An engineer is standing at a packaging machine that has stopped. Year of manufacture unknown, make one of the nine the company services. He needs a wiring diagram.

What happens then: a call to the office. The office searches. Which type was it again, which year, and was that modification in it or not. Six to seven minutes, every single time.

That is not the repair. That is the hunt for the drawing of the repair.

Why it takes so long

There is no shortage of documentation. The manuals are all there, digitally.

The problem is that each supplier organises them differently. One by serial number, another by year-of-manufacture ranges. There are three thousand customers with machines from 1995 to now. Of one old type there are only three hundred left in the country. One engineer knows that.

On top of that, customers swap motors and modify machines after delivery without it being recorded anywhere. What is actually there, you only know when you are standing next to it.

The clock that is really ticking

The managing director said it himself and it stuck: the real machine knowledge sits in the heads of the engineers, and they are ageing.

That stops being an HR question the moment the knowledge sits nowhere but in a head. Every engineer who retires takes a piece of the service with him. You can hire a new engineer. You cannot hire twenty years of feel for machines.

What the company worked out itself: an engineer costs five thousand euros a month and brings in a thousand euros a day. Every fault that is fixed faster is money straight away. But so is every engineer who leaves without his knowledge being recorded — you only notice that one a year later.

What we are building

All the manuals in one searchable store. On top of that a web app where the engineer gets answers by chat, per machine, on type number and year of manufacture.

One thing is not negotiable there: every answer comes with a clickable reference to the original page in the PDF. If the diagram is on page 44, then page 44 opens. An engineer who has to rely on an AI summary at a machine that is live will use it once.

Beyond that: we process PDFs in software, because that is cheap. We only bring in AI where it genuinely adds something. That keeps the monthly cost down, and those monthly costs decide whether something like this is still running in two years.

The order

Internal first, with their own engineers. Prove that the answer is correct, on real faults, with people who spot it immediately when it is not.

Only after that a QR code on the machine and a customer portal. And if that works, a weekend assistant — because customers are asking for support outside office hours and there is none at the moment.

But in that order. Whoever starts with the customer portal is building a shop window for a shop with no stock in it yet.