The consumer
The metamodel gives an agent a grammar it cannot step outside of. The instance graph gives it facts it did not write. With both, answering a question about your organisation stops being an act of composition and becomes an act of traversal — and the difference shows up in the one place that matters: whether you have to check the answer before you forward it.
Retrieval and traversal
Both start the same way: find the thing being asked about.
retrieval
stops there
traversal
starts there
Ask what a supplier failure reaches, in the Adventure Works model:
typed traversal · three hops · four types, three departments, one query
Supplier
Victory Bikes
Product SKU
Touring Tire Tube
Sales Order
SO70282
Retail Shopper
Dalton Adams
Three hops, and the agent inferred none of them. It did not work out that a tyre tube belongs in a bicycle or that an order implies a customer — each hop is a typed relationship somebody or some system already asserted. The chain crosses a supplier, a product SKU, a sales order and a person: four types, three departments, one query.
And one hop further out is the part a paraphrase never reaches. The same component either has a second agreement or it does not — this one does not: one agreement, and no other — so the answer is a single point of failure, named. For the tyres the same supplier also provides, the traversal returns a second agreement instead, and the answer becomes an impact with a mitigation attached. That difference is not judgement. It is the second edge on the same node, present or absent.
Why the answer holds
Three jobs, and only the first and the last are language:
01 · language
Translate the question into a traversal
02 · deterministic
Execute it deterministically
03 · language
Narrate the result
Organisational intelligence is not a language problem with a graph behind it. It is a graph problem with a language interface — and asking a model to reason about impact from prose is asking it to do the one thing it cannot be held to.
Three things follow, and none of them is a prompt.
what you can check
what gets recorded
what improves it
At the edge of the model
Asked who else could cover the purchase orders assigned to one buyer in Purchasing, the agent in the demo tenant did not answer that nobody could. It answered that there were
no secondary co-assignees modeled in the graph.
Read the qualifier. It is not a hedge — it is the whole claim. There may well be a colleague who covers her every August; what the model states is that nobody has recorded it. One of those sentences you can act on, and the other you would have had to verify before you did.
An agent over documents cannot make that distinction, because there is nothing for it to be silent against. Absence in a document store is indistinguishable from a page nobody wrote. Absence against a type is a finding: the relationship exists in the grammar, this node does not have it, and that gap has an owner — because the responsibility matrix belongs to the type, so every node knows who answers for what is missing on it.
Which is also the honest answer to the fair question about any model: what if ours is incomplete? It will be. The agent will tell you where.
Unprompted
A typed graph with state is something that can be monitored, and monitoring is where an agent earns a standing seat rather than a chat window. It notices:
that a process lost its owner
that an approver holds a mandate they have never used
that hundreds of orders now route through one person with no co-assignee recorded
that a skill has exactly one holder on the evening shift
None of those is a question anyone would have thought to ask on the day it became true. All of them are states the model can be checked against — and each one routes to a named person, because the responsibility matrix is a property of the type and not a notification rule somebody configured.
Autonomy of attention, not of authority. The agent identifies, reports, monitors and escalates. It does not change your organisation, and the boundary is not a policy in a prompt: what it is able to assert at all is limited by a grammar it did not author and cannot extend.
One more objection
So far this describes a model that is true today. You have seen this part fail before — the catalog nobody updated, the CMDB that was wrong within a year. That objection deserves its own answer.