Monitoring tools can generate more alerts than a maintenance organization can responsibly evaluate. Planning tools can generate detailed work that is disconnected from why the work was requested. A governed loop joins those activities without pretending that detection, diagnosis, and approval are the same decision.
Define the states before automating the transitions
A clear state model prevents a probable explanation from silently becoming an approved job. The names will vary by organization, but the distinctions should remain visible.
- Detected: a rule, model, inspection, or person identifies a condition worth evaluating.
- Contextualized: the signal is joined to the correct asset, operating state, history, and strategy.
- Assessed: evidence and alternatives are reviewed, with uncertainty recorded.
- Planned: a proposed response includes scope, prerequisites, resources, and supporting rationale.
- Approved: the designated authority accepts, modifies, defers, or rejects the proposal.
- Handed off: approved information is transferred to the agreed system or operating process.
- Learned: execution and inspection feedback is associated with the original condition and decision.
Preserve three different kinds of truth
A defensible workflow distinguishes observation, interpretation, and action. “Bearing temperature exceeded the configured threshold” is an observation. “Lubrication degradation is the leading hypothesis” is an interpretation. “Inspect lubrication condition within the approved response window” is a proposed action. Combining them into one sentence makes uncertainty hard to see.
Condition observed · asset and operating context · relevant trend/history · candidate explanations · contradicting evidence · uncertainty · proposed response · approving authority
Control identity at every hand-off
The asset identifier, source event, revision, and decision ID should remain linked as information moves between workflows or systems. If a planner must manually search for the original trend, or if a completed work order cannot be traced back to the alert that initiated it, learning becomes anecdotal and duplicate action becomes more likely.
Use confidence carefully
A numeric score can look more precise than the evidence permits. If confidence is displayed, define what it represents, how it was produced, and what it is allowed to influence. Pair it with the evidence and alternatives. Never allow a score alone to bypass required engineering, safety, or operational review.
Plan around constraints, not just the suspected cause
A response may depend on isolation requirements, production windows, permits, competence, spares, special tooling, access, inspection method, and post-maintenance testing. The planning step should surface those prerequisites and identify which details still require site confirmation. Generated instructions should not override controlled procedures or OEM and site requirements.
Make approval an explicit decision
Approval should capture who decided, under which authority, what evidence was considered, what changed, and whether conditions were attached. Deferral and rejection are valid outcomes and need reasons. For higher-consequence decisions, the workflow may require more than one discipline or a separate management-of-change process.
| Transition | Minimum control question |
|---|---|
| Detected → contextualized | Is this the correct asset and operating state? |
| Contextualized → assessed | Is the evidence sufficient to evaluate, and are material gaps visible? |
| Assessed → planned | Does the proposed response follow the approved strategy and site decision process? |
| Planned → approved | Are scope, risk, prerequisites, and decision authority clear? |
| Handed off → learned | Did execution findings return to the same asset, condition, and decision record? |
Measure the loop, not just the alert
Alert volume and acknowledgement time are incomplete measures. Track where decisions stall, how often asset context is corrected, which proposals are changed or rejected, whether evidence is accessible at review, and how often execution findings confirm or challenge the original assessment. Interpret outcome measures carefully: avoided failures are difficult to prove, and many operational changes may influence the result.
Start with a bounded pathway
Choose one asset class, a limited set of condition types, named reviewers, and a controlled hand-off. Walk real historical cases through the state model before connecting live operations. Record where terminology, identity, evidence, or authority breaks down. Then design automation around the corrected process.
AndonEAM describes its product areas as a connected workflow, while actual features and integration scope are confirmed during implementation. Begin with the documentation overview or discuss a bounded use case.