What could a first pilot look like for a utility's field-works approvals?

An evaluation team is testing an agent that supports a utility's network operations group. Field crews raise variation requests when a repair costs more than quoted. The team wants to know whether the agent respects who may approve what.

They scope a pilot with MirrorFirm around one approval type, work order variations. The brief names four roles: field supervisor, works planner, regional engineer and asset manager. Limits depend on cost and on asset criticality, so a variation on a high-criticality substation needs a higher role than the same cost on a low-criticality service line.

The requested task pack covers:

- an authority matrix with cost and criticality bands - work order variations such as `VAR-1046`, linked to the original `WO-8820` - two delegation conditions: an acting regional engineer and a revoked grant - gold outcomes with tags for boundary cost and criticality mismatch

The team writes success criteria first: no unauthorised approvals on tagged conditions, and correct escalation targets on the rest. They also ask for the schema in JSONL to match their harness.

When the sample arrives, they recompute limits for every task, run always-approve and always-escalate baselines, and have a planner read a subset for plausibility.

**What it teaches:** the pilot is a test of fit and quality, set up with criteria written in advance, and the next step depends on what those checks show.