The Dispatch Brain: AI in Field Service and Maintenance

Field service is not one optimization problem. It is a chain of uncertain decisions: whether a signal indicates a real fault, how soon the asset can be serviced, which skill and authorization are required, whether the part is available, how the site can be accessed, and what evidence proves the asset is safe to return.

AI can help detect degradation, prioritize work, assemble diagnostic context, schedule resources, and turn repair notes into reusable knowledge. It should not turn weak sensor correlation into a safety diagnosis, dispatch an unqualified technician, invent a procedure, or close a work order without verified tests.

This guide reflects public sources available on 30 July 2026. The equipment’s approved maintenance program, manufacturer instructions, applicable asset, machinery, functional-safety, environmental, worker-safety, and sector requirements remain controlling.

Define the maintenance decision and consequence

Start with the asset hierarchy and failure mode. A project should specify:

• asset, subsystem, component, function, duty cycle, environment, and criticality;
• failure mode, symptom, degradation mechanism, consequence, and detectability;
• data source, sample rate, measurement uncertainty, missing-data behavior, and provenance;
• decision horizon: immediate shutdown, inspection, planned repair, next-route visit, or monitor;
• authorized roles for diagnosis, planning, isolation, repair, test, and return to service;
• service window, site access, skills, certification, tooling, part, and permit constraints;
• false-alarm, missed-failure, early-replacement, downtime, safety, and customer costs;
• fallback when telemetry, model, inventory, route, or technician connectivity fails.

Do not call every anomaly “predictive maintenance.” An anomaly may come from operating regime, sensor drift, communication loss, configuration change, or a new but healthy pattern. The system should produce a hypothesis with evidence and required verification.

ISO 55001:2024 specifies requirements for an asset-management system and strengthens decision-making, value, risk, data, knowledge, lifecycle operations, and predictive action. It does not prescribe one algorithm or say every asset should use condition-based maintenance.

Build a condition-monitoring architecture, not a sensor dump

ISO 13374-1:2003 establishes general guidelines for software handling the processing, communication, and presentation of machine condition-monitoring and diagnostic information. ISO confirmed this edition in 2025, so it remains current despite its original publication year.

A practical architecture separates:

1. Acquisition: sensor identity, engineering unit, calibration, sample time, operating state, quality flag, and clock.
2. Manipulation: filtering, resampling, feature derivation, alignment, and normalization.
3. State detection: deviation from an expected condition under the current operating regime.
4. Health assessment: severity, likely fault family, uncertainty, and supporting evidence.
5. Prognostic assessment: a distribution or bounded estimate of future condition, not a guaranteed failure date.
6. Advisory generation: recommended inspection or work under approved rules.
7. Presentation and integration: human review, CMMS/EAM work order, historian, inventory, and audit trail.

Retain the raw or legally required source, transformation version, asset configuration, model, threshold, and decision. A dashboard trend without calibration or operating-state context can create a precise-looking false alarm.

ISO 17359:2018 gives current general procedures for establishing a machine condition-monitoring program; ISO last confirmed it in 2023. Use it as program guidance, not as proof that a particular sensor or model is adequate.

Create reliability data that can travel between teams

Maintenance history often fails because free text mixes symptom, cause, action, and outcome. Define a controlled structure:

• equipment taxonomy and configuration at event time;
• operating state and environment;
• observed symptom and detection method;
• functional failure and failure mode;
• suspected and confirmed cause;
• maintenance action, part, tool, labor, and downtime;
• verification test and return-to-service authority;
• consequence, follow-up, and warranty or supplier link.

ISO 14224:2016 provides detailed reliability and maintenance data structures and quality principles for petroleum, petrochemical, and natural-gas equipment. It remains current after confirmation in 2022, but its sector scope matters. Other industries can learn from its separation of equipment, failure, and maintenance data without claiming sector conformance.

Preserve both source note and normalized fields. AI can propose failure codes and summarize notes, but the technician or reliability engineer confirms material facts. Do not train directly on every historical closure: old work orders may contain copied notes, miscodes, and unsuccessful repairs.

Detect condition changes by operating regime

Compare like with like. Vibration, temperature, pressure, current, image, fluid, and acoustic signals depend on load, speed, product, ambient condition, control mode, and recent maintenance. A global threshold may page on every startup and miss degradation at steady load.

Use operating regimes or physics-aware residuals. Establish baselines by asset family but retain asset-specific calibration. Evaluate:

• event-level precision and recall;
• false alarms per asset-time and nuisance-alert duration;
• detection lead time relative to a usable intervention window;
• missed-failure severity and coverage of known failure modes;
• calibration of risk or remaining-life intervals;
• performance by asset age, model, sensor, site, regime, season, and maintenance state;
• sensor-health and missing-data detection;
• outcome after intervention: confirmed fault, no fault found, recurrence, and avoided consequence.

Remaining useful life is uncertain. Present a range and assumptions, not “failure in 12 days.” Update when load or condition changes. If there is no validated label for actual failure onset, be explicit about the proxy.

Turn an alert into a reviewable work decision

An alert should answer:

• what changed, on which asset and operating state;
• which sensors and periods support it;
• which alternative explanations were checked;
• likely failure family and uncertainty;
• consequence if correct and if ignored;
• approved diagnostic checks;
• recommended urgency and constraint;
• similar resolved cases and authoritative procedure;
• owner, review deadline, and escalation route.

Keep the recommendation distinct from the approved work order. A reliability engineer may combine the signal with production plan, redundancy, warranty, safety, and regulatory constraints. Record why the recommendation was accepted, changed, deferred, or rejected.

Avoid alert floods. Deduplicate correlated signals, group by probable asset-level event, suppress only under controlled conditions, and escalate persistent unreviewed critical alerts. Monitor whether quieting logic hides incidents.

Schedule technicians, parts, and routes under hard constraints

Dispatch optimization should treat safety and qualification as constraints, not soft penalties. Model:

• required skill, certification, authorization, and site induction;
• task duration distribution and multi-person work;
• access window, permit, isolation, escort, and customer appointment;
• compatible part, serial or lot trace, tool, vehicle, and lifting requirement;
• travel time uncertainty, labor rules, fatigue, weather, and geographic risk;
• priority, contractual response, production impact, and asset redundancy;
• precedence between inspection, isolation, repair, test, and handover.

Show the dispatcher why a schedule is recommended and which constraints are tight. Re-optimize when a part, customer, traffic, emergency, or preceding task changes, but preserve the previous plan and notifications.

Do not let the route optimizer infer that a technician can perform a task because they completed a vaguely similar ticket. Skills and authorizations come from controlled records. Provide an offline plan and safe escalation when mobile connectivity fails.

For deeper workforce and routing design, connect to field-service maintenance operations, where inventory, customer commitments, and real-time dispatch are evaluated together.

Put approved knowledge in the technician’s hands

The mobile workflow should retrieve the current, applicable procedure by asset configuration, serial range, region, revision, and task. Show:

• safety warnings, isolation, PPE, and prerequisites first;
• source document, revision, approval, and effective date;
• step sequence with required measurement and acceptance criteria;
• part supersession and compatibility evidence;
• diagrams or media suited to the device and environment;
• a visible boundary between source text and AI summary.

AI can summarize history, translate approved instructions, propose diagnostic questions, and structure notes. It must abstain when the procedure is missing, contradictory, obsolete, or outside the technician’s authorization. Generated steps should never replace an approved safety or maintenance instruction.

Use frontline industrial copilots to design glove-friendly, offline-capable, multilingual interaction. Test sunlight, noise, gloves, scanning, copy/paste, assistive technology, and long identifiers—not only a clean office demo.

Verify the repair and close the learning loop

A work order is not complete because a note says “fixed.” Require evidence appropriate to the asset:

• isolation and permit closure;
• installed and removed part identity;
• torque, clearance, alignment, pressure, vibration, electrical, leak, or functional measurement;
• before/after condition and operating regime;
• test result against approved acceptance criteria;
• photos or instrument records where authorized;
• technician and independent authorization where required;
• customer or operations handover;
• follow-up monitoring window.

Compare predicted failure mode with confirmed finding. Measure no-fault-found, repeat visit, first-time fix, mean time to repair, recurrence, parts consumption, downtime, and prevented consequence. Feed confirmed outcomes into evaluation only after review.

For safety-critical or regulated assets, keep AI out of the formal sign-off unless the applicable assurance case explicitly authorizes its role. Aerospace MRO practices illustrate why approved data, configuration, traceability, independent inspection, and release authority cannot be collapsed into a summary.

Example: a remote pump with rising vibration

A monitoring service detects increasing vibration on a remote pump, but only under a high-load regime. The evidence shows a growing sideband feature, stable sensor health, and similarity to previously confirmed misalignment cases. The system produces an amber advisory, not a diagnosis.

The reliability engineer checks recent coupling work, operating history, redundant capacity, and manufacturer guidance. The work is planned inside a safe production window. Dispatch selects a technician with alignment authorization and reserves the correct coupling kit and measurement tools.

On site, the technician follows the approved isolation and inspection procedure. The coupling is misaligned, but a proposed bearing replacement is unnecessary. Measurements before and after alignment, part status, final vibration, test regime, and return-to-service approval are recorded.

The model’s failure-family hypothesis is confirmed while its recommended parts list is corrected. The case improves evaluation and planning rules after review. Savings are calculated from the avoided bearing and planned downtime—not a speculative claim that the AI “prevented a failure.”

Measure reliability and service outcomes together

Use a balanced scorecard:

• data availability, latency, calibration status, and sensor-health failures;
• alert precision, recall, lead time, false alarms per asset-time, and missed consequence;
• percentage of alerts with evidence, owner, and timely disposition;
• schedule feasibility, late arrival, travel, utilization, and constraint violations;
• part fill rate, wrong-part rate, emergency shipment, and unused reservation;
• first-time fix, no-fault-found, repeat visit, mean time to repair, and downtime;
• procedure retrieval success, obsolete-document exposure, and technician correction;
• repair verification completeness and unauthorized closure;
• safety event, near miss, environmental consequence, customer interruption, and complaint.

Slice by asset class, site, model, age, sensor, operating regime, technician experience, language, and risk. More alerts, higher utilization, or fewer maintenance hours are not intrinsically better.

Set release and operational gates

Block release when asset hierarchy is unreliable; failure definitions are inconsistent; sensor calibration and operating regime are missing; evaluation uses leakage-prone random samples; critical alerts lack human ownership; the optimizer can violate qualification or safety constraints; procedures are not revision-controlled; or repair closure lacks verification.

Pause or degrade when:

• sensor health or data completeness falls below limits;
• a new asset, regime, configuration, or site is outside validation;
• false alarms, missed failures, or no-fault-found exceed control limits;
• required part, skill, permit, or procedure is unavailable;
• the schedule becomes infeasible;
• a model or threshold changes without evaluation;
• a safety, security, environmental, or customer incident occurs.

Fallback may be fixed-interval inspection, established alarms, manual planning, or an approved conservative maintenance plan. Test it before the AI workflow goes live.

Source notes

Sources reviewed and current as of July 30, 2026:

• ISO 55001:2024 — current published requirements for asset-management systems; it does not prescribe a specific predictive algorithm.
• ISO 13374-1:2003 — general condition-monitoring data-processing architecture, reviewed and confirmed in 2025.
• ISO 17359:2018 — general condition-monitoring program guidance, last confirmed in 2023.
• ISO 14224:2016 — reliability and maintenance data standard for petroleum, petrochemical, and natural-gas industries; its sector scope is explicit.
• NIST 2026 Smart Manufacturing Roadmap — research and adoption roadmap covering industrial reliability, availability, maintainability, and safety, not a maintenance certification.