Role Playbooks4 min read

Use AI to Triage Maintenance Work Orders Without Inventing the Fault

A role playbook for facilities teams: organise symptoms, evidence and missing fields with AI while keeping diagnosis, safety priority and authorisation with competent people.

Bokili Editorial· Verified November 11, 2026
ShareX
A maintenance coordinator moves work orders through observed symptom, evidence, human priority and technician diagnosis checkpoints.

Maintenance coordinators often receive work orders that mix a symptom, a guess and a requested fix: “air handler failed again—replace the motor urgently.” An AI assistant can help normalise the notes, group likely duplicates and expose missing fields. It should not turn that request into a diagnosis or a safety priority. The useful workflow keeps what was observed separate from what a competent person still needs to inspect.

The operating rule

Let AI organise the request. Let a competent person diagnose the fault, assess the hazard and authorise the work.

Why triage needs a visible evidence boundary

The UK Health and Safety Executive says maintenance must be planned and carried out safely by people with the necessary skills, knowledge and experience. It also notes that faulty equipment can create hazards and that maintenance records must stay current. That makes a work-order queue more than an administrative list: errors in symptom, location, asset or urgency can send the wrong person to the wrong problem.

The UK government’s facilities-management asset-data standard reaches the same practical conclusion from a data angle. Maintenance and investment decisions need complete, accurate and up-to-date asset information, consistent change control and named accountability. NIST’s Generative AI Profile adds the model risk: generative systems can state false or unsupported content confidently. A plausible diagnosis is not evidence that the fault exists.

The SEAT triage card

1

S — Symptom

Record only what the reporter observed: sound, smell, temperature, leak, alarm, loss of function or visible damage. Preserve the reporter’s words where they matter.

2

E — Evidence

Attach the asset ID, location, time, photo, alarm code and previous work-order references. Mark missing evidence instead of filling the gap.

3

A — Action and safety

Apply only pre-approved immediate actions, such as isolating an area or stopping use. A competent person sets safety priority and authorises maintenance.

4

T — Technician question

Turn every unproven cause into a question for inspection: “Is the motor drawing abnormal current?” rather than “motor failure.”

Use AI for structure, not fault finding

A controlled work-order workflow

  1. 1

    1. Prepare a safe batch

    Export only the fields needed for triage. Remove unnecessary personal information and use the organisation’s approved AI environment.

  2. 2

    2. Extract observations

    Ask for one row per request with asset, location, time, observed symptom and source. Require “missing” when a field is absent.

  3. 3

    3. Group without diagnosing

    Group exact duplicates or similar symptoms for human review. Keep different assets and conflicting reports separate.

  4. 4

    4. Surface the queue questions

    List missing asset IDs, repeated faults, safety words, access constraints and overdue inspections. Do not ask the model to assign a legal or safety classification.

  5. 5

    5. Have the coordinator decide

    A named coordinator checks the source, applies the approved priority matrix and sends technical questions to the competent maintenance person.

Worked example: three warm-room reports

A fictional office submits three requests. One says “meeting room 4 is hot again.” A second says “fan broken on level 2.” A third records a temperature of 28°C beside asset AHU-2-17 and notes an alarm code. A rushed summary might combine all three and diagnose a failed fan. The evidence does not support that conclusion: the rooms, asset and timing may not match, and one report contains no asset reference.

Reading is a start. Practice makes it stick.

Start learning
Unsafe AI triageEvidence-preserving triage
IssueFan failure affecting level 2Three reports of heat or airflow symptoms; possible relationship not established
EvidenceRepeated complaintsOne temperature reading and alarm code; two reports need asset/location confirmation
PriorityUrgent motor replacementCoordinator applies the approved priority matrix after checking occupancy, hazard and asset criticality
Next stepDispatch replacementTechnician inspects AHU-2-17; coordinator asks whether the other reports map to the same system

The second version is deliberately less decisive. It gives the coordinator a usable queue, keeps the alarm code attached to its source and stops a guessed cause from becoming a purchase or repair instruction.

A prompt that preserves the diagnosis boundary

Triage the records, not the fault
Using only the approved work-order records below, create a SEAT card for each request: observed Symptom; Evidence and source; permitted immediate Action or safety flag from the supplied policy; Technician question. Write MISSING for absent fields. Do not diagnose a cause, merge different assets, assign a safety category or recommend a repair. After the cards, list possible duplicates for a human coordinator to review and explain which fields still need confirmation.
A reviewable queue with source-linked symptoms, explicit gaps and questions for a competent person.

Use fictional or approved data. The prompt is a drafting aid, not a maintenance or safety assessment.

Release gate before dispatch

  • Every request has a source, location and asset reference—or a visible missing-field flag.
  • Observed symptoms are not rewritten as causes.
  • Possible duplicates retain their original work-order IDs.
  • Safety and urgency come from the approved matrix and a named human decision.
  • The technician receives the evidence, access constraints and unanswered questions.
  • The final diagnosis, repair and closure note are added by an authorised person.
  • Changes to asset data and priority are logged with an owner and time.

Try the method in ten minutes

Build a six-order practice queue
  1. Write six fictional work orders: two duplicates, one missing asset ID, one ambiguous symptom and one unsupported diagnosis.
  2. Run the SEAT prompt in an approved AI tool.
  3. Check whether every output preserves the source and uses MISSING rather than guessing.
  4. Remove any diagnosis or priority that did not come from an approved human rule.
  5. Ask a maintenance colleague which technician question would make the queue more useful.

Connect triage to the wider work system

This workflow complements Bokili’s guides to verified incident timelines (https://bokili.com/en/learn/ai-incident-notes-verified-timeline), evidence-tagged budget commentary (https://bokili.com/en/learn/ai-budget-variance-evidence-commentary) and AI work handoffs (https://bokili.com/en/learn/ai-literacy-training-business-ai-work-handoff). Each keeps structure useful without allowing fluent language to erase uncertainty or ownership.

The maintenance coordinator’s contribution is not a machine diagnosis. It is a clean, traceable queue that sends the right evidence and question to the person qualified to decide. Use AI to reduce clerical friction, then keep technical judgement, safety priority and authorisation where they belong.

Practise this in Bokili

Turn what you just read into a ten-minute mission.

  • Build a SEAT work-order card

    Practise separating an observed symptom, evidence, approved action and technician question in ten minutes.

    Start learning

Sources

  1. Maintenance of work equipment — Health and Safety Executive
  2. Facilities Management Standard 002: Asset Data — UK Government
  3. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — NIST
ShareX

Reading is a start. Practice makes it stick.

Bokili turns skills like this into ten-minute missions for your whole team, with instant feedback and progress you can see.

Start learning

Keep reading