The Data Model Is the Product

Ask a data team what they shipped last quarter and you'll hear about dashboards, a migration, maybe an LLM experiment. You'll rarely hear "we fixed the model." Yet the model is the thing all of that sits on — and when it's wrong, everything built on top inherits the error, quietly, at scale.

The data model is the product. The dashboard is just its most visible window.

Everything downstream is a view of the model

A metric is a query against the model. A dashboard is a layout of those metrics. A self-serve tool is a friendlier way to write the same queries. An AI agent is an autocomplete for them. None of these things create meaning — they inherit it from the tables underneath. If "revenue" is ambiguous in the model, it's ambiguous in all four, and no amount of front-end polish fixes that.

This is why teams that pour effort into the presentation layer while neglecting the model end up with beautiful dashboards nobody trusts. The trust problem was never in the chart. It was three layers down.

Why modeling gets skipped

Modeling has terrible optics. It's invisible when it goes well and only shows up in the demo as "the same numbers, but now they're right" — which doesn't demo. The work is also genuinely hard: it requires understanding the business, not just the schema, and making decisions that are easy to defer and expensive to revisit.

So it gets skipped. Someone wires a dashboard straight onto raw tables, ships it, and moves on. The model never gets written down — it just lives, half-formed, in the WHERE clauses of forty different queries. Every one of those queries is a small, private fork of "what the business means," and they drift.

What a good model actually buys you

• One definition, reused. "Active user" is computed once, in one place, and everything inherits it. New questions get answered by joining, not by re-deriving.
• Legibility. Someone new can read the model and understand the business. The table names are the documentation.
• Cheap change. When the business shifts, you change the model and the change propagates. You're not editing forty queries and hoping.
• A foundation that compounds. Every clean model makes the next question cheaper to answer. Every shortcut makes it more expensive.

That last point is the whole game. Modeling is the rare investment where the interest compounds in your favor instead of against you.

The industry stopped treating this as an opinion

For years, "the model is the product" was a thing careful practitioners believed and hurried teams ignored. That's changing, and it's worth knowing why — because the change is the strongest evidence the argument was right all along.

In September 2025, Snowflake launched the Open Semantic Interchange — a vendor-neutral, YAML-based specification for how semantic models are defined and shared — with Salesforce, dbt Labs, BlackRock, and RelationalAI. Within months the coalition passed thirty-seven members, including Databricks, Google, AWS, Cube, Hex, and ThoughtSpot. dbt Labs open-sourced MetricFlow under Apache 2.0 as the reference implementation, and the v0.1 spec went live in January 2026. When that many competitors agree metrics need a common language, the question stops being whether the model is the product and becomes how fast you can write yours down.

The forcing function is AI. A dashboard with a wrong number presents it passively — someone eventually notices. An agent presents it confidently, at scale, with an explanation that sounds authoritative. So the platforms have quietly conceded the point: Microsoft's Copilot grounds its answers on the semantic model schema — the tables, measures, relationships, and linguistic metadata — and is retiring the older report-surface Q&A feature in favor of that model-grounded path. Without a governed model underneath, you don't get intelligence — you get confident confusion, automated.

None of this is a new argument. It's the same argument, now with a spec, a license, and a vendor coalition standing behind it.

The model is the house

I've argued elsewhere that an AI agent pointed at a messy schema just industrializes the confusion — it's a new front door on a house with no foundation. The same is true of every other tool. Dashboards, notebooks, semantic layers, agents: they're all front doors.

The house is the model. Build that first, build it well, and everything you hang on it gets better for free. Skip it, and you spend the rest of the project paying interest on a debt nobody remembers taking on.

Here's the test: pick one metric your team argues about — active user, churn, revenue, whatever it is — and go find every place it's actually computed: dashboards, notebooks, that one script finance keeps in a shared drive. Count the definitions. More than one, and you don't have a metric — you have a model that never got written down, living instead in forty different WHERE clauses. That's the debt this piece is describing, and it's already due.

This series is about building the house — starting with the two decisions that shape everything else: the grain and the shape.