Analysis | Published 20 November 2026. Product documentation checked on that date.
An AI assistant can return a plausible revenue figure while answering the wrong business question. Bookings, invoiced sales and recognized revenue may all appear in the same warehouse. A query can execute successfully and still select the wrong definition. For leaders deploying conversational analytics, the first acceptance test should therefore be whether the system uses the approved metric, period and scope.
The question is timely as enterprise platforms connect more assistants to governed business data. At Ignite, held on 17–20 November 2026, Microsoft described Fabric IQ as combining OneLake data, Power BI semantic models and operational context from ontologies. Its announcement distinguishes generally available capabilities from previews, including new ChatGPT and Claude connectors. This is a direction of travel, rather than evidence that connecting an assistant automatically makes its answers reliable.
A semantic layer gives the question a business definition
A semantic layer stores reusable definitions between underlying data and the applications asking questions of it. Instead of leaving each interface to reconstruct what revenue means, an organization can specify the calculation once and expose that definition to approved consumers.
dbt’s documentation describes centrally defined metrics powered by MetricFlow, with joins handled by the Semantic Layer. It also documents access through its MCP server for AI tools. That architecture can reduce duplicated calculation logic across applications. It does not decide which accounting or operating definition a business should approve; that remains a management responsibility.
Consider a hypothetical sales assistant asked whether a region met its target. Before evaluating its prose, establish whether the target covers signed contracts or recognized revenue, whether cancellations are excluded, and whether the reporting period follows a calendar or fiscal schedule. These are different questions, even when employees use the same short phrase.
Names and synonyms are part of the interface
Databricks’ metric-view documentation explains how display names, formatting specifications and synonyms provide business context for visualization and natural-language tools. It lists runtime and YAML requirements, making compatibility a concrete implementation check. Descriptive metadata helps connect everyday language with technical fields; it is not a substitute for validating the underlying calculation.
The practical implication is to treat naming changes as product changes. If one department calls a measure net sales and another calls it revenue, someone must resolve whether those are true synonyms. Adding both labels without resolving the difference could make the interface more fluent while leaving the disagreement hidden.
Put durable rules in the governed model
Google’s Looker guidance makes a useful distinction between shared model definitions and agent instructions. It recommends clear field descriptions, moving important calculations into LookML, and keeping field-specific metadata and hard requirements in the model. It also warns that free-form instructions can become stale as models change, and points to verified question-and-query pairs.
For buyers, this suggests a procurement question: where does the durable business rule live, and which channels inherit it? A rule maintained separately in every assistant creates another reconciliation task. Ask vendors to demonstrate how a changed definition reaches dashboards, APIs and conversational tools, rather than accepting a demonstration of one correct answer.
Build a small acceptance set before broad access
Our recommendation is to start with one decision process and a compact set of questions approved by its business owner. Include an ordinary query, an ambiguous term, a period boundary and a case where the available data cannot support an answer. Document the expected measure, filters, freshness and acceptable explanation before running the assistant.
Then compare results with the existing approved reporting process. Record definition mismatches separately from execution failures and unsupported explanations. Repeat the exercise after changes to models or metric definitions. This applies the principle behind workflow-specific AI evaluation to a narrower, inspectable business outcome.
Give the metric a named owner and a review path. An unresolved dispute about what sales means should return to that owner, with the conflicting definitions visible. Faster access to numbers only improves data-driven decisions when the organization can explain what those numbers represent.
Illustration: Original AI-generated editorial artwork created for WiredBusiness, depicting the calibration of business metrics. It does not depict a vendor interface, measured dataset or customer deployment.