AI is only as good as the data it’s fed

Barely a week passes without a new AI tool launching for the property sector. Consumption forecasting models, automatic anomaly detection, self-learning energy optimisation. The promises are big, and there are good reasons to believe in them. What’s rarely mentioned is what needs to be in place for any of it to actually work.

Why AI punishes poor data harder than you do

Incomplete energy data isn’t a new problem. A gap in March, a meter reading that arrived a week late, a utility company that changed format without warning – that’s the kind of thing you already spend time correcting, meter by meter, and in the best case you can explain what you did and why after the fact. Let an AI model guess at the same gaps instead, and the guess becomes a black box; you see the result, but not the assumptions made along the way.

A model that needs to forecast consumption or flag anomalies builds its conclusions on patterns in the historical record. If the pattern has gaps – gaps that look like genuine drops, or meter changes that look like real shifts in demand – the model draws a wrong conclusion and does so with exactly the same certainty as if the data had been correct. And you won’t notice the error until it has already influenced a decision.

Four things that have to be correct before AI has anything to work with

Coverage without gaps Not “most of the time”, but continuously, across the entire period you want to analyse. A model trained on the first quarter with a gap in February will unsurprisingly learn the wrong things and draw the wrong conclusions about February.

Consistent structure: The same meter point logic needs to be traceable over time, even when the meter is replaced, the utility company changes format, or a property is added to your portfolio. If the structure shifts partway through the history, comparisons between periods become meaningless.

Traceability: If a model flags an anomaly, you need to be able to see where the value came from and when it was collected, so you can judge whether the anomaly is real or a collection error. Without that traceability, every AI-generated conclusion is a guess you can’t verify.

Validated values: Incorrect readings that were never caught don’t become less incorrect when they’re fed into a model. They just become harder to spot, because the model presents them alongside everything else, with no warning flag.

None of these four points are AI questions. They’re data collection questions that have been there all along (at Metry, we’ve been working on precisely this for over a decade). The difference now is that the consequences of ignoring them become vastly more expensive when an automated system is acting on the data instead of a person who can read between the lines.

But if the foundation isn’t in place, who lays it?

Fill the gaps before AI tries to fill them

Building that foundation manually, meter by meter and utility by utility, is exactly the work that already takes too much time today. It doesn’t make sense for it to take even longer just because the goal is now an AI tool rather than an energy management system. Luckily, that’s what Metry’s data collection can help you with!

With Metry you get visibility into exactly how much data is actually arriving per meter – gaps are identified and resolved without you having to hunt for them manually. Every value that’s collected is also quality-assured and logged with source and timestamp, so you can always trace where a value came from and trust that it’s correct.

Before you invest in the next AI tool for your properties, it might be worth checking first how your current data collection actually looks. Get in touch if you’d like to see how your data really stacks up, meter by meter.