The data layer most teams skip

July 2026

Every org I have worked in — a $5B satellite program, a directed-energy startup, a lighting company, a submarine — eventually hits the same wall. The dashboards disagree. The AI agent hallucinates. Two systems both claim to know how many parts exist, and they are both wrong.

The wall is never the dashboard tool or the model. It is that nobody decided, precisely and in writing, what the underlying things are.

"A part" is not a number

My favorite example, because it cost real money everywhere I have seen it: a 185-line bill of materials is 185 unique items, not a quantity of 185. ERP thinks a BOM line is an item. MES thinks it is a count. Sum across the two and you get numbers that look plausible and mean nothing. The fix is not a smarter report — it is a schema-level convention for unique-part-versus-unit-count that every downstream consumer inherits whether they think about it or not.

That is the job. Not pipelines. Definitions that pipelines cannot violate.

What the skipped layer actually contains

• Raw structures and schemas — how data is shaped at the source, with constraints that make invalid states unrepresentable.
• Entity relationships — what points to what, with keys that survive systems disagreeing.
• Metadata and lineage — every record self-describing and traceable to source, so "where did this number come from" has a boring answer.
• A governed semantic layer — meaning captured once, consumed everywhere, so operational and analytical views stay consistent.
• Access as attributes — ABAC-style rules on user, resource, and mission attributes, decided at the data-product level, not per-dashboard.

Sonar taught me this before Foundry did

On the submarine, a contact was an entity with a position, a time, and a lineage of observations. If your track history was sloppy, you did not have a data quality problem — you lost the contact. There was no meeting where we decided whether data definitions mattered.

I now build the same discipline into enterprise data: spatial-temporal meshes of domain-owned, versioned data products where entities carry position, time, and lineage. Then, and only then, AI on top — retrieval over governed datasets is delightful; retrieval over vibes is a liability with a chat interface.

Skip the layer and every project above it inherits the debt. Build it and the dashboards stop arguing, the agents stop inventing, and the org starts trusting its own numbers. That trust is the actual deliverable.