- In the AI projects we take over, the model is rarely the problem. The data is, in the vast majority of cases.
- Four defects recur: data that cannot be found, data that contradicts itself across systems, no owner, no history.
- Four weeks are enough to know whether a project is feasible, and at what cost, before committing a budget.
- The deliverable that matters is a table of sources with, for each, an owner, a quality level and a decision.
What we find when we arrive
The scenario repeats. A company launched an ambitious AI project, often with a competent contractor. The prototype worked on a curated sample. Then, at the moment of scaling, everything seized up: production data does not look like the sample, two systems give different figures for the same customer, nobody knows who is allowed to correct a value, and the history needed to train or evaluate only goes back six months.
The project is then paused "while we clean the data", which, without a method, never ends. It is not fate; the order of operations was wrong.
The four defects that kill a project
Data that cannot be found
It exists, somewhere: in a monthly export on a shared drive, in an old application nobody opens any more, in the emails of someone who left last year. The project needs regular, reliable access, not a treasure hunt.
Contradictory data
The CRM counts 4,200 active customers, billing counts 3,800, the marketing tool 5,100. Each is right by its own definition. Until the company decides which one is authoritative, no model can learn anything consistent.
No owner
When a value is wrong, who fixes it? If the answer is "IT" or "nobody", quality will never improve. Every data source needs a named business owner who answers for its quality.
No history
Forecasting sales requires years of sales. Detecting anomalies requires examples of anomalies. If the history does not exist or was overwritten at each update, the project must start by collecting, and that takes the time it takes.
"Show me last month's real data, as it is, without cleaning." What you see in ten minutes is worth more than thirty pages of scoping.
The four-week plan
| Week | Goal | Deliverable |
|---|---|---|
| 1 | Inventory of the sources the use case needs | List of sources, access, frequency, volume |
| 2 | Quality measurement on real data | Empty fields, duplicates, inconsistencies, depth of history |
| 3 | Owners and definitions | One name per source, one arbitrated definition per indicator |
| 4 | Decision | Feasible now, feasible after (costed) remediation, or to be postponed |
The final deliverable fits on one page: a table of sources with an owner, a quality level (good, acceptable, blocking) and a decision. A board can read it in five minutes and decide. That document is what almost every project we take over is missing.
What governance really changes
The word "governance" scares people, rightly when it means a committee and a binder. What we put in place is simpler: one owner per source, one definition per indicator, an automatic quality check that raises an alert when a threshold is crossed, and a thirty-minute monthly review. This light setup does more for an AI project than any choice of model.
"You do not clean data for a project. You organise the company so that data stays clean."
Conclusion
AI projects do not fail because AI is immature. They fail because they started from the end. Four weeks of inventory, measurement and arbitration before any development avoid months of drifting and make it possible to say, with figures, whether a project is reasonable. It is less spectacular than a demo, and it is what separates projects that deliver from projects that are forgotten.
FAQ
Do we need a data warehouse before doing AI?
Not necessarily. For a first bounded use case, clean access to two or three sources is enough. The warehouse becomes useful when several projects share the same data.
Who should own a data source?
The business person who depends on it most and has the authority to get an error fixed. Rarely IT, which hosts the data but does not know its meaning.
How much does remediation cost?
It depends on the defect. Deduplicating a customer file takes a few days; reconstructing three years of history may be impossible. That is exactly what the four-week diagnosis costs out.
Can we do generative AI without this work?
For an assistant over well-organised documents, yes, with modest expectations. For anything touching structured business data, no: the model will answer confidently from wrong data.
