Your internal data is worth more than the AI you put on top of it

[PLACEHOLDER ESSAY] This is a stub. The published version is on the launch shortlist in content/03_thinking.md and should be written before the site ships. Outline below — replace with real prose.

Thesis

Most companies have terrible internal data and great external AI. The leverage is in the data layer — the AI is the easy part. Invest in semantic clarity, audit trails, and queryability before you invest in models.

Outline

1. The temptation: a CTO sees ChatGPT, signs an RFP for a “GenAI initiative,” then discovers their warehouse is a graveyard of half-documented dim_* tables.
2. What “data layer” actually means in this context: curated views, semantic naming, row-level boundaries, audit trails.
3. The SQL gateway example — the curated view layer is ~80% of the work; the LLM is ~20%. Walk through where the time actually went.
4. A concrete failure mode: the LLM that answers “how many active users” five different ways because there are five definitions of “active” in the schema. The model is fine. The data is not.
5. What to do first: a semantic layer pass. A definition pass. An access-pattern audit. Then a model.
6. What to do never: tell the model to “be careful” about the data and consider the problem solved.

Hook to write

A short, real (anonymized) example from one of our engagements where the data work made the model unnecessary, or where skipping it produced an answer the client trusted by accident.

---

Stub published 2026-05-25. Replace before the site is launched publicly.