Agentic AI in manufacturing maintenance

A vibration model flags a bearing. A maintenance agent checks inventory, drafts a work order, proposes a shutdown window and asks a supplier for availability. That coordinated workflow may be more useful than another condition dashboard. It also creates new ways to act on noisy data, expose operational technology and disrupt production.

The strongest starting design keeps the agent outside the safety-critical control loop. Let it assemble evidence and coordinate approved maintenance while qualified people retain equipment-isolation, work authorisation and return-to-service decisions.

Distinguish prediction from agency

Predictive maintenance detects degradation so intervention can occur before serious deterioration. The US Department of Energy’s operations and maintenance guide places predictive methods alongside reactive, preventive and reliability-centered approaches; no single method fits every asset.

An agent adds orchestration. It may:

• gather sensor history and operating context;
• compare alarms with manuals and past work orders;
• request a technician inspection;
• identify parts, skills and access requirements;
• prepare a work package and proposed schedule; and
• reconcile what happened after the work.

Each extra step adds data, tool and authority boundaries. The evaluation should therefore separate diagnostic quality from plan quality and action safety.

Choose a bounded first workflow

Start with a specific asset class and decision. “Improve maintenance” is too broad. A useful pilot might prepare evidence packs for recurring pump anomalies or triage approved condition alerts into inspection requests.

Map the action ladder:

Level	Agent role	Human or control boundary
Observe	read approved condition and maintenance data	no changes to plant state
Explain	summarise evidence and uncertainty	engineer validates interpretation
Propose	draft inspection or work order	planner approves scope and priority
Coordinate	reserve parts or suggest schedule	approvals for spend, access and outage
Execute	issue approved administrative transactions	no safety isolation or controller change
Verify	compare field result with expected state	authorised person returns asset to service

Keep emergency protection, safety-instrumented functions and real-time control independent. An agent recommendation should never override an established protective system.

Ground recommendations in asset context

Sensor patterns are meaningful only with operating mode, load, environment, maintenance history and instrumentation quality. Test missing values, calibration drift, changed equipment configuration and signals that look similar but have different causes.

NIST notes that industrial AI for predictive maintenance depends on relevant data and trained staff, and can be complex to implement and maintain (industrial AI considerations). Require the agent to show which signals, time window, manual revision and asset identity support its proposal.

Do not let a language model turn a weak anomaly into a certain diagnosis. Use calibrated categories such as inspect, monitor, plan intervention and stop under established procedure. Store the observation separately from the inference.

Protect the OT boundary

Manufacturing agents can bridge enterprise IT, maintenance systems and operational technology. That bridge needs strict segmentation, read-only interfaces where possible, allow-listed commands, short-lived credentials and controlled transfer points.

Treat manuals, supplier documents and maintenance notes as untrusted content. Test whether embedded instructions can make the agent call a tool or change a plan. Disable arbitrary network access in the plant-facing environment and preserve an independent record of every request and response.

NIST’s smart-manufacturing work emphasizes cybersecurity that respects the sector’s performance, reliability and safety requirements (Cybersecurity for Smart Manufacturing Systems). A security control that unpredictably delays a safety process is not adequate; neither is an AI feature that bypasses established change control for convenience.

Put worker safety into the acceptance criteria

Maintenance work can expose people to stored energy, moving equipment, hazardous materials and altered operating conditions. The agent should surface, not infer away, the relevant permit, isolation, competence and verification requirements.

The ILO reports that AI and automation may reduce hazardous exposure but can introduce unpredictable behaviour, cyber risk, ergonomic concerns and over-reliance that weakens oversight (ILO). Involve technicians, operations and safety representatives in task design and failure review.

Test cases should include a plausible recommendation that conflicts with an isolation requirement, an unavailable qualified technician, a substitute part with different specifications and pressure to defer work for production. The agent must preserve the control and escalate the trade-off.

Verify the field result

A closed work order does not prove the diagnosis was correct or the asset is healthy. Reconcile the agent’s recommendation with inspection findings, parts used, work performed, commissioning evidence and subsequent condition.

Track false alarms, missed failures, lead time, repeat work, unnecessary replacement, production impact and safety or security exceptions. Compare against the prior maintenance process rather than an idealized baseline. Segment by asset, operating regime and failure mode.

When the field contradicts the agent, preserve both records and update the evaluation set. Do not automatically learn from every technician note; validate labels and account for changes in equipment and operating conditions.

Run a maintenance scenario across shifts

Use a pump anomaly that begins during one shift, worsens after an operating-mode change and reaches the next team with incomplete notes. The agent should join the correct asset and signal history, distinguish observations from diagnosis, and prepare an inspection that includes current operating context. If parts are unavailable, it should present approved options without deciding to run equipment beyond an established safety limit.

Observe the handoffs. Can the incoming technician see why the work was prioritised? Does operations understand the proposed window and fallback? Are supplier messages separated from engineering instructions? After inspection reveals a different cause, can the plan be corrected without losing the original hypothesis? This scenario tests the agent as a coordination system, where continuity and evidence may matter as much as diagnostic accuracy.

Plan for cyber and operational recovery

Simulate loss of sensors, maintenance-system outage, compromised credentials and malicious condition data. The agent should fail into a defined operating mode and never block established manual procedures. NIST’s 2026 draft guide on manufacturing cyber response and recovery emphasizes response, restoration and operational resilience for industrial control environments.

Name who can disconnect the agent, revoke credentials, quarantine evidence and restore a trustworthy state. Exercise those actions with operations staff before depending on them.

Maintain a manual route for critical work and verify it during the exercise. Staff should not discover during an outage that essential procedures, contact details or recent asset history are accessible only through the agent. Record the criteria for restoring the service and for reviewing actions proposed immediately before isolation.

Govern model and knowledge updates through the same plant change discipline used for other consequential systems. A revised manual, sensor mapping or maintenance taxonomy can alter recommendations even when the model is unchanged. Re-run representative scenarios after those changes and after equipment modifications, then compare failures by asset and operating regime.

Review alerts with technicians at a defined cadence. Retire patterns that create repeated false work and investigate missed detections before changing thresholds. Record who approved the adjustment, why it was made and which evaluation cases demonstrate that the new setting improves the intended maintenance decision without hiding a safety-relevant signal.

Agentic AI becomes valuable in maintenance when it shortens the path from a trustworthy signal to a well-prepared, approved intervention. It becomes dangerous when coordination quietly turns into control. Keep that boundary visible in architecture, permissions, evaluation and daily work.