Not all semantic layers are equal

The real differentiator is how many types of questions the semantic layer can answer natively.

Most conventional semantic layers handle the OLAP basics well: lookups, slice-and-dice, simple aggregations. These are the pivot-table equivalents that resonate with end users familiar with self-serve spreadsheet workflows, and they're what most vendors demo during evaluations.

The hidden gap.

There's a non-trivial class of questions that falls outside what conventional semantic layers can express: running totals by segment, rolling windows by region, nested aggregations, multi-step calculations, retention cohorts.

The unspoken workarounds: build a derived table, push calculations outside the semantic layer, or ask an analyst to write custom SQL. These gaps don't surface during PoC evaluation. They appear later, as end users push the tool further.

In the world of AI, the variety of questions users ask will only grow more sophisticated, and will inevitably hit these same boundaries. When that happens, AI has to generate raw SQL from scratch, losing all the governance and reliability the semantic layer was supposed to provide.

This is why some vendors quietly position AI as "best suited for lookup questions" to avoid exposing where their semantic layer falls short.

Holistics' metrics-centric semantic layer.

Our semantic layer answers a wider variety of self-service questions than what most conventional semantic layers currently support: running totals, rolling windows, nested aggregations, period comparisons, all as first-class metric definitions.

It's easier for both AI and humans to parse, understand, and debug because it operates at a higher level of abstraction than SQL, while remaining SQL-native. Every metric is composable, governed, and visible to AI.