AI in Maintenance Only Works When You Know Why

Somewhere in every facility management meeting this year, someone has said a version of the same sentence. We should use AI for this. It arrives with a lot of hope attached. The hope is that AI will look at the maintenance operation, see what nobody else can see, and hand back a plan that pays for itself. An oracle, more or less.

That is not what happens. What usually happens is that the team spends three months connecting tools, generating dashboards, adding a scorecard here and a recommendation panel there, and at the end someone quietly asks what all of it was for. The models were fine. The data was thin, and the direction was missing.

This post is about the question that has to come before any AI initiative in maintenance, which is simply: why? Not why AI, but why this problem. Everything else follows from getting that part right.

Solutions Are Cheap Now. Problems Are Not.

For most of the history of facility management, the expensive thing was producing the output. A monthly report took days. Consolidating work orders from a dozen sites took a coordinator and a lot of Excel. A trend analysis needed someone who knew both the equipment and the spreadsheet.

AI has collapsed the cost of all of that. Reports, summaries, recommendations, consolidation, and pattern spotting are now close to free. Any problem you can describe has more candidate solutions than it did two years ago, and they can be produced in an afternoon.

When solutions become cheap, the scarce thing shifts. It stops being the ability to build and becomes the ability to choose. Every AI initiative now needs an honest answer to one uncomfortable question: are we building a solution to a problem that does not exist?

Listen for how the initiatives get phrased. We should create predictive maintenance schedules using AI. Can the system predict when the next asset will break down? Can AI generate a set of initiatives to improve staff efficiency? Each of these sounds like progress, and each of them is ambiguous. Predict which asset, with what data, so that who can do what differently? Improve efficiency measured how, against what baseline? Without answers, the initiative is a solution looking for a home.

What AI Says When Nobody Gives It a Goal

Try it. Ask an AI assistant what the best next step is to give customers a better experience in your outlets. It will tell you to keep the stores clean, train the staff, respond to complaints quickly, and invest in marketing. All of it true. None of it useful. You could have written that list yourself in the time it took to type the question.

The answer is generic because the question was generic. There was no goal, no metric to measure against, no context about which outlets, which assets, or what “better” would look like in numbers. AI does not invent intent. It sharpens whatever intent it is given, and when it is given none, it produces the average of everything it has ever read.

This is the part that cannot be delegated. Deciding the scope, the direction, and the reason a problem is worth solving is human work. Not all problems are worth solving. Some are annoying but cheap. Some are expensive but nobody will act on the answer. The judgement about which ones matter is exactly where the thinking has to happen, and it is the step most teams skip on their way to the tooling.

A Worked Example: One KPI Across a Hundred Outlets

Here is what a purposeful version looks like. Imagine a retail chain with a hundred outlets across the country. Management has decided that the customer's experience of the brand is physical: the store should feel maintained, professional, and pristine, every time. So they set one golden metric. Any maintenance issue raised in any outlet must be resolved within two hours.

Notice how much intent is packed into that one number. It says what the company believes about its customers. It says what minimal disruption means in practice. It gives every outlet in the country the same target, and it connects a technician fixing a flickering light in one branch to the mission of the whole business. This is a KPI worth measuring, and everyone can say why.

Now look at the cost of measuring it. Every issue needs a timestamp when it was raised, a timestamp when it was closed, and the repair updates in between. Across a hundred outlets that is thousands of events a month, tracked in WhatsApp groups, phone calls, and an Excel sheet per region that nobody fully trusts. The metric is right. The process of accumulating it is maddening.

That is the problem. Not “we need AI” but “we have a metric that matters and we cannot see it without a week of manual work.” Once the problem is stated that plainly, the sequence of what to do about it becomes obvious.

Digitise First, Automate Second, AI Third

The order matters, and it is the same order every time.

• Digitise the data. Every issue raised in every outlet goes into one system, with the time it was reported, who picked it up, what was done, and when it was closed. No Excel, no chat groups as the system of record. Until this exists, there is nothing for AI to work with. A model cannot predict anything from a spreadsheet that is half filled in.
• Automate the process. Once the data is captured at the source, the two-hour clock runs itself. Escalations fire when a job is at risk of breaching. The monthly KPI report assembles itself from the records instead of from a coordinator's weekend. The manual accumulation that made the metric so expensive disappears.
• Then apply AI, with a target. Now the question is sharp. Which outlets are trending toward breaching the two-hour target, and why? Which asset types account for most of the misses? Which regions are improving, and what did they change? Draft the monthly narrative, flag the anomalies, recommend where to focus next month. Every one of those outputs is measured against a metric the company already believes in.

Compare the two versions side by side. In one, AI is asked to recommend the best next step for customer experience and produces a list of platitudes. In the other, AI is asked to explain movement in a specific KPI, on clean data from a hundred sites, in service of a mission everyone can state. Same technology. Completely different value. The difference was never the model. It was the clarity of the problem it was pointed at.

The Blade and the Purpose

A useful way to hold all of this: AI is a sharpener, not a blade. It will make whatever you hand it sharper, faster, and cheaper. It will not tell you what the blade is for. Hand it a vague wish and you get a very polished vague answer. Hand it a defined problem, a metric, and clean data, and it becomes one of the most powerful tools a maintenance team has ever had.

So before the next AI initiative gets a budget, run it through three questions. What problem is this solving, stated in one sentence? Why is that problem worth solving, in terms the business already cares about? And do we have the data to see the problem today, or is the first job still digitisation? If the answers are not clear, the initiative is not ready. That is not a failure of AI. It is the thinking that has to happen first, and it is the part that stays with us.