Context
A question such as “why are customers contacting us?” needs more than a fluent summary. Service teams need to compare patterns, inspect the underlying interactions, and understand what a reported number represents. Gilles built a production agentive BI capability for customer experience analytics, with client validation. The engineering task connected natural language questions to structured metrics and the evidence behind them, supporting exploration from grouped patterns to individual contacts.
Approach
The delivered work combined SQL-based analysis, tool orchestration, charts, and reports. A useful general pattern is to separate question interpretation, data access, calculation, and explanation so errors can be investigated at the right step. The companion article on shared metric definitions for BI agents explains why a correct query still needs an agreed business meaning. For example, a rate needs both a counted event and a denominator, and its explanation should make that distinction understandable.
Quality controls
Implemented controls included read-only queries, tool and query validation, execution limits, and tenant-scoped access. Client annotation and internal validation sets supported checks of the NLP-derived metrics used by the product. Model changes were tested for regressions. Tenant isolation across AI workflows and agent permission boundaries explain the wider design considerations; neither a prompt nor a plausible answer substitutes for enforced access controls.
Outcomes
The result was a production analysis capability that clients could validate against their own questions and interaction evidence. Its value here is the verified combination of analytical functionality and bounded execution. No financial return or client-specific performance gain is claimed.
- Delivered conversational analysis with charts, reports, and interaction drill-down.
- Combined client validation with regression checks and restricted data access.
