Note ·
The bottleneck is read access
Every part of the loop assumes that the systems of record can be read, and in a large company that assumption is where the project goes to die. Data has owners. Ownership feels like standing, and a request for read-only access to feed a warehouse reads, to the owner, as a request to be made replaceable. Then the practical limits behind the political one: APIs that exist in the licence but not in practice, APIs that work but nobody documented, MCP servers that are demos, exports that need a ticket and a person who is on holiday.
I think the honest thing is to say what this is. The capability that AI has democratised, the ability to build and change things, runs here into the oldest form of power there is, holding the data. The warehouse is a political project before it is a technical one, and pretending otherwise wastes a year. The same dynamic at the scale of a continent is the subject of The Bottleneck Is Data, and Europe Is Sitting on It; inside a company it is the same people with smaller titles.
The practical thing is to measure it. The time from “we would like to read X” to the first record from X in the warehouse, per source, is the number that tells you how far a company is from any of this, and it is a number a leadership team can be shown. Start read-only, start with the sources whose owners want to be read, and treat each refusal as a line on the roadmap rather than a reason to stop. Tim Berners-Lee’s five-star ladder for open data works as a maturity scale for internal sources too: a file on a share is one star, linked data with stable identifiers is five, and most companies sit at two.
References
- Five-star open data: the ladder from a file on the web to linked data.
- Model Context Protocol: the standard the MCP servers in question implement.