Essential Energy Deploys AI for Safety Inspections

Essential Energy is rolling out an AI system that reads field staff notes to surface safety issues across its distribution network serving 900,000 regional and remote premises. The company, led in this work by principal data scientist Andrew Slack-Smith, found the most valuable signals lived in the free-text comments on iPads. The AI algorithm is expected to identify safety-related data in 76% of cases versus 59% for the incumbent rules-based checks, and the program is designed to scale to "hundreds of other data sources."

The team moved beyond structured telemetry and built NLP-driven pipelines to extract meaning from unstructured field notes. Practitioners should expect a stack that includes:

• •text ingestion and normalization from mobile devices (iPads) with timestamp and asset linkage
• •supervised classification and named-entity extraction to label hazard types and asset identifiers
• •lightweight ensemble models or transformer-based classifiers for short, noisy notes

The company plans to add voice-to-text capture so crews can speak observations; the system will enrich those transcripts using its asset knowledge and produce three concise paragraphs to fill reporting gaps, reducing manual entry time. Operationalization considerations include annotation guidelines, label-imputation for sparse classes, latency budgets for near-real-time alerts, and versioned model deployment with monitoring for drift.

Utilities historically rely on structured sensor data and rule engines; Essential Energy's pivot to unstructured field intelligence reflects a broader trend where tacit human observations provide early-warning signals that sensors miss. Extracting actionable insight from workers' notes can shorten mean-time-to-detect and prioritize maintenance spend. The measured uplift from 59% to 76% is commercially meaningful for risk reduction in distributed infrastructure.

Expect familiar ML operational risks: false positives/negatives, hallucination if generative summarization is used, privacy and consent for voice transcripts, explainability requirements for safety-critical decisions, and the need for continuous retraining as field language evolves. Human-in-the-loop review and auditable decision logs will be necessary for regulatory and safety governance.

Track pilot results on real incident reduction, the choice of model family for noisy short-text classification, rollout of voice-to-text, and how the team links textual findings to maintenance workflows and asset telemetry.