AI in production
From pilot to a system the business can operate: prioritised use cases, deployment and risk control.
The pilot worked. Nobody uses it.
It is the most repeated story of the last three years. The demo impressed the committee, the budget was approved, and six months later the system is switched on and empty.
The model is not what failed. What failed is underneath it: data with no owner, no lineage and no quality, and an organisation nobody told what to do differently on Monday.
From pilot to system.
What separates a demo from something the business can still operate next year.
- 01Prioritise by value and by available data — A valuable use case whose data does not exist is not a use case: it is a data project in disguise. We separate the two and sequence them.
- 02Prepare the data — This is where the real work goes, and the part no demo ever shows. If this phase looks short, the project has already gone wrong.
- 03Deploy with control — Traceability of what was answered and from what; explicit limits on what the system may decide alone; and human review wherever a decision affects a person.
- 04Operate — Who maintains it, how you measure that it still helps, and under what condition it gets switched off. A system with no shutdown criterion has no owner.
Why this order.
Because it is the only one that holds. Artificial intelligence is the layer that reaches the budget, and also the only one that does not work on its own: it rests on applications, which rest on data, which rest on what the business actually does.