Ask a supplier whether their platform is predictive and the answer is always yes. Ask whether it estimates a time to failure, and with what confidence interval, and the conversation usually changes shape. Most of what is sold as predictive maintenance in commercial property is something else, and the something else is often the better purchase.
What this comes down to
- Rule-based fault detection is mature, cheap and available today. Genuine prognosis is not, on most building plant.
- The useful design parameter is the warning time an asset gives you, not the number of sensors on it.
- ISO 17359, ISO 13379-1 and ISO 13381-1 already define this work, and they long predate the dashboards.
- Before any procurement, settle who acts on an alert and within what deadline. Systems fail on that, not on analytics.
Detection and prediction are not the same product
The distinction is not academic. It determines what you can promise a tenant, what you can put in a budget, and whether an alert has a deadline attached to it.
What fault detection actually does
Automated fault detection and diagnostics compares live plant behaviour against rules describing how the system should behave, and raises a flag when reality diverges. Simultaneous heating and cooling, a damper commanded shut that is clearly open, a valve hunting, a fan running outside occupancy: these are detectable with the points a building management system already has, and no additional sensors at all.
This is well-trodden ground. ASHRAE Guideline 36, High-Performance Sequences of Operation for HVAC Systems, whose 2024 edition absorbed twenty-three addenda to the 2021 version, exists precisely to write sequences that behave consistently enough to be monitored automatically. Control sequences designed for fault detection are the unglamorous prerequisite that most buildings skip.
What prognostics claims to do
Prognostics estimates the remaining useful life of an asset and attaches a confidence level to that estimate. It needs degradation data, meaning a measurable parameter that drifts in a known way toward failure, and ideally a population of similar assets that have already failed so the drift can be calibrated. Manufacturing plants have that. A commercial building with four air handling units and two lifts, monitored for eighteen months, does not.
The interval that decides whether any of this is worth doing
Reliability engineering has a name for the useful window: the interval between the point at which a developing fault becomes detectable and the point at which the asset actually fails. That gap is the entire product. If it is six months, condition monitoring buys real planning freedom, letting a replacement be ordered, scheduled out of hours and carried out without disruption. If it is four hours, a sensor tells you a failure is imminent that you cannot do anything about, and a spare part in a cupboard would have been the better investment.
This is the question to put to every asset on a list before deciding what to instrument. How much warning does this equipment give, in practice, and what could we usefully do in that time? It reorders priorities quickly, and it usually removes several items that looked obvious.
A standards chain that predates every dashboard
None of this was invented by proptech. The ISO condition monitoring family has structured the same work for two decades, and citing it in a specification is a fast way to separate serious suppliers from the rest.
- ISO 17359:2018, Condition monitoring and diagnostics of machines: general guidelines, sets out how to establish a programme at all, starting from equipment criticality and working through parameter selection, alarm criteria, diagnosis and the maintenance decision.
- ISO 13379-1:2012 covers data interpretation and diagnostics techniques, which is the discipline of working out what a symptom means.
- ISO 13381-1 covers prognostics, defining terms including estimated time to failure and the confidence level attached to it. Its latest edition, published in 2025, addresses machine systems rather than isolated machines.
One point in ISO 17359 deserves highlighting because it is routinely lost in software demonstrations: the correct response to a detected fault is a risk-based judgement, not an automatic repair. Increasing monitoring frequency, scheduling work into the next planned shutdown, or accepting the condition and watching it are all legitimate outcomes. A platform that only knows how to raise tickets will quietly generate work you did not need.
Five things to settle before signing anything
- Which assets, ranked by consequence of failure. Not by cost, not by age. By what breaks in the building, and for whom, when that item stops.
- What data already exists. Most commercial buildings have thousands of unused points in the management system. Exhausting those costs far less than a sensor retrofit and answers the question of whether anyone will act on the output.
- Who owns the data, and in what format it leaves. A monitoring contract that cannot export its full history in an open format has quietly bought your building’s operating record.
- Who receives an alert, and within what deadline. Nominate a person and a response time. Programmes rarely fail because the analytics were wrong; they fail because nobody was contractually obliged to read the output.
- How false alarms are handled. Agree in advance what the tuning process is and who pays for it. A system that cries wolf for a quarter is dead for good, however sound its models.
Why the pressure is coming anyway
Even for owners with no appetite for analytics, the monitoring layer is arriving by law. European building performance rules now require automation and control systems in larger non-residential buildings to continuously monitor, log and analyse energy use, benchmark it, detect losses of efficiency and inform whoever runs the plant. Those functions are the exact data foundation that condition monitoring needs, which means many portfolios will hold the raw material before they have decided what to do with it. The implementation side of that sits in our guide to implementing building energy management systems.
Do we need new sensors to start?
Usually not. Rule-based fault detection runs on existing building management system points, and starting there establishes whether the operational discipline exists before any capital is committed. Added sensors make sense once a specific asset has been shown to need a parameter the system does not already capture.
Is machine learning necessary for this?
No, and on most building plant it is premature. Deterministic rules catch the majority of recurring HVAC faults, are explainable to an engineer, and can be argued with. Statistical models earn their place where a parameter drifts gradually and the pattern is genuinely hard to describe in rules.
How long before a programme produces anything useful?
Fault detection typically surfaces findings within weeks, because most buildings are already running with faults nobody has noticed. Anything resembling prediction needs a full seasonal cycle at minimum, simply to know what normal looks like in both summer and winter operation.
The business case, in numbers
If the question is what this does to an operating budget rather than how it works, we have taken that apart separately.
