Most AI projects start with the tool or the use case. The data question comes later, often at the very moment the project should be producing its first results. The reverse order is the one that works: first check that the data can answer the question being asked, then choose how to answer it.
The question also has to be asked at the right level. There is no such thing as data maturity in the abstract. An organisation can have flawless invoicing data and unusable logistics data. AI readiness is therefore assessed use case by use case: for a given decision, is the right data available, in the right state?
The ten questions below are designed for exactly that. They require no particular technical background: a business owner and the person who knows the source system can answer them together in a single meeting.
The business question
1. Which decision should the project improve, and who makes it?
A useful AI project changes a decision: chasing a customer, resequencing production, adjusting a cash forecast, routing a request. If you cannot name the decision and the person who makes it, there is no use case yet, only an intention.
2. How is that decision measured today?
Without a current indicator, there is no way to tell whether the project adds anything. On-time delivery rate, actual payment delay, request handling time: the existing measure serves both as a baseline and, often, as the target the model will learn to predict.
The data available
3. Does the data exist, and where does it live?
List the sources: ERP, CRM, data warehouse, shared files, documents. Data that only exists in manually consolidated spreadsheets is not ruled out, but it calls for preliminary work that should be scoped and costed from the outset.
4. Is there enough history, and does it cover the situations to anticipate?
A model learns from the past. The history must be long enough to cover the business cycles (seasonality, closings, peaks), and you need to know where the breaks are: an ERP change, a reorganisation, a new catalogue. A long history interrupted by a migration is sometimes worth less than a short, consistent one.
5. Are the definitions shared?
What is an order delivered on time: against the requested date or the confirmed date? What is an active customer? If two departments answer differently, the model will learn an average of two definitions. Stabilising the vocabulary is often the project’s first deliverable.
6. Can you isolate the populations that behave differently?
Credit notes, returns, partial payments, unallocated entries, one-off orders, test data: not every row in a table tells the same story. If they cannot be identified and set aside through justified, traced rules, they blur the learning. This is the step that weighs most on the result, and the one most often underestimated.
Quality and accountability
7. Who is accountable for this data?
When a value is wrong, who corrects it, and is it corrected at the source or in an export file? Data without an owner degrades, and a model fed by degrading data degrades with it. The point is not to launch a full governance programme, simply to name, for the scope of the use case, who is responsible for what.
8. Is quality measured, or discovered through incidents?
Missing-value rates, duplicates, inconsistencies between systems: a handful of indicators tracked over time is enough to know whether the data can be trusted. If they do not exist, the first weeks of the project will be spent producing them, which is a good thing provided it was planned.
9. Are you allowed to use it for this purpose?
Personal data, confidential information, contractual clauses, third-party documents: some data cannot be used for any purpose or sent anywhere. The question is especially pressing with generative AI, which handles entire documents and often relies on hosted services. It is better settled with the relevant people in the company before building than discovered afterwards.
Operation
10. Where will the answer be delivered, and who will keep it running?
A prediction nobody sees changes no decision. It has to reach the place where teams already work: the dashboard they check, the report they open, the tool they use. You also need to know who will monitor the model once in service, because data and behaviours change over time, and the quality of a prediction changes with them.
Reading the answers
The aim is not ten perfect answers. The answers mainly tell you where to start.
- Questions 1 and 2 remain unanswered: there is no use case yet. Go back to the business decision rather than launching an experiment.
- Questions 3 to 6 reveal gaps: data preparation is the first workstream. This is not a project failure; it is often the most valuable part, and it benefits uses well beyond AI.
- Questions 7 to 9 raise issues: governance work is needed, limited to the scope of the use case and run in parallel.
- Question 10 has no answer: the risk is delivering an accurate model that nobody uses. Delivery is designed from day one.
- The answers are broadly positive: a proof of concept can be scoped, with its perimeter, duration and success measure defined before it starts.
What we see on assignments
In our predictive projects, the difficulty rarely lies where it is expected. At an industrial equipment manufacturer, the order history already existed in the Operations datamart; the work consisted in defining the exclusion rules with the business, then delivering the prediction in the dashboard the teams already used, next to the actual indicator. In the finance department of an industrial group, settled and open invoices had to be separated, and partial payments, credit notes and unallocated entries excluded, before any modelling. In both cases, questions 5, 6 and 10 made the result.
This is why our Data & AI Assessment evaluates the data readiness of each selected use case, with a value / feasibility matrix and a roadmap. For a use case already identified, the two-to-three-week scoping described on the Artificial Intelligence & Automation page ends with a scoping note and a reasoned decision to launch, or not, a proof of concept. And when the data still lives in manually consolidated spreadsheets, Data Ready Express sets up a centralised, automated data foundation in ten days to build on.
Useful AI starts with reliable data. Answering these ten questions honestly costs one meeting; skipping them often costs a proof of concept.