Dex puts analytics-agent governance below the prompt—but its abstention claim still needs a benchmark

The Modern Data Company has launched Dex, a conversational analytics interface for DataOS that is available now to DataOS customers. The product accepts business questions and follow-ups in natural language, returns answers or charts, and says it will decline when governed data cannot support a reliable answer. Its more important design choice is where trust controls live: identity and access policies are evaluated by DataOS as queries execute, rather than delegated to instructions inside the language model. The Modern Data Company launch

Provenance is part of the answer

Dex retrieves a relevant governed data product and its definitions, relationships, lineage and quality signals before DataOS runs the query where the data already lives. The response can expose the metric and definition used, the selected data product, applied filters and underlying sources. DataOS can operate across Snowflake, Databricks, BigQuery, Iceberg, operational databases and APIs, while customers can select supported commercial, enterprise or private language models. Dex product page

That separation matters operationally. The model handles interpretation and conversation, while the data layer handles authorization and execution. A prompt can ask a model not to reveal restricted information; a runtime policy can prevent the restricted rows from being returned in the first place. The same architecture also centralizes metric definitions instead of rebuilding them in every prompt. Dex product page

What teams should verify before rollout

The launch describes the right controls, but it does not publish an evaluation set, abstention precision or recall, SQL execution accuracy, latency distribution, or evidence that equivalent questions consistently select the same governed definition. Those omissions do not negate the architecture; they define the acceptance test. The Modern Data Company launch

Before production access, teams should build a domain-specific test set with answerable, ambiguous, unauthorized and deliberately unsupported questions. For each response, reviewers should verify the chosen metric, filters, source product and access outcome—not just whether the prose sounds plausible. Follow-up tests should also change user identity mid-scenario and confirm that conversational context never preserves data the second identity cannot access.

Dex therefore makes a useful architectural bet: governed context and authorization belong in the data plane, while the model remains replaceable. The remaining burden is empirical. “Admits when it doesn’t know” is only a control once customers can measure when it abstains, when it should have abstained and what evidence supported the answer it did return.