Agree on what the number means
A question about complaint rates needs a denominator, a time period, and a definition of a complaint. A syntactically valid query can still answer a different question. Begin with reviewed metric definitions and representative questions. Treat business meaning as maintained application knowledge rather than assuming the model can infer it from column names.
Make the route to evidence visible
Gilles delivered a production agentive BI capability with charts, reports, interaction drill-down, and client validation. That experience illustrates the connection between a business question, a calculation, and supporting records. A general design recommendation is to show enough calculation context for a user to understand the answer, including important filters and qualifications.
Enforce access outside the model
Read-only queries limit changes but still need restrictions on accessible data. Tenant isolation describes production work on scoped querying, validation, and execution limits. PostgreSQL’s row-security documentation explains one public mechanism and its exceptions. It does not establish that an application is secure merely because the feature exists. Agent instructions should never be the authority for assigning access.
Evaluate meaning as well as execution
A useful review compares the intended calculation, selected evidence, and explanation. Include empty results, ambiguous terms, and questions outside the permitted scope. Client validation and regression checks were part of the delivered work. Broader test recommendations belong to the evaluation design; they are not claims of measured revenue gains or universally correct automated analysis.
Sources & further reading
Written by Gilles Jacobs.
