Data readiness for AI: how to find out where your organisation stands

Why AI projects fail on data, not on technology

In July 2024, Gartner asked 1,203 data management leaders whether their organisation has the right data management practices for AI. Sixty-three per cent said no, or did not know. In the projects we see, three data problems keep coming back.

Data you cannot rely on

Missing values, inconsistent formatting, duplicate records, information that is out of date. A rule or a model that runs on polluted data produces polluted results, and nothing in the output shows that the input was wrong. That is what makes this the problem you pay for later rather than catch early.

You know the symptom: figures in your Power BI reports that do not add up, reports that contradict each other. Usually the dashboard is not the problem; the data underneath it is.

Data that lives in separate systems

Finance works in the ERP, sales in the CRM, operations in Excel. That is not sloppiness: Excel ended up there because it worked. The problem starts when the three do not talk to each other.

An application that has to predict customer behaviour cannot do it if the same customer is called something different in three systems and the three are never synchronised. What you build on that works in the demo and stops working a month later.

Data that nobody owns

Ask who is accountable for the quality of your product data. In the projects we see, that is the question without a clear answer: data gets used by everyone and owned by no one. Without an owner your data quietly degrades: definitions drift, formats diverge, and nobody decides which version is right.

That cost lands on someone. It lands on the person who rebuilds the monthly report by hand. And it lands on the data or IT lead who has to explain at the next management meeting why two systems disagree.

In 2024, RAND Corporation looked into why AI projects fail, drawing on 65 interviews with people who do the work. They identified five root causes. Data comes second on that list. Of the five, exactly one is technical, and it sits at the bottom.

These three problems are solvable. You do have to know which of them you have before you invest in AI.

What data readiness for AI actually means

Ready does not mean perfect. No organisation has perfect data, and waiting for it is a way of never starting. It means your data is good enough to let an application work reliably on the questions you actually want answered.

Whether your data is ready for AI comes down to four things: quality, accessibility, governance and integration. Read them as four questions about your own organisation rather than as a model.

1. Quality. Are your records correct, complete and current? The test is simple and uncomfortable: when your figures and a colleague’s figures differ, do you know which of the two to believe?
2. Accessibility. Can the people who need data find it without asking IT first? If every question about last quarter goes through the one person who knows where things are, your data is not accessible. It is accessible to that person.
3. Governance. Is someone accountable for the quality of your data, and are there rules for how it gets entered, stored and used? Governance sounds like something for a large organisation with a chief data officer. In the SMEs we work with, it comes down to naming an owner per data domain and writing down the handful of rules everyone follows.
4. Integration. Do your systems talk to each other, or does someone merge them by hand every month? Every manual merge costs time and carries a fresh chance to introduce an error.

None of the four stands on its own. They pull at each other: data that nobody owns drifts in quality, and data that is hard to reach gets copied into a spreadsheet where no rule reaches it.

If you score well enough across these four dimensions, your data can support an AI application. That is not a guarantee that the project will succeed: the use case, the people who have to work with it and the ongoing management decide the rest.

Which leaves the one question the four dimensions cannot answer for you: where does your organisation stand on them today?

Count your yes answers.

Wherever you land, the count tells you where to look first. It does not tell you what to buy.

Five questions give you an indication, not a verdict. There is a longer, ten-question self-assessment that goes further, and it is the better instrument when your answers were close calls.

Land in the lower two ranges and the next question is which gap to close first, and in which order. That is what a Data & AI Readiness Assessment is for. In two weeks you know where you stand: an honest picture of your data maturity and a concrete action plan.

If you answered yes four or five times, an assessment with us is probably not your next step. Pick a use case and get going. And the first steps of this route you can take yourself, without us.

Four steps to data readiness

Four steps: map what you have, pick one domain to start with, get the quality right there, and connect it to the rest afterwards. In that order, because each step leans on the one before it. The first steps are organisational work, which is why you can begin without a data team.

1. Map what you have

Begin with an inventory. That exercise takes a week or two and needs no technical expertise: you map which systems you have, which data sits where, and who uses what.

The inventory is something you make yourself. Our assessment goes deeper: interviews, a technical review and a report. Both start in the same place: knowing what you have.

2. Pick one domain

Not everything at once. Choose the data domain with the most pain or the most potential: customer data, product data, financial data. A broad programme never finds an owner, because everyone carries a piece of it and nobody carries the result. A narrow one does, and that is the difference between a plan and a change.

3. Get the quality right there

In that one domain, make the data correct, complete and consistent. Appoint an owner: a named colleague who decides what a valid record looks like and who gets asked when two systems disagree. Agree on rules for entry and maintenance, and keep them short enough that people follow them.

This is organisational work rather than technology work. For a data or IT lead who has to build the case internally, that is the strongest argument on the table: the first improvement costs agreements, not a platform.

4. Connect it and take the next step

Once one domain is in order, connect it to the other systems, so there is one place where the right figure sits and everyone reads the same number. That is a single source of truth in practice, and it can be a modest data platform rather than a multi-year warehouse programme. An integration is not a one-off exercise either: it needs maintaining, and that maintenance belongs in the cost from the start.

Then, and only then, does an AI investment make sense for the heavier use cases. We guide SMEs through this route, and the assessment is where it starts.