AI enablement
AI is only as trustworthy as the layer underneath it.
Most organizations buy the assistant and skip the layer it reads from, then wonder why two people asking the same question get two different numbers. I build the layer.
The ontology is the product. The assistant is the interface.
A language model asked a business question does not fail because it reasons badly. It fails because nothing told it which table is authoritative, which join is legitimate, or what the business calls a metric this quarter. Define that once, in a governed semantic and metrics layer, and the same model starts answering correctly. Skip it, and you have automated the disagreement.
Assistant on raw tables
Three teams ask the same question. Three numbers come back, and the argument is about the data instead of the decision.
Assistant on a governed layer
One definition of the metric, one set of joins the assistant is allowed to make, one answer. The layer did that, not the model.
What I built on it, how I run my own work the same way, and the limits - below.
What I built on it
Two Databricks Genie assistants, in a global rare-disease pharma organization, sitting on the curated semantic and metrics layer I own. They work because of the layer, which is the entire point.
Web-analytics reporting assistant
Self-serve conversational analytics for people who should not have to write SQL to get a number. Curated entities, canonical metric definitions, synonyms for the words the business actually uses, and trusted example queries, so the same question returns the same answer to whoever asks it.
Governance assistant
Checks new and existing tables against the standard: naming conventions and column comments. Governance holds when conformance is cheap to check and expensive to skip, and this is what makes it cheap.
The same discipline, applied to my own work
I run my own analysis, research and writing through an agentic, file-based system on Claude Code. What makes it useful is not that it drafts quickly. It is that it is governed: work enters one door, is routed to a specialist role, and nothing changes state until it clears an approval gate. Every approved action is written to an append-only log, every fact lives in exactly one file, and nothing reaches another person without me sending it.
Anything that leaves the system exits only through a person.
That is the same argument as the section above, one level down. A capable model plus an ungoverned workspace produces confident work nobody can audit. The gates and the record are what make the output worth using.
Why a semantic layer beats prompt engineering
A prompt is a per-question fix. It lives in someone's head or someone's chat history, it is not versioned, and it does not survive the person who wrote it leaving. A semantic layer is a per-organization fix: the definition of active patient, or qualified session, or net revenue is written once, owned by someone, versioned, and read by every assistant, dashboard and query that touches it.
The practical test is boring and decisive. Ask the same question three different ways and see whether you get the same number. If you do not, no amount of prompt tuning fixes it, because the ambiguity is in the data model, not the wording.
This is also why the governance work and the AI work are not two separate things on this site. They are the same work, and the assistants are just the part you can see. How the foundation is built
What I do, and what I do not
I do: build the governed semantic and metrics layer that conversational analytics reads from, define and own the metric definitions behind it, stand up assistants on top of it, and set the standards and stewardship that keep it true. I build hands-on, not by directing a vendor.
I do not: train or fine-tune models, build serving infrastructure, or do machine-learning research. My value is upstream of that, and pretending otherwise would be the first untrustworthy thing on this site.
For the statistical and machine-learning modeling I do build, including a decision tree you can watch fit itself, see analytics and ML