Nobody writes blog posts about field mappings. We do.
Field notes on the layer nobody demos: getting records out of SAP, MES and PLM in a form a retrieval system can actually use. This hub collects the extraction paths, the masking of personal data inside the pipeline, and the permission model that has to survive retrieval for an answer to pass a security review.
The layer everyone skips and then rebuilds.
This area is about making enterprise data reachable: reading SAP, MES, PLM, SharePoint and thirty-year-old SQL schemas, working out which field is actually authoritative when four of them disagree, masking personal data before anything is indexed, and carrying the real permission model through to retrieval.
It is written for the people who will do that work — data engineers, integration architects, and the one person in the company who still understands the custom Z-tables.
The argument running through this area is that AI projects rarely fail on the model. They fail because a demo was built on a curated folder of two hundred clean documents, and the production corpus turned out to be four million files with three naming conventions, no revision discipline and an access-control list that nobody has audited since 2019. Fixing that is not preparation for the project. It is most of the project.
Common questions
Why does the data layer decide whether an AI project ships?
Because the model is the interchangeable part. Retrieval quality, field-level authority and entitlements determine whether an answer can be trusted, and none of those improve when you swap in a better model. A pilot that skips this layer works on a curated folder and stops working on the real corpus.
Do we need a data warehouse before any of this works?
Usually not. A warehouse is optimised for aggregates, and most AI use cases need the specific record with its context and its permissions attached. We generally read from the source system or a replica and build a retrieval layer beside the warehouse rather than waiting for one.
Not a sales call. An architecture call.
Thirty minutes with the architect who would actually run the engagement.