An industrial method,
not a series of workshops
Ten steps, sixteen deliverables, twelve control domains, and acceptance criteria written before the first access to data. The method is the same for a bank pilot and for a public administration: what changes are the use cases.
Scope before buying
The technology decision follows requirements and the benchmark. No hardware is ordered before the proof of concept.
Prove it on your data
Comparative tests use your data, your easy and hard questions, and your security constraints.
Willing to say no
The benchmark report may conclude “do not continue”. That is written into the contract — and it has happened.
What makes a project succeed — and what makes it fail
Success factors
- A sponsor at executive level, not only in IT.
- Use cases chosen for measurable value, not for ease.
- Named data owners accountable for quality.
- A deliberately small pilot scope.
- Users trained and involved from the design stage.
- A run budget planned from the outset.
Warning signs
- “Give us a free demonstration first.”
- An AI project handed solely to the IT department, with no business owner.
- A hardware budget committed before the benchmark.
- A scope that grows at every meeting.
- No value indicator defined at the start.
- No data governance in place.
Responsible AI, in practice
- Every answer is traceable: question, executed query, data consulted, model used.
- Sensitive data does not leave: filtering before any external call, or fully sovereign deployment.
- Critical actions require human validation — without exception.
- Entitlements mirror those of the existing information system rather than being recreated.
- Model consumption is measured, capped and alerted on.
- Answer quality is tested at every change, not only at delivery.
Skills transfer included
Two deliverables are dedicated to it (guides and training). The objective is not to make you dependent: it is for your teams to administer the platform and for us to be able to step away.