The challenge
The group's executive committee received a weekly pack of 40 dashboards and a 60-page deck every Monday, stitched together from four production sites and two service business units. No single human read all of it. The controlling team spent 220 person-days a quarter producing ad-hoc answers to executive follow-ups. Decisions were being made on gut feel while the numbers were waiting in Excel.
The solution
We built a natural-language reporting agent on top of the group's existing data warehouse. First we encoded the company's metrics and definitions into a strict semantic layer — revenue by site, gross margin by product line, capacity utilisation, working capital, customer segments — with full row-level permissions. Then the agent. Executives ask questions in the board room ("Show me gross margin by site this quarter against last quarter, on a constant-currency basis") and receive a formatted answer with the SQL that ran, the rows that were permitted, and a one-click drill-down into the source.
The problem with 40 dashboards
Dashboards freeze the questions you thought you wanted yesterday. The questions you actually want in the board room are different — sharper, more contingent. The group’s executives were asking them, and the controlling team was answering them 45 minutes later on a good day.
We replaced that loop.
The semantic layer was the hard part
The agent is the easy part. The hard part was codifying the company’s metrics and definitions into a semantic layer strict enough that the agent cannot invent numbers. We spent seven of the fourteen project weeks with the controlling team, nailing down every definition, every exclusion, every site-level adjustment. The agent was then built on top of a foundation the controlling team trusted.
Governance
Every question the agent answers is logged with the generated SQL, the permissions that applied, the tables hit and the confidence score. The governance dashboard shows the executive committee exactly what the agent has been asked and how it has answered. Analysts can flag an answer as wrong in one click; the flag feeds a weekly tuning cycle.
What the CFO said
“I used to prepare for the board by reading a deck the team had built for me. Now I prepare by asking the agent the questions the board is actually going to ask. It has completely changed my confidence going in.”
Services deployed
- AI-Driven Reporting (semantic layer + reporting agent)
- Knowledge Engineering (data ontology and document grounding)
- AI Governance (controls, dashboards, audit)
- Custom AI Agents (reporting agent runtime)