™Automatisation de processus · Industrial operations
Automating a recurring reporting cycle
A monthly report assembled by hand from four sources. Automating the assembly was straightforward; agreeing what the numbers meant was not.
Cette étude de cas n'est disponible qu'en anglais pour le moment.
Le problème
An operations team produced a monthly report by exporting from four systems, combining the results in a spreadsheet and formatting the output manually. The cycle consumed several days each month and always fell at the busiest point.
Because it was manual, the report was occasionally wrong, and because the spreadsheet had been inherited twice, nobody could fully explain how two of its figures were derived.
Démarche
Before automating anything, we reconstructed the derivation of every figure with the team. Two could not be justified and were removed; one was being calculated differently from how it was labelled.
This is the uncomfortable part of process work, and it is usually where the value is.
Ce qui a été construit
An automated pipeline that pulls from the four sources on a schedule, applies the agreed calculations, and produces a draft report.
The draft is not distributed automatically. A named reviewer approves it, with any figure that moved unusually since last month highlighted for attention.
The calculation logic is documented alongside the code, in language the operations team can read.
Technologie et processus
- Scheduled extraction from the source systems with validation on each input.
- Deterministic calculation, no model involvement anywhere in the numbers.
- Variance flagging against the previous period to direct review attention.
- Version history, so a published report can be reproduced later.
Résultat
Assembly time fell from several days to a review session. The team described the bigger gain as confidence: the figures are now defined somewhere other than in one person's memory.
The variance flagging caught a source-system problem in its second month, which would previously have gone unnoticed.
Ce que nous avons appris
- Never automate a calculation nobody can explain. Explaining it first is the work.
- Keeping the model out of the numbers entirely was the right call, the team needed to be able to defend every figure.
- Highlighting what changed is more useful to a reviewer than presenting a clean, complete report.
L'identité des clients et les détails commercialement sensibles ont été retirés. Les résultats sont décrits qualitativement lorsque des chiffres mesurés ne peuvent être publiés.