A business tool is only worth as much as the data it consumes. For a site selection consultancy, that data comes from reference datasets: companies registered in France, the places and shops on record, the mapping of roads and public facilities.
These datasets are public, large, and updated regularly by the bodies that publish them. Loading one for the first time is a few days of work. Keeping it current for three years is a different job, and it is the one that decides whether the tool stays trustworthy.
What the chain does
For each dataset, an automated job downloads the new release, checks it, lands it in the data warehouse, and then feeds the application databases behind the business tools.
The checking is the part that matters. A public file can arrive truncated, change format without warning, lose a column, or halve in size because of an incident at the source. Without verification, those faults spread quietly all the way into the studies delivered to clients.
So the jobs apply blocking checks before anything is made available: volume compared with the previous release, expected structure, consistency of identifiers, share of missing values. A release that fails a check is not published, and the previous one stays in service.
Versions rather than overwriting
Every load keeps the previous version. It sounds trivial and it solves two concrete problems.
First, you can roll back if a fault is found later, without waiting for the source body’s next publication. Second, you can compare two versions, which makes real movements visible: openings, closures, relocations. That comparison becomes information in its own right, sometimes more useful than the dataset itself.
The monthly watch
The whole chain runs unattended on the cloud. A monthly check confirms that each job ran, that the volumes are consistent, and that no step is sitting in a failed state. Anything abnormal raises an alert addressed to a named person.
It is a modest discipline and it is what separates useful automation from dangerous automation. A job that stops without saying so leaves everyone believing the data is current when it is six months old. By the time somebody notices, trust has to be rebuilt on top of the work.
What transfers
The same architecture applies to any regular data source: a file received from a partner, an export from a business system, a supplier catalogue, a price feed.
The useful rule is simple. Data that feeds a decision needs a job that refreshes it, checks that reject it when it looks wrong, and an alert that tells someone when the chain stops. Without those three things it is not a pipeline, it is a debt.
The end client is not named. The figures quoted are those of the engagement.