A strong answer. “I would not build the architecture first. I would find the two or three things people wait for, because that is what determines whether anyone uses it.

Usually it is: getting a new source in, getting access to data they can see is there, and knowing whether a number is trustworthy. So the first thing I build is ingestion as configuration — adding a source is a file, not a bespoke pipeline — because that removes the largest queue and buys the credibility to do everything else.

Then access derived from classification rather than from tickets, so the answer to ‘can I see this’ is automatic and the security team gets a model instead of a backlog.

Then the layered model — raw append-only, staging, core, marts — because without it we get six definitions of customer and the platform becomes a place numbers are argued about.

Then cost attribution from day one. Not optimisation — attribution. Tag warehouses and queries by team, because without it nobody owns anything and every cost conversation is about the total.

What I would deliberately defer: the catalogue, which is worth little until there is something worth cataloguing, and the semantic layer, which people build early and then find nobody adopts. And I would measure adoption — percentage of sources on the framework, time from request to data available — because a platform nobody uses is indistinguishable from one that does not exist.”