Industry application

How to Evaluate Agentic AI for Facilities and Maintenance Workflows in Dubai

A facilities agent should convert asset evidence into a controlled maintenance response without confusing an alarm with a diagnosis, a dispatched technician with a completed repair or an acknowledged work order with restored service. Evaluation must connect every proposed action with the correct asset, fault evidence, safety classification, authorised person and verified return-to-service condition.

Separate fault detection from maintenance coordination

A predictive model and an agent perform different functions. The model may detect an anomaly, estimate failure risk or classify a likely fault from sensor and maintenance history. The agent may then collect supporting evidence, create a work order, find an authorised technician, check parts and access, coordinate approval and follow the case to completion. A useful evaluation tests each function separately before testing the complete sequence.

This distinction matters because a strong prediction can still lead to a poor maintenance decision. The agent may attach the alarm to a similar asset, ignore a conflicting inspection, dispatch someone without the required competence or close the order after acknowledgement. Conversely, weak or missing sensor evidence may make confident coordination unsafe even when the agent follows every administrative rule.

Define the asset and action boundary

The release claim should cover one asset family, a defined set of sites and a limited range of faults. The organisation should document the asset hierarchy, alarm sources, maintenance history, operating conditions, access restrictions, contractor pool, spending limits and evidence required to confirm restoration. Without that boundary, an average result can combine assets with different failure modes and safety consequences.

The authority boundary must be equally clear. The agent may prepare a work order, suggest a priority under approved rules, identify eligible technicians and request parts or access. Competent people must retain safety classification, isolation, permits, statutory inspection and safety-critical return-to-service decisions. The evaluation should test every attempted action outside the agent's authority.

Recreate time-ordered asset evidence

Maintenance decisions depend on sequence. Three minor alarms across separate shifts may have a different meaning when followed by a temperature excursion or repeated operator complaint. The test library should therefore preserve raw timestamps, asset identifiers, sensor readings, inspection notes, prior repairs, overrides and environmental or occupancy conditions.

Construct cases with late evidence, duplicated alarms, missing intervals, out-of-order events and simultaneous faults on similar assets. The agent should update its assessment when new evidence arrives while preserving the earlier record. A summary that removes the weak signals may prevent an engineer from understanding why the later event requires escalation.

Test asset identity and sensor uncertainty

Facilities portfolios often contain similar equipment names, nested asset structures and identifiers that differ across building-management, maintenance and finance systems. Evaluation should include records in which a sensor, work order or parts record points to the wrong parent or sibling asset. The agent should request resolution rather than merge evidence on the basis of a similar label.

Sensor evidence also needs qualification. A missing reading may indicate a healthy asset, a communications failure or an unavailable sensor. An intermittent alarm may reflect a developing fault, an operating transition or measurement noise. The agent should expose these alternatives and route uncertain cases to the responsible engineer instead of presenting one unsupported diagnosis as fact.

Test safety and authority before dispatch

A fast dispatch is not a successful result when the assigned person lacks the required competence, permit, access or isolation authority. Tests should vary technician qualifications, contractor status, site access, hazardous conditions, permit requirements and statutory inspection needs. The agent must block or escalate any assignment that fails a mandatory condition.

Safety signals should override ordinary efficiency targets. Create cases in which a low-cost or nearby technician is available but cannot perform the task safely, and cases in which several weak signals together require urgent professional review. Record missed escalation, false emergency, unauthorised assignment and attempted permit bypass as separate failures because they require different controls.

Test parts, access and changing constraints

Maintenance coordination often fails after diagnosis because the required technician, part, access window or shutdown approval is unavailable. The test environment should change these conditions after the agent has proposed a plan. A robust agent should recalculate the route, preserve ownership and explain which constraint forced the change.

The agent should not solve a resource conflict by silently substituting an incompatible part, unqualified contractor or prohibited time window. It should compare only approved alternatives and refer unresolved trade-offs to the responsible facilities manager or engineer. These tests show whether the agent coordinates the operating system or merely produces a plausible schedule.

Verify restoration before closing the work order

Work-order acknowledgement confirms receipt, not repair. A technician's completion note may also be insufficient when the asset requires a functional test, sensor recovery, inspection or stable operation over a defined period. The organisation should specify the evidence that closes each fault class.

Evaluation should include premature completion notes, repeat alarms, unsuccessful repairs and a receiving system that fails to update. The agent must keep the case open, request the missing verification and connect repeat faults with the previous intervention. Wrong closure can hide downtime, distort reliability measures and delay the response to a continuing hazard.

Combine model, agent and workflow measures

The predictive model should be assessed on the fault or risk claim it makes. Relevant measures may include detection, missed failure, false alarm, calibration, warning time and performance across operating conditions. The agent should be assessed on asset matching, evidence use, authorised routing, dispatch quality, state accuracy and recovery from changing constraints.

The complete workflow then needs operating measures such as verified time to restoration, first-time fix, repeat failure, avoidable dispatch, technician correction, downtime and total cost. Report results by asset class, fault pattern and data condition, with denominators and uncertainty. Rare safety misses and wrong-asset actions require individual review even when aggregate performance appears high.

Define release and monitoring rules

Initial deployment should limit the agent to the asset family, fault types, sites and actions supported by the evidence. Shadow triage or work-order drafting may provide a suitable first boundary. Manual dispatch, safety decisions and verified rollback should remain available until the organisation has evidence for wider authority.

Predefined pause rules should cover wrong-asset intervention, missed safety escalation, unauthorised assignment, permit bypass and premature closure of a continuing fault. Monitoring should add new failures, engineer overrides and changing operating conditions to a time-ordered regression set. Expansion should follow demonstrated control of a wider boundary, not elapsed time in service.

What Marketways can evaluate

Marketways can define the asset and authority boundary, assemble a time-ordered fault library and test the predictive model, coordinating agent and complete maintenance workflow. We can replay sensor gaps, similar identifiers, unavailable skills, missing parts, access failures and unsuccessful repairs, then trace every recommendation and action to its evidence.

The evidence pack can report model performance, missed escalation, wrong-asset action, unauthorised dispatch, premature closure, first-time fix, repeat failure, downtime and verified restoration. Marketways supports technical and management evaluation. Responsible facilities professionals, competent persons and applicable authorities retain safety, permit, statutory inspection and return-to-service decisions.