Querying company data through dialogue shifts the requirement backwards. It is not the language model that determines the quality of an answer, but whether the organisation has agreed what a term actually means. Whether revenue is counted net or gross, whether intercompany transactions are included, which date applies: without these definitions, a well phrased answer rests on an uncertain basis.
Our projects show the same pattern that market observation describes. Initiatives rarely fail because of technology, but because master data is inconsistent and nobody owns the metrics. That effort is unglamorous, predictable and the precondition for everything else.
We recommend a narrow start. First, the most important metrics are defined in writing, each with a responsible person. Then a single analysis case that regularly causes debate is carried through completely, from source to answer. Only then does the question of tooling arise, and the existing platform often covers more than expected. From the outset it should also be clear who is accountable for a decision based on a model answer.