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

A maintenance case must identify the affected asset

An occupant may report a warm room while the building-management system records a temperature alarm and a technician notes a recurring valve fault. The three records may describe one problem. However, inconsistent room, equipment and system identifiers can split them into separate cases.

Facilities teams must first establish the affected location, asset and service before deciding how to respond. Otherwise, a technician may inspect a similar unit while the failing asset continues to deteriorate. Reliable asset identity therefore determines whether later diagnosis, dispatch and verification refer to the same physical system.

Maintenance evidence is divided across systems

Building models describe the asset and its design. Sensors record changing operating conditions. Maintenance systems record faults, work orders, parts and technician observations. Occupant reports describe the service consequence. No source alone establishes the complete condition.

The maintenance team must connect these records while preserving their source and time. A recent alarm may contradict an older work-order note, and a technician may have recorded a temporary repair rather than a permanent resolution. The combined record should expose the disagreement instead of selecting the most convenient status.

Use the agent to prepare the maintenance case

An agent can retrieve the verified asset record, assemble recent alarms and service history, check technician skills and spare-part availability, and prepare a work order. It can also coordinate access and keep the case active while staff complete diagnosis and repair.

This role addresses a specific coordination gap: employees often spend time finding records and transferring the case before a qualified person can decide what to do. The agent should accelerate that preparation. It should not classify a life-safety risk, isolate critical equipment, approve regulated work or authorise material expenditure without the competent person required by the facility's procedures.

Define the states from detection to verified restoration

The workflow should distinguish detection, validation, diagnosis, approval, dispatch, repair and verified restoration. Each state requires named evidence. A technician acknowledgement does not prove that the technician arrived, a completed work order does not prove that the asset recovered, and a cleared alarm does not prove that occupants received the required service.

Management must define the evidence that moves the case from one state to the next. For example, restoration may require stable temperature and valve performance over a specified period rather than a closure note. These state rules prevent the agent from describing administrative completion as physical recovery.

Keep safety and engineering authority outside the agent

Safety permits, statutory inspections, lockout procedures and engineering approvals impose fixed constraints on the maintenance route. The workflow engine should enforce those constraints before any technician or contractor receives an instruction.

The agent may retrieve the relevant procedure and present the evidence to the authorised engineer. However, it must not infer that a missing permit is unnecessary or treat an earlier approval as valid for a changed condition. The case record should identify the person who approved the action, the equipment state covered by that approval and the time at which the approval was valid.

Test failures that could produce an unsafe or incomplete repair

The evaluation team should test wrong-asset matches, stale sensor readings, suppressed alarms, conflicting maintenance histories, unavailable parts, access denial and contractors without the required competence. The team should also test an apparently successful repair followed by renewed deterioration.

Each case must end in an observable asset and service state. The evaluator should inspect the alarms, work order, permit, technician action and post-repair measurements. A fluent explanation cannot compensate for a missed safety escalation, a repair on the wrong unit or a case closed before performance recovered.

Example: a recurring chilled-water valve fault

A building records warm-room complaints on two floors. The building-management system shows intermittent temperature alarms, while the maintenance system contains three earlier work orders for valves with similar names. One previous case was closed after a manual reset.

The agent links the room complaints to the verified air-handling unit and retrieves the relevant alarm history, valve record and earlier technician notes. It identifies that the previous closure did not include a sustained performance check. The agent prepares a case for the facilities engineer, checks technician and part availability, and waits for the required access and safety approval.

After the repair, the workflow keeps the case open until temperature and valve behaviour remain within the approved range for the required period. If the alarm returns, the agent reopens the same asset case rather than creating another unrelated complaint. The facilities engineer retains the decision about isolation, repair method and return to service.

Pilot one asset class and measure verified recovery

The first pilot should use one asset class with reliable identifiers, sufficient maintenance history and a measurable restoration state. The team can replay known faults and then operate the agent in shadow mode while staff follow the current route.

The comparison should measure time to accountable ownership, wrong-asset intervention, missed escalation, technician preparation, first-time repair, repeat fault, downtime and verified restoration. Results should be separated by asset criticality and operating condition. Management can then decide whether to expand the workflow, correct the asset data or retain the existing process.

How Marketways designs the maintenance workflow

Marketways maps the selected asset family, reconciles identifiers and reconstructs the path from signal to verified restoration. We define the required evidence, case states, safety boundaries, approval points and failure scenarios before designing the agent.

Marketways then builds the pilot workflow, controlled test cases and decision scorecard. The resulting evidence shows whether the agent reduces avoidable coordination delay without obscuring asset condition or transferring safety and engineering authority away from competent people.