Step by step: which parts of planned preventive maintenance (PPM) are rules, where AI helps with sensor data and fault reports, and what competent people must keep.
Partly, and mostly with plain rules rather than AI. Planned preventive maintenance (PPM) software creates recurring jobs on a schedule, AI can read sensor data and engineers' notes to flag problems early, and a competent person still sets the intervals and signs off anything safety-critical. If you are weighing up preventive maintenance software for a workshop, site or fleet, the useful question is which steps to hand over and which to keep. Below, every step from the asset register to the final sign-off gets an honest label, the conditions that must hold first and a small pilot for one type of equipment. The aim is fewer surprise breakdowns and less paperwork for the maintenance team, not a site that nobody watches.
What each step needs
01● Standard automation
Keep an equipment register with maintenance history
Everything starts with a clean list of what you own: asset, location, model, serial number, install date and every past job. This is a database task, not an AI task. A rule creates the record when an asset is bought, links each work order to it and refuses a job that names no asset. AI can help one time, by reading old invoices, manuals and spreadsheets to propose the first draft of the register, but a person checks it. Without this register every later step guesses, and a model asked to plan maintenance for assets it cannot identify will simply invent plausible ones.
⛨Works when each asset has a unique ID, a named owner and a location, and every job is logged against that ID.
02● Human review
Set maintenance tasks and intervals for each asset type
Someone who knows the equipment decides what gets checked and how often. Manufacturer manuals, past failures, the operating environment and legal inspection duties all feed the answer, and the cost of getting it wrong is real: too little causes breakdowns, too much wastes engineer time. AI can summarise a manual into a draft task list and point out where two sources disagree, which saves reading time. It should not set the interval. A maintenance lead reviews the draft, signs it off once per asset type and revisits it after every unexpected failure.
⛨Works when manuals and failure history are available to the reviewer and one named person owns the plan for each asset type.
03● Standard automation
Generate and schedule recurring work orders
Once the plan exists, creating the jobs is plain scheduling. A rule turns each task and interval into a work order on the due date, or after a set number of running hours or cycles, assigns it to a team and reminds the owner when it is overdue. Most maintenance tools already do this and it is predictable, which is exactly what you want. Do not hand this to a language model: a missed inspection because a model got a date wrong is a compliance problem. Keep the rule simple, log every generated order and review the overdue list weekly.
⛨Works when intervals are stored as data, not free text, and the tool can trigger on dates and on meter readings.
04● AI candidate
Spot early warning signs in sensor and usage data
Where machines produce data such as vibration, temperature, current draw or run hours, AI can compare today's readings with the asset's own history and flag drift before a breakdown. This is the condition-based part of maintenance and the area where machine learning has a real record. It also needs the most care. A model trained on too little history raises false alarms until people ignore it, and a model that stays quiet gives false comfort. Treat each alert as a suggestion with the underlying reading attached, and let an engineer decide whether to inspect.
⛨Works when an asset has months of reliable readings, a clear failure history to learn from and an engineer who reviews every alert.
05● AI candidate
Triage fault reports and draft work orders from technician notes
Operators report faults in messy words: a photo, a voice note, 'machine 4 is making a noise again'. A language model can match the report to the right asset, suggest a category and urgency, pull related past jobs and draft a clear work order. After the job it can turn an engineer's rough notes into a tidy close-out record with parts used and time spent. The output is a draft that a person confirms. Models misread urgency, so anything that mentions smoke, leaks, guards or injury skips the queue and goes straight to a person.
⛨Works when assets can be identified from the report, urgency categories are written down and safety words trigger an immediate human escalation.
06● Standard automation
Reorder spare parts and book technician time
Parts and people are constraints, and constraints are rules. When a planned job needs a part, the system checks stock, reserves it or raises a purchase request when stock falls below the minimum. Scheduling then matches the job to an engineer with the right skills and certificates and a free slot. AI can suggest a better order of jobs when many are due at once, but the hard limits, such as certified staff only, come from rules. A planner approves the weekly schedule and changes it when production needs a machine at a bad moment.
⛨Works when stock levels, minimum quantities and engineer skills are kept up to date in one system.
07● Keep human
Do safety-critical work and sign off the result
Isolating energy, working on pressure systems, lifting equipment, electrical work and statutory inspections under PUWER and LOLER stay with competent people. They carry legal responsibility and need judgement in the room, not a generated checklist. AI may help by preparing paperwork, but the competent person verifies the machine is safe, signs the record and decides when it returns to service. Keep this step explicit in the plan so nobody assumes the software has cleared a machine. The sign-off is also the moment to record anything surprising, which feeds the next review of intervals.
⛨Applies to every job covered by PUWER, LOLER or an insurer's inspection rules, and to any job where a mistake can injure someone.
A sensible first experiment
Pick one asset type, for example your forklift trucks or compressors, and run for six weeks. Load the register, let rules generate the recurring work orders, and have a model draft work orders from operator fault reports for a planner to confirm. Measure three things: the share of planned jobs done on time, the number of unplanned stoppages for that asset type, and how many AI drafts the planner had to correct. Continue only if on-time completion goes up and corrections are mostly about wording, not about the wrong asset or urgency. Then add sensor alerts for the one asset where you hold the most history.
The trap to avoid
The most common mistake is buying a predictive maintenance model before the basics work. If the register is incomplete and half the jobs are never logged, the model learns from gaps and its alerts mean little. The second mistake is letting the software mark safety checks as done. Fix the register and the recurring schedule first, then add AI to one step.
Questions teams ask
Can AI plan equipment maintenance on its own?
Not safely on its own. Rules generate the recurring jobs, AI can read sensor data and technician notes, and a person owns the plan and signs off safety-critical work. A fully autonomous plan would mean nobody checks the intervals or the exceptions, and that is where breakdowns and compliance gaps appear.
Do I need AI for planned preventive maintenance?
Often not. Most of preventive maintenance is calendar or meter-based scheduling, which a maintenance management tool does with plain rules. AI adds value on top where you have sensor data to read or a lot of free-text fault reports to sort. If you have a small fleet of machines, a good register and recurring work orders may be all you need.
What data does predictive maintenance need?
Reliable readings over a long enough period, plus a record of when the asset actually failed or was repaired. Without failure history the model has nothing to learn from. Start with one critical asset type, check that sensors are calibrated and log every alert with the outcome so you can see whether the warnings were right.
How is maintenance software different from AI?
Maintenance software stores assets, schedules work orders and tracks parts and history, and it behaves the same way every time. AI is an optional layer for reading messy text and noisy data. Buy or set up the software first, then add AI to one step where the input is unstructured and a person reviews the output.
Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.