Answers in minutes, not in tickets.
An international consumer goods group wanted to give sales and marketing direct access to its own business figures - in plain language, without going through the BI team. The decisive constraint: the answers had to be reliable enough to go into board decks without further checking.
The starting point
The client’s BI team handled several hundred ad-hoc analyses every quarter for sales, trade marketing and category management. The wait was frequently several days - often enough for a question that could have been answered in minutes. In parallel, Excel exports had established themselves as a shadow stack, with the familiar effect: every department was working from slightly different numbers.
What we deliberately did not do
We did not let the AI query the database directly. A solution in which the language model writes and executes the query itself looks impressive in a demo - and fails in daily operation at the latest when the answers are meant to end up in board decks. Unnoticed misreadings, confused columns, silent omissions: in a group environment, none of that can be repaired after the fact. We put the AI where it is reliable - understanding the question - and executed everything that has to be deterministic in a controlled layer.
What we actually built
- A natural-language input for the business units - no SQL knowledge, no new UI hurdle.
- A deterministic step between language and data: the AI understands the question, and a controlled layer translates it into a validated query against the official data model.
- Results in the form the recipients are used to: a table in the browser, a chart, or an Excel export.
- An auditable record: for every answer, the source, the filters and the calculation are documented and visible to the business units.
- Several AI providers behind the scenes - the platform is not tied to a single contractual partner.
What did not work
The first version let the AI choose the presentation as well. The result: the same question produced a bar chart one time, a table the next, a heatmap after that - and the business units lost confidence, because "the numbers looked different". We fixed the presentation deterministically by data type and aggregation, and limited the AI to what it does reliably: understanding the question.
The strategic lever
The decisive lever was the clean separation: AI for understanding, a controlled layer for execution. The AI is given responsibility only where it is robust - with language. Everything that has to be deterministic (definitions, filters, calculations) runs through clearly validated logic. That cuts misinterpretation off at the source rather than filtering it out afterwards - which is the precondition for sales and board reporting trusting the answer.
What we would do differently today
We would have made the handling of synonyms robust earlier. In the first weeks in production the system regularly returned the wrong metric when terms diverged ("revenue" vs. "net proceeds") - with the corresponding loss of trust. Today we maintain terminology agreements close to the schema; on a next project we would align that layer with the business units from the very beginning.
Self-service analytics without wrong numbers?
45 minutes. We show you where AI can be given responsibility - and where it cannot.