AI in a construction company: where it pays off and where it is expensive decoration

When we built Materia we had one easy option: put “AI system for construction production management” on the website and let the customer fill in the blanks. It would have worked. That sort of thing sells.

We did not do it, because nine tenths of what the system does contains no artificial intelligence whatsoever. And that is a good thing.

The rule we use

We reach for AI on tasks that have many possible inputs and one correct answer that a human can verify instantly.

Everything else we do with rules.

It sounds like a detail. It is the entire difference between a system people trust and a system they quietly stop using after six months.

Where AI pays off: reading documents

A supplier’s price quote might arrive as a clean PDF export from their system, as a scanned sheet with a stamp on it, as an Excel file with its own layout, or as a photo taken on a phone. The inputs are effectively infinite.

But the correct answer is always the same structure: article code, description, quantity, unit price, total price, validity. And when the model gets it wrong, a human spots it immediately — the number either matches the PDF or it does not.

This is exactly the case for it. Standard quotes are read by an ordinary parser. Only when that struggles does a language model step in — and the result is badged in the app so you can see where the data came from.

Where AI does not pay off: decisions that must be explainable

“What should we order next week” looks like a job for AI. It is not.

The answer can be computed: look at the construction phase, the order of items, the date of the last quote request, and the lead time from the catalogue. It is a handful of rules. More importantly — when the system gets it wrong, you have to be able to say why.

On a three-hundred-thousand order, “the system recommended it” is not enough. You need to be able to say “the system proposed this because the roof phase starts in nine days and the item has no quote request date”. The second can be checked and corrected. The first can only be believed or ignored — and after a few weeks people start ignoring it.

Where AI does not pay off: price prediction

We have heard this from several people: “can it predict where material prices are going?”

It cannot, and we do not want it to. The price of roof covering in three months depends on things that are not in the data — exchange rates, energy costs, whether the supplier decides to run a promotion. A model would produce a number, because models always produce a number. It just would not be worth anything.

What is worth something is knowing what the price is now and what it was last month. That is not a prediction, it is a measurement. And in a supplier negotiation it is a far stronger argument than an estimate.

A practical test before you automate anything

Before putting AI on a task in your company, try answering three questions:

1. Can it be computed with rules? If yes, compute it with rules. It will be cheaper, faster and explainable.
2. Can a human tell within ten seconds that the output is wrong? If not, you have no way to catch errors — and errors will come.
3. Does something leave the building based on that output, or is it just displayed? If it leaves, a human with a button has to stand in between.

What this means for us

AI made our own development dramatically faster — Materia took four months instead of a year. That is a real benefit and we are not going to pretend otherwise.

Inside the product, though, it does exactly one thing: it reads non-standard PDF quotes. The planner, the status derivation, the reply matching, the risk flags — all rules, because on orders worth hundreds of thousands you want to know why the system suggested precisely that.

Duller on a website. Better in daily use.