Predictive maintenance is one of the easiest AI use cases in real estate to understand. Find the warning signs before equipment fails. Fix the problem in a planned service window instead of at 2 a.m. Reduce downtime, emergency calls and premature replacement.

It is a compelling idea, and it is something we have been working on directly.

The practical work exposes a dirty secret fairly quickly. Before you can predict the failure of a piece of equipment, you have to know what equipment you own.

When we document an existing building, we are often the first people in years to walk every mechanical room and write down what is actually in it.

The list we come out with rarely matches the list the owner started with.

The asset inventory problem

Ask a portfolio owner for a standardized inventory of major mechanical equipment across every building and the answer arrives from several directions at once.

The CMMS has part of it. The building automation system has another part. A condition assessment has more. Someone has a spreadsheet. The drawings show equipment that may or may not still be there. Names and asset IDs do not reliably match between any two of those sources.

None of that is a criticism of building teams. It is what happens when systems are installed over decades, for different purposes, often under different owners, by people who had no reason to think the records would one day need to agree with each other.

Predictive maintenance depends on joining those fragments and reconciling them with the physical building.

A sensor reading needs context

Sensors are cheap and easy to deploy now. Vibration, temperature, current and pressure all reveal changes in how a piece of equipment is behaving. On its own, though, a vibration number means very little.

Which motor produced it? What is that motor driving, and what does that equipment serve?

How old is it, and what is normal for that model? Was it serviced last week? Are five similar units in other buildings showing the same pattern? The sensor gives us a signal. The building record is what gives the signal meaning.

In this country, consequence has a calendar

New technology always tempts us to put sensors on everything, and for most existing portfolios that is the wrong starting point. Start with consequence instead. Which failures are expensive? Which interrupt tenants or critical operations? Which failure modes actually give us a warning we can detect?

In Canada those questions have an answer that most predictive maintenance material never accounts for, because here the consequence depends heavily on the date.

A heating plant that fails in January is not the same event as the same plant failing in June. One is a maintenance item. The other is a tenant relations and life safety problem that can escalate within hours.

The same logic runs through the year.

A chiller that goes down during a July heat warning carries a cost no maintenance budget captures. Freeze-thaw shortens the life of almost everything on a Canadian roof, and it does its damage in a narrow band of the year. Make-up air and glycol systems fail in ways a portfolio in a mild climate never has to plan for.

So the instrumentation priority is not simply the most expensive equipment. It is the equipment whose failure lands in the worst month, and it is worth knowing which of your assets those are before you spend anything on sensors. The technology should follow the risk. Here, the risk has a season.

The portfolio is where this gets interesting

One building produces experience. A portfolio can produce a learning system. If a pump fails in Building A, the history leading up to that failure may identify risk in similar pumps in Buildings B, C and D. Maintenance records, sensor trends, age, operating conditions and outcomes all become training data.

This is the point where building intelligence becomes more than another dashboard, because the portfolio begins learning from itself. It is also how an operator’s knowledge that a particular unit always does something odd before it fails becomes information the whole organization holds, rather than something that walks out the door at retirement.

Given the age of much of our building stock, and the age of many of the people who have been running it, that transfer is more urgent than it sounds.

A prediction has to lead somewhere

A prediction by itself does not save anyone money. Someone still has to decide what to do about it.

A mature system should observe a change, assess the risk, recommend an action, schedule the intervention, record the result and then learn whether the prediction was right. That closed loop matters far more than the AI label on the front of it. It also keeps people in the process, which is the part most often missed.

Experienced operators bring judgment algorithms do not have. What AI gives them is the ability to watch more equipment, across more buildings, than any individual could.

Start with the boring work

The first step toward AI-driven predictive maintenance may not be an algorithm at all. It may be an equipment inventory.

Verify what exists. Standardize the names. Connect assets to the spaces and systems they serve. Bring the maintenance histories together. Identify the critical equipment, then decide what deserves monitoring.

That work is considerably less exciting than saying a building can predict the future. It is also how the future gets built.