Before deploying AI
AI inherits every error in your data.
Before deploying artificial intelligence you need to know whether the data that will feed it can carry it. That's what the audit is for: it's the pre-op. We don't deploy AI without it, because a pilot that fails on the data costs far more than checking first.
Why AI projects fail.
Almost never because of the model. They fail because the data doesn't exist or lacks the granularity, because each department defines it its own way, or because nobody has authority over it. The vendor shows the demo on toy data; the problem shows up when it's connected to yours.
The 3-D rule.
- Data: does it exist, with what quality, how much history and which privacy constraints?
- Definition: is there a single way to calculate it, agreed between business and technology?
- Decision-maker: is there someone with authority to decide on it and someone who answers for its implementation?
- An initiative missing any of the three is an orphan: it gets dropped or put in order before a euro is spent.
The Viability Traffic Light.
Every AI initiative lands in one of three boxes — go today, needs foundations or drop it — with the technical reason in writing. The ones that need foundations become a budget line justified with your own data.
Then, who builds it.
Your team, a partner or our engineers. The fractional CDO fee makes sure what gets built is what was decided, and that it complies with GDPR, the Data Act and the AI Act by design.
Frequently asked questions
- How do I know whether my data is ready for AI?
- Run each initiative through three questions: whether the data exists with the quality and history needed, whether there's a single agreed definition, and whether someone has authority over it. If any of the three fails, the AI project will inherit the problem.
- Why do AI pilots fail in companies?
- Mostly because of the data, not the model: data that doesn't exist at the needed granularity, metrics defined differently by each department, or nobody with authority to decide. And vendor demos run on test data, which never has those problems.
- Can I use personal data to train or feed an AI?
- It depends on the legal basis, the purpose and which layer the data is in: identifiable, pseudonymised, anonymised or aggregated. Many use cases work on aggregated or anonymised data, which cuts the risk drastically. The legal side is signed off by your counsel; the technical design to comply with it is part of the work.