How should a plant set approval thresholds for AI maintenance recommendations?
An AI system can summarize evidence and propose a maintenance action without being authorized to release that action. Plants need a written decision boundary that distinguishes information retrieval, draft recommendations and operational changes. The threshold should depend on consequence and uncertainty, not on how confident the AI sounds.
Separate suggestions from actions
Define permission levels before the pilot. Read-only answers may help staff locate documents. Draft work orders can be reviewed by a planner. Changes to priority, schedule, protective settings, operating limits or equipment configuration need the site's established approval authorities. An AI-generated statement of confidence is not a substitute for engineering evidence or an approved procedure.
Use a simple decision matrix
- Low consequence, clear evidence: AI may propose a draft and identify supporting records; an authorized person confirms it.
- Medium consequence or incomplete context: require planner or supervisor review and a check for existing work, material and operating constraints.
- High consequence, safety or environmental exposure: require qualified operations and engineering review under the applicable site process; no autonomous release or control action.
- Unknown asset, conflicting documents or inaccessible source: stop the automated workflow and request clarification or verified evidence.
Make evidence visible at the decision point
Show asset identity, the alert or observation, relevant trend window, source documents with revision dates, work history, open work orders and the proposed action. Identify what the system did not find. The reviewer should be able to reject, edit or escalate the recommendation and record why. A source link that opens an unrelated or unauthorized document is not adequate evidence.
Example: two alerts on one critical pump
The system sees an abnormal vibration trend and drafts a bearing inspection work order. The CMMS already contains a planned inspection for the same pump. The workflow should flag the potential duplicate and show both records, not silently create another job. If the suggested action requires a trip setting change, it must be routed to the established engineering and operations process rather than handled as ordinary planning.
Test the gate before a live pilot
Use representative historical cases, including false alerts, duplicate work, missing documents, wrong asset tags and high consequence scenarios. Measure the share of suggestions correctly escalated, unsupported claims, reviewer time and unauthorized actions attempted or prevented. Document a stop mechanism and an incident path. NIST's AI Risk Management Framework provides voluntary guidance for governing, mapping, measuring and managing these risks.
Maintain a traceable record
Store the recommendation, evidence references, model or workflow version, human decision, timestamp and final work order outcome. Review rejected and accepted recommendations to improve alert quality. Approval should remain meaningful; a one click rubber stamp defeats the control.
Explore KANE consulting
Discuss practical maintenance and reliability improvements for your facility.
Request a consultation →Related answers: verify ai maintenance answers · measure ai pilot success