Every grid operations manager can name the decision that hurts. It has a clock on it, it costs money in both directions, and the answer lives in six systems that do not agree. Here is how we model that, and why the model is what makes an AI agent useful.
A distribution transformer runs hot on the second evening of a heatwave, and someone decides tonight whether to load-balance it, defer it, or roll a crew. The answer is split across the asset system, the GIS, SCADA, the outage system, the customer register and the ERP. Six identifiers for one transformer, no registry declaring them the same unit, the reconciliation happening in the head of someone who happens to know.
What the ontology buys
Point an agent at those schemas and it is guessing. Not for want of a better model: the meaning of a utility does not live in its tables. It lives in historian tags named after the instrument a commissioning engineer installed, and in the head of the switching engineer who knows which feeders the asset register is wrong about. Frontier models exceed ninety percent accuracy turning a question into SQL on clean academic schemas, and fall to roughly one in ten on real enterprise warehouses.[1] The schema is the binding constraint, not the model.
Declare the ontology and the same agent stops guessing. It knows that one asset record, one map feature and one telemetry point are a single transformer, and that what that transformer delivered on the fourteenth of March has a settled answer and a metered one. From there it reads across the six systems in place, in the vocabulary the control room already uses. That is the difference between AI that drafts text and AI that answers a decision: the answer arrives with the evidence under it, and with the act that follows from it.
What we declare first
One decision with a clock and a named owner, never a source system inventory. Then identity, before any attribute: for each object, which system is authoritative for its existence, its key in every other system, and the rule that resolves them. Where that link exists nowhere, and meter to transformer assignment is the classic case, either infer it and carry the inference with its confidence, or commission the field survey.
Then time, because a grid runs three clocks: the as-built network in the GIS, the as-operated one in the outage system including the jumper installed at three in the morning, and settlement, mutable for months. A question about last Tuesday’s loading must resolve against last Tuesday’s network, or the answer comes back confident, against today’s topology, and right often enough that nobody catches it for a year.
Every action is staged
Competence is half of it. The other half is that the agent never executes.
Nothing here produces a switching order. It produces the packet the operator signs. The control system boundary is never crossed in the write direction. Writes land in the enterprise systems of record, the asset registry, the work order, the inspection: proposed with a real diff of what would change, then committed or rolled back by a person. Writability is default deny, and a source held as a managed copy is refused outright, because the next sync would erase the write.
Permissions live in the database, not in the prompt. A settlement analyst and a wildfire planner query the same asset table and see different columns because the database says so, and an agent composing a query it may not run takes a privilege error, not a polite refusal. The trail is append-only and hash-chained, so a filing eight months later re-derives the same computation and names the delta.
Where this is wrong
Real-time control is not ours: anything closing in milliseconds belongs to the energy management system, which a decision system reads and never joins. And no ontology conjures a meter-to-transformer mapping nobody recorded.
Utilities do not lack data. They lack a declared model of their own operations that a machine can reason over, and a record that stands up two years later.