A question that should take ten seconds took two days

The data was never the problem. Safe access was

Enterprise databases and semantic models already held the answers to most business-critical questions. Getting anything out of them required SQL or DAX, schema knowledge, and an understanding of the business rules governing how the numbers should be read. For most people that barrier is insurmountable, so every question routed through an analyst who knew both the schemas and the access rules.

• The answers lived in two places. Operational data sat raw in the warehouse; business logic sat separately in semantic models. Users did not know which held a given answer, let alone which to trust when both could produce a number.
• The schema is too large to hand to a model. Thirty tables cannot be passed on every request: accuracy degrades and cost rises as irrelevant context crowds out the tables that matter.
• Nobody types the value that is in the database. People ask about "the north region" when the column holds NORTH_AMERICA. A query built on the literal phrase returns nothing, and an empty result reads as "no data" rather than "wrong filter".
• Access control is fragmented by region, partner and role, and write access could not be risked at all. A conversational layer returning unauthorised results is the fastest way to have it switched off permanently.

An empty answer costs more than a slow one

A semantic model that lacks the requested concept returns nothing, and to the person asking, nothing reads as "no data exists" rather than "wrong engine". A single-path system quietly loses trust that way, one unanswered question at a time.

So the platform is built trust-first and coverage-second. Semantic models are consulted first because they hold business-approved logic. The warehouse guarantees that no question goes unanswered. If no model covers the concept, the query returns zero rows, or execution errors, it routes to the warehouse automatically: no retry, no error message, no intervention.

Route it, ground it, then check who is asking

Database knowledge is split across two layers. A lightweight catalogue of plain-English table descriptions identifies which tables are relevant; only then does a rich metadata store (column definitions, synonyms, valid filter values, metric formulas) come into play. Thirty tables never become thirty tables of schema on every query; one to three relevant ones pass through.

Business terms are resolved to canonical dataset values before any query is built: a region abbreviation to a Global Region, a vendor shorthand to a Provider. Ambiguity is settled in semantic context rather than guessed at, and nothing is generated until every filter value has matched a real entry. Resolution runs in parallel for both paths, so neither engine waits on the other.

Access control is enforced before execution rather than filtered afterwards, and every answer discloses how it was produced. The core system is domain-agnostic: what changes between deployments is the metadata catalogue and the business logic configuration, which is why updating that logic costs no code.