The data work that decides whether your AI works

Organisations usually have more data than they can use. The obstacle to AI is not quantity; it is that the same metric is defined three different ways, that nobody can say where a figure originated, and that access is governed by convention rather than by a system.

A language model placed on top of that will answer confidently and inconsistently, which is worse than not answering at all.

Natural-language analytics is appealing precisely because it promises to skip the hard part. It does not. If "active customer" means something different in finance and in product, a natural-language interface will simply make the disagreement faster and harder to notice.

A governed semantic layer — one agreed definition per metric, with the logic visible — is what makes conversational analytics trustworthy. Build it first, then put an interface on it.

Retrieval systems frequently bypass entitlements because the index was built without them. The result is a system that can surface content a user has no right to see, which is a serious problem discovered at the worst possible moment.

Permission-aware retrieval means the access decision is made at query time against the same entitlements as the source system. It is more work to build and it is not optional.

When an AI-produced figure is questioned, the useful answer is a chain: this number came from this field, in this table, populated by this pipeline, from this source system, at this time.

Without lineage, every dispute becomes an investigation. With it, most disputes resolve in minutes — which is the difference between a system people rely on and one they quietly stop using.