Note ·
A taxonomy agents can read
A warehouse of exported files is a pile until something says what the things in it are and how they relate: business capabilities, applications, data objects, interfaces, owners, controls, risks, and the edges between them. With that in place, “what depends on the policy system” is a query with one answer, and an agent can reason about a change instead of pattern-matching on names. Without it, retrieval returns everything that mentions the word, which is the failure I described in Retrieval Is a Tool, Not a Layer.
The taxonomy is also a map, in the sense the state’s cadastral map was one: it makes a messy organisation legible to something that acts on it, and that is where the caution comes from. The cadastral-map argument is James C. Scott’s, in Seeing Like a State, which is on my reading list, and the Legibility topic on this site is where his readers and these architecture notes meet. The map will be wrong exactly where the company runs on informal knowledge, the arrangement two teams have had for years that no system records, the integration that works because one person watches it. Agents acting on the map will be wrong there too, and confidently. The ontology therefore needs a way for people to annotate where the map and the territory diverge, and the annotations are not a defect of the model; they are the most valuable records in it.
The alternative, redrawing the organisation to fit the map, is the high-modernist mistake, and it has been made before with less capable tools. A small example of an ontology that earns its keep is the topic vocabulary of this site, where every topic has broader parents and the map is drawn from them rather than by hand.
References
- James C. Scott, Seeing Like a State (Yale, 1998).