Three distributors, two ERP migrations, and data nobody could use.

Two ERP migrations is where operational data goes to die. Fields get remapped, meanings drift, historical records arrive in a format nobody planned for, data gets dropped, and the reconciliation that should have happened gets deferred until it's someone else's problem.

For these three distributors, the result was that the business could only be examined at the highest level — by vendor, and not much below it. Segment-level questions couldn't be answered, which meant the questions that actually drive profitability weren't being asked. Delivery was the clearest example: they were charging for it and paying for it, and had no reliable view it's volume or potential.

We enriched and restructured their data so it could be segmented the way the business is actually run, then built analysis on top of it — including the delivery fees and pricing segmentation that had been invisible.

The core of the work was classification. The fields that would have answered their questions didn't exist in the data, so we inferred them from transaction-level records — deriving the segmentation the business needed from what the transactions could tell us.

Delivery was the clearest case. There was no classification of what went out on the company's own trucks versus what shipped another way, or in what volume. That had to be reconstructed from transaction patterns before anyone could ask what self-delivery actually earned. Once it existed, it was answerable for the first time since migration.

Built on BigQuery, with Tableau for data visualization to present to the people making decisions rather than sitting in a data lake. Delivered in weeks.

Making the segmentation match how the business operated and understood itself.

This sounds like a technical problem and isn't. In most cases, the categories that exist in an ERP are the categories someone configured during implementation, and they frequently don't correspond to how the people running the company actually think about their product lines, their customers, or their margins. In this case, product segmentation was lost entirely in migration. Getting that right meant working out how the business genuinely viewed what was happening and then building the data to reflect it — not the other way around.

Inferring a field the system never recorded means deciding what counts as evidence, and then being right often enough that people trust the resulting numbers. For the delivery classification, that meant working out which transaction patterns reliably indicated an own-truck delivery — and validating that against what the business suspected to be true, because a segmentation that's confidently wrong is worse than one that doesn't exist.