Six weeks until the next reporting cycle. The group's ERP data was barely extractable, and the reporting template — modified for years — could no longer be explained by anyone, including prior advisers. German GAAP and IFRS charts of accounts did not match one-to-one, intra-group revenue and cost allocations were invisible, and a genuine consolidation had never been performed at this group of German and French entities.
This was not a grand transformation but a precise intervention in a live process — under time pressure and with a result that had to hold.
The dirt first, then the report
The instinct in a data-cleaning journey is to rush to the clean report. We did the opposite. The first report was a diagnosis of the dirt itself: which postings carry no cost centre, which centres have no name, where do duplicates and phantom accounts hide? Only with that map did we cleanse — fuzzy-matching against duplicates, clustering to identify which customers belonged to the group.
Sound familiar?
If your group figures from different accounting standards don't line up, let's look together in a free first conversation at where the differences come from.
AI takes on the translation work
For the account mapping between the German and French books, AI did the heavy lifting: every account was described in plain language, translated into both languages and matched semantically through a vector database. The result was a similarity index rather than a fragile manual table. The pipelines ran daily and GDPR-compliant over APIs.
What no one could reconcile before suddenly lay clean on the table.
Bank-grade in six weeks
Within six weeks the group had a reporting it could put in front of its banks. German GAAP and IFRS views now switch at the push of a button; differences surface immediately and can be explained line by line. For the first time, intra-group revenue and cost allocations were disclosed properly — the foundation of a genuine consolidation.
Much of the account matching came back one-to-one automatically; what remained was a short, explainable review rather than weeks of manual work. A pragmatic intervention — not a showpiece, but a clean, defensible result delivered on time.
Three things you can take away
- Diagnose before you clean. Rushing straight to the clean report skips the real source of error. Only a map of the dirt makes the cleanup targeted.
- AI as translator, not black box. A similarity index built from plain-language descriptions is more traceable and maintainable than any manually kept mapping table.
- Pragmatism beats perfection. Under real time pressure, a defensible, explainable result matters more than a perfect concept that misses the deadline.