AI in hotel technical service: work orders, repeat faults and the limits of predictive maintenance

Predictive maintenance uses sensor readings and history to flag a fault before it becomes a breakdown. Most hotels do not have that data. The first useful step is reading closed work orders to find repeat faults by room, asset and season, and adding sensors only where someone can act on the alarm.

The chief engineer knows which air conditioning unit will fail again in August. The work order history knows too, written in free text that nobody reads. The gap between those two facts is what this article is about. After reading it, you will be able to tell preventive from predictive maintenance, run a repeat-fault analysis on your own work orders, judge whether an asset justifies sensors, and keep the statutory calendar intact while you do it.

What is the difference between preventive and predictive maintenance?#

Preventive maintenance is work done on a schedule. Predictive maintenance is work done because a measurement says a fault is developing.

Preventive work follows fixed intervals: monthly filter changes, annual boiler service, seasonal start-up checks. It is cheap to plan and easy to audit, but it replaces healthy parts and still misses faults that appear between two intervals. If you are building or reviewing that calendar, a hotel preventive maintenance plan: what to schedule, check and record is the place to start.

Predictive work follows condition: vibration, temperature, pressure, current draw or run hours compared against a baseline. The decision moves from the planner's calendar to a measurement plus a rule. This is not a smarter calendar. It changes who decides and when, and it only works where a failure gives warning before it happens.

Planned maintenance is the umbrella term that covers both, plus reactive repair. Whatever the method, the output is the same object: a work order assigned to a person, with parts and time. Both methods depend on a work order system that records what was found, what was done and how long it took. Without that record, neither method can be reviewed or improved.

A hotel can run good preventive maintenance for years with no sensors at all. It cannot run predictive maintenance for a week without condition data.

What maintenance data does a hotel actually hold?#

Mostly text and paper: work orders written by staff, meter readings taken by hand, plant round logs, invoices and guest complaints. Continuous condition data is rare.

Work orders are the largest dataset and the messiest. The same fault appears as "AC not cooling", "klima arıza" and "room 412 too hot", with no asset number and no failure code. If you want to see how a complaint becomes a closed job, read hotel maintenance work orders: from a guest complaint to a closed job.

Meter readings exist for electricity, water and gas, usually daily or weekly, often on a clipboard before being typed into a spreadsheet. They describe the building, not the individual asset. Plant round logs carry boiler pressure, chiller temperatures and pool chemistry, written by the operator at fixed hours. They are a real condition record, but sparse and in handwriting.

Manufacturer manuals sit in a folder as PDFs, frequently in another language and for a model the hotel no longer runs everywhere. What is usually missing: an asset registry, consistent failure codes, downtime records and any sensor time series long enough to define normal.

The practical consequence is that any AI step here starts with structuring, not with modelling. A model fed inconsistent text returns confident nonsense. The same principle applies across hotel systems, and it is worth reading about hotel data quality before AI: what to clean up and why before any tool is bought.

How do you find repeat faults in closed work orders?#

Export the closed work orders, tag each one with an asset and a location, group identical faults, and count them by room, asset and month.

Build a short asset list first: chillers, boilers, lifts, pumps, air handling units, and room-level items such as fan coils, minibars and bathroom fittings. Free text cannot be grouped without a target to group it against. Then export everything closed over at least two years, with date opened, date closed, room or area, the description, the resolution and the time to close.

A language model reads each description and proposes an asset, a fault type and a location. This is a draft. A person reviews a sample, corrects it, and the corrected sample tells you how far to trust the rest. Count repeats: the same fault on the same asset three or more times in a season is a repeat fault, not bad luck. That is the signal a replacement or a root-cause check is cheaper than another repair.

Separate repeat faults from repeat complaints. A guest reporting the same noise twice may be describing two different causes, and a technician closing the same ticket twice may be closing a symptom. The output is a short list of candidates, not a dashboard: assets to replace, areas to redesign, procedures to change. A dashboard nobody acts on is a cost with no return.

Can AI draft maintenance procedures from a maker's manual?#

Yes, as a draft. A language model can turn a manufacturer's manual into a short plain-language procedure for your staff, and a competent person must check it against the original before anyone uses it.

What it does well: summarising long documents, ordering steps, translating, turning dense specification tables into a checklist, and writing the result in the language your technicians actually read. What it gets wrong: it drops torque values and warning lines, mixes instructions from two models, and occasionally invents a plausible step that appears nowhere in the manual.

The manual stays the authority. Every draft carries which manual, which revision and which date it was derived from, so a later reader can check it. The value sits in the routine procedures nobody has ever written down, such as filter change intervals, seasonal start-up checks and cleaning routines for equipment that has run for years on habit.

A procedure is a document with an owner and a review date. Without those two fields it becomes folklore within a season. And this is a documentation gain, not a maintenance gain: it reduces the time a new technician spends asking questions, but it does not find faults.

What does predictive maintenance need before it works?#

Three things: sensors on the asset, enough history to know what normal looks like, and a person with the authority and the parts to act when the alarm fires.

Sensors come first: vibration, temperature, pressure, current or run hours, plus a data path from the plant room to wherever the alarm is read. The device is the easy part. History comes second: a baseline of weeks or months before a threshold means anything. A sensor installed last Tuesday cannot say what is abnormal, because it has never seen normal.

Then a rule: who sets the threshold, what counts as an alert, what counts as noise, and who reviews the rule when it is wrong. Then a person: an alarm at three in the morning with nobody on shift is a log entry. It becomes maintenance only when someone owns the response and can stop the asset.

Finally, a loop: alarm, work order, repair, and a record of what the alarm said before the repair. That record is how the rule gets corrected instead of ignored. False alarms are the usual cause of death. After enough of them, technicians stop reading the screen, and the system is paid for and unused.

Which hotel assets could justify sensors, and how do you test before buying?#

Assets where a failure closes a revenue area or breaks a legal duty: chillers and boilers, pool plant, laundry, lifts, and the water and electrical distribution that serves the whole building. Test on one asset before buying for twenty.

Two criteria decide it: the cost of failure, and whether the failure is gradual or sudden. Bearings, fouling, belt wear and valve drift give warning. Control boards and electronics usually do not. Gradual failures suit condition monitoring. Sudden failures are better handled with a spare part on the shelf and a response contract.

Design the test: pick one asset, record its current condition, consumption and repair history, install monitoring, run it for a season, then compare every alarm against what actually happened. Count both directions: alarms that preceded a real fault, and alarms that preceded nothing. Ask for the second number explicitly, and in writing. Write the exit condition before you start, for example that the test ends if the false alarm rate makes the alerts unreadable to the shift.

The same discipline applies to any vendor promise about meters and savings: ask for the mechanism and the measurement, not a percentage. The approach is set out in detail in AI energy management in hotels: how to test a vendor's promise on your own meters.

What stays fixed: periodic inspections and safety sign-off#

Legally required periodic inspections of lifts, boilers, pressure vessels, electrical installations and fire systems stay on their calendar, with a qualified person's signature. No model replaces them.

These are duties with a document and a responsible person attached. They are not maintenance optimisations and they do not move because a sensor says the asset looks fine. Condition data can help you prepare: a lift with rising door-cycle faults is worth flagging before the inspection, and a boiler with drifting pressure is worth a conversation with the service firm.

Keep one calendar that holds statutory inspections, manufacturer service intervals, condition alerts and seasonal openings, so that none of them hides behind another. The structure is described in periodic inspections in hotels: boilers, lifts, electrics and fire systems on one calendar. Record who signed, when, and what was found: the record is the evidence, and it is also the input to next year's plan.

Scope, dates and penalties differ by asset type and by facility. Confirm your own case with your technical and legal advisers rather than relying on a general article. AI can summarise the inspection reports and draft the follow-up work orders. It does not inspect, and it does not sign.

What to do next#

Export the last two years of closed work orders and tag them by asset, room and season before you buy a single sensor. That one step tells you which faults repeat, which assets are candidates for replacement, and whether condition monitoring has anything to monitor. The calendar of statutory inspections and the safety sign-off stay exactly where they are, with a qualified person's name on them.