Utility decision scenarios expose citation and contextual-reasoning weaknesses

Engage the utility director, capital-program manager, engineering lead and municipal risk team around the quality of decision records. Ask where staff already use AI, which references are checked, and who can reject a recommendation when evidence is incomplete. A bounded engagement could audit a small set of historical decisions using sanitized local cases. The value hypothesis is a more defensible review process, not a promise of safer infrastructure or cheaper capital projects. Do not use this study to rank today's vendors. Smaller utilities may need a shared specialist reviewer, but funding, turnaround and access must be established before recommending that arrangement.

Develop a local rubric before selecting a model. Include operating constraints, asset condition, alternatives and evidence validity, and retain the source record for each scored output. Run repeated trials across prompts and model versions, using blinded domain reviewers and an unassisted baseline. Test whether retrieval changes the failure profile rather than assuming grounding solves it. The prototype must not write to control systems or approve work orders. Validate authorization, confidential-data handling and audit logs. Report disagreement and error severity as well as average scores. Require reviewers to detect seeded citation errors and contextual omissions; a fluent rationale or matching recommendation alone should not pass.

The utility's responsible engineering manager should own the decision boundary and acceptance standard. Delivery work includes selecting cases, recruiting qualified reviewers, recording evidence, training staff and integrating review into capital or maintenance workflows. Dependencies are trusted asset data, professional judgment, security approval and funded review time. Proposed acceptance criteria include verified support for every consequential claim in the sampled decisions, no unauthorized operational action, and successful detection and correction of seeded errors. These are proposed controls, not observed improvements. Maintain a conventional decision path and record overrides. Reassess after model or data changes; monitor reviewer workload so formal oversight does not become a rubber stamp.