figures live in three systems that disagree, customer records are duplicated, nobody is sure who is allowed to see what, and the most important report is a spreadsheet one person rebuilds every Monday. Put an AI model on top of that and you get confident answers to the wrong question.

This checklist is the one we use when a client asks whether they are ready. It is written for the people who approve the budget, not for engineers.

Why data comes first

An AI model does not know your business. It knows what your data tells it. If that data is incomplete, inconsistent or out of date, the model will repeat those problems quickly and at scale, and it will sound very sure of itself while doing so.

There is also a quieter risk. Connecting AI tools to company data can expose information to people who were never meant to see it. A search assistant that reads every shared folder will happily summarise salary files for anyone who asks, unless someone has thought about permissions first.

The checklist

Work through these seven points honestly. You do not need a perfect score, but you should know where the gaps are before you commit money to a pilot.

1. Every important dataset has an owner

Someone in the business, not only in IT, is responsible for each key dataset: customers, sales, products, transactions. When a number looks wrong, everyone knows who to ask. Without owners, problems get noticed and then ignored.

2. Your key numbers mean the same thing everywhere

Ask three managers how many active customers you have. If you get three answers, you have a definition problem, and an AI tool will use whichever one it happens to find. Agree the definitions of your most important measures, write them down and build reports from that single version.

3. The data is reasonably clean

Nobody has perfect data. What matters is that you know the error rate and it is low enough for the task. Duplicates, missing fields, free text where there should be a list, dates in four formats: each of these is fixable, but only once someone has measured it.

4. It arrives automatically and on time

If the data behind a report is copied by hand, it will eventually be late, wrong or both. Reliable pipelines that load data on a schedule, check it on the way in and alert someone when it fails are the plumbing every AI feature depends on.

5. Access is controlled by role

You can say who is allowed to see which data, and the systems enforce it. This matters more with AI, because a model that can read everything can also reveal everything. Settle access rules before connecting any assistant or model to company data.

6. Privacy and regulation are covered

Personal data must be handled under the GDPR: a lawful basis, a clear purpose, minimal collection and sensible retention. The EU AI Act adds duties that depend on how an AI system is used, with stricter rules for areas such as hiring and credit decisions. If you are regulated by CySEC or a similar authority, expect questions about how AI outputs are checked and recorded. None of this needs to stop you, but it needs to be designed in from the start.

7. You know what it will cost to run

AI features have running costs: storage, compute, model usage and the people who keep it all working. A pilot on a sample of data can look cheap and then grow expensive when it meets the whole company. Estimate the monthly cost at full scale before you decide the business case works.

A quick check you can do today

If you want a rough reading in five minutes, score each statement below as true or false.

Four or five true answers means you are in good shape for a focused first project. Two or three means some groundwork first will pay off. Fewer than that, and an AI project now is likely to become an expensive way of discovering your data problems.

If you are not ready yet

That is normal, and it is good news in disguise. The work that makes data ready for AI also makes every other report more reliable, and it usually pays for itself before a single model is trained.

Pick one use case that would clearly save time or money, and fix the data that use case needs. Do not try to clean everything at once. A narrow, useful first project builds confidence and gives you the patterns to repeat across the rest of the business.

How a data review works

When clients ask us to assess their readiness, we run a short data review. We map where your data comes from and how it moves, measure quality on the datasets that matter, check access and privacy, and look at what the platform costs today. You get a written report ranking the gaps by effort and value, and a realistic plan for the first AI use case, including rough running costs.

Our team holds Databricks platform architect accreditation and has spent years with financial data, where accuracy is not optional. If you would like to know where you stand, read about our data engineering work or tell us about your situation. We will be honest about whether you need us at all.