The AI-Native Data Warehouse: Keep Truth Atomic, Make Aggregation Disposable

Last week, I led a conversation with Databricks product managers and their team about where Databricks should go next as AI changes how people interact with data.

One idea continued to evolve through that discussion: traditional data architectures are optimized around questions we already know, while AI analytics is increasingly valuable for exploring questions we did not anticipate.

That led me to a stronger view: in an AI-native architecture, pre-aggregated serving models should become optional accelerators, not the canonical source of truth.

The role of DWS

Traditional data warehouses often introduce a summary or serving layer between atomic data and end users. In some environments, this is called DWS. Elsewhere, it appears as aggregate fact tables, data marts, Gold-layer models, or materialized views.

These models were designed for a good reason: make known questions faster and easier to answer.

If teams repeatedly need website visits over the last 30 days, we can precompute that metric at the customer or lead level. Dashboards become faster, queries become cheaper, and business users do not need to reconstruct the logic every time.

But that convenience comes with a loss of flexibility.

What happens when someone asks for 31 days instead of 30? Or wants to understand the sequence of visits, compare different windows, or combine those events with a dimension that was not included in the original model?

Once the detail has been aggregated away, AI cannot reason its way back to it.

This matters because the value of AI analytics is not only answering known questions faster. It is allowing people to explore questions we did not anticipate when the data model was designed.

Change the role of aggregation

The answer is not to eliminate aggregation. Performance, cost, and consistency still matter.

Instead, I think the role of aggregation needs to change.

In an AI-native data architecture, pre-aggregated serving models should be treated as optional accelerators, not the canonical source of truth.

The durable foundation should be:

1. Clean, atomic fact tables that preserve business events at the lowest useful grain.
2. Structured semantic metadata defining entities, relationships, joins, metrics, ownership, and freshness.
3. Curated context and skill files explaining analytical conventions, business logic, and known exceptions.
4. Heavy offline and online evaluation covering result correctness, join paths, data quality, latency, and cost.

On top of that foundation, materialised aggregates and caches can accelerate recurring, expensive queries. They remain optional and replaceable, without restricting access to the underlying facts.

What changes for the data team

This shifts the work of the data team.

Instead of trying to anticipate and physically model every future question, we invest more deeply in the atomic foundation, semantic context, and evaluation system that allow AI to answer new questions safely.

Keep the truth atomic.

Govern the meaning.

Make aggregation disposable.

Evaluate continuously.