Why most enterprise AI pilots stall at the semantic layer

Six months in, the slide deck still says 'AI-powered.' The model card is fine. The dashboards almost work. And yet the project has quietly stopped moving — adoption is flat, the steering committee is patient, and the next quarter's budget is being negotiated as a 'consolidation' phase.

I have walked into this room more times than I can count. The diagnosis is almost always the same, and it is almost never the model.

The bottleneck is the semantic layer. Nobody in the building has agreed on what a customer is, what an order means once it is partially refunded, or which of the seven 'revenue' fields the CFO actually trusts. The model is hallucinating because the data is hallucinating first.

The framing that unlocks these engagements is brutally unfashionable: AI is a downstream beneficiary of governance, not a substitute for it. If your conceptual model is fuzzy, your logical model will leak, your physical model will drift, and your LLM will confidently cite the leak.

Senior leaders rarely want to hear this. The promise of AI is that it skips the boring work. The reality is that AI makes the boring work load-bearing. A clear glossary, a curated entity catalog, and an MDM spine are not a tax on AI — they are the foundation that lets AI compound instead of embarrass.

Five recurring failure modes:

1. Definition drift — 'active customer' means three different things in three different dashboards, and the model has been trained on the average of all three.

2. Source-of-truth ambiguity — every domain claims authority over master entities; nobody owns stewardship.

3. Glossary debt — the business glossary is a Confluence page from 2023, half-deprecated, never enforced at the data layer.

4. Pipeline-first thinking — engineering ships pipelines before semantics, then patches semantics with one-off views.

5. RAG over chaos — retrieval-augmented generation is layered on top of a knowledge base nobody curates, so the model retrieves, faithfully, the wrong thing.

The reframe I push: the semantic layer is the product. The LLM is a feature of the product. Adoption follows trust, trust follows a single coherent definition surface, and the definition surface is the deliverable that justifies the AI budget.

Operationally that means three commitments. Conceptual model first — agreed by the business, not just by data engineering. Logical model published in version control with diffs everyone can read. Physical model that is a disciplined implementation of the logical model, not a parallel universe.

Days 0–14 — entity audit. Inventory the top 10 entities the business actually decides on. Cluster the 30+ definitions into a single canonical list, with deltas annotated.

Days 15–45 — glossary + stewardship. Stand up a working glossary with named owners for each entity. Make 'unowned definition' an open ticket, not an anonymous footnote.

Days 45–75 — semantic platform. Land a thin semantic layer (dbt + a metric store, or an MDM spine) that publishes the canonical definitions to consumers — BI, RAG, ML features.

Days 75–90 — re-pilot. Restart the AI use case on the new semantic layer. Measure adoption against the pre-cleanup baseline, not against the original promise.

A canonical entity catalog with 8–12 entries, each with a one-paragraph definition, a steward, and a contract test.

A metric store or semantic layer that BI, RAG, and ML feature pipelines all read from — no parallel SQL.

A weekly definition review with the business — the meeting where 'should we add this?' is answered before engineering ships it.

A retrieval index built on top of curated semantic objects, not raw documents — RAG queries the catalog, not the chaos.

The teams I have seen ship AI at scale share one trait: they treat semantics as a craft. Their definitions are short, their stewardship is named, and their LLMs cite a small set of curated entities — not the entire data lake.

If your AI pilot has stalled, the answer is rarely 'a better model.' It is almost always 'a smaller, sharper definition of what we are talking about.' That is unfashionable, expensive in calendar time, and the only thing that compounds.