The box of delivery notes

At Tomás's building-materials company, on the outskirts of Jaén in southern Spain, there was a box. A cardboard box, the kind fruit comes in, on a table at the back of the office. Into the box went the delivery notes. Delivery notes are those slips that go with each delivery: what was supplied, to whom, how much, when. They came from two places. The suppliers' ones, when they brought material to the company. And the company's own, the ones clients signed when they received an order and the driver brought back in the van, sometimes crumpled, sometimes with a coffee stain, almost always handwritten or printed on those thermal printers that fade in the sun.

Every afternoon, Merche took the box and emptied it. Note by note, she entered into the computer what each one said: the supplier or the client, the date, the products, the quantities, the prices. To cross-check them against the orders, to balance the warehouse, so the invoice matched what had actually been delivered. Fifty, sixty notes a day. It took Merche the whole afternoon, and part of the morning when the box came in full from the weekend.

It was quiet work, the kind you don't see until it isn't done. Nobody at the company thought about the delivery notes, just as nobody thinks about water until it's cut off. But if Merche was away for two days, the box overflowed, the warehouse stopped balancing, and the invoices went out late. The whole company depended, without knowing it, on someone moving paper into the computer every afternoon.

Paper isn't the problem; typing it is

Tomás had been told to ditch the paper, switch to digital delivery notes, make suppliers and clients sign on a tablet. It sounds good in a talk. On the Jaén industrial estate, it doesn't. His suppliers were who they were, his clients had always signed on paper, and the driver wasn't going to wrestle a tablet in the rain while unloading sacks of cement. Paper, in his world, wasn't going anywhere. And forcing it would have been fighting reality to solve a problem that wasn't the paper.

Because the problem wasn't the paper. Paper is fine: it's fast, it doesn't crash, everyone understands it, it needs no battery. The problem was Merche's time spent moving it across by hand. That step, taking what a piece of paper says and writing it into a computer, was the one that cost hours, produced typos, and depended on a single person.

A machine that reads paper

What's done well today, and wasn't before, is exactly that: teaching a machine to read a piece of paper and pull out what it says. Not to digitise it, which is just taking a photo and saving it — that already existed and wasn't much use, because a photo of a delivery note still has to be typed in. But to understand it: look at the photo of the note and pull out, cleanly, the supplier, the date, each product, each quantity, each price, ready to go into the system.

We built something simple to use. The driver, on getting back, instead of throwing the notes into the box, took a photo of them with their phone, right there, in the van. The photo arrived in the system. The system read the note, pulled out the data, and cross-checked it against the matching order on its own. If everything matched — the note said the same as the order — it passed it without anyone touching anything. If something didn't match — a different quantity, a product not on the order, an odd price — it set it aside and showed it to Merche to look at.

The difference from before is enormous. Before, Merche typed in a hundred per cent of the notes. Now, the machine processed on its own the ones that were clear, which were most of them, and Merche only looked at the ones with something odd, which were the interesting ones, the ones that really needed an eye and a head.

The honesty of sometimes reading it wrong

Here we have to tell the truth, because promising that a machine reads a hundred per cent of delivery notes without fail is a lie, and lies in this come at a high price. A note handwritten in bad handwriting, or half-faded, or with a stain over the number, the machine sometimes reads wrong, just as a person would read it wrong.

That's why the important thing wasn't that it read. It was that it knew when it wasn't sure. The system we built, when a piece of data wasn't clear, didn't make it up: it marked it in yellow and passed it to Merche. "I've read this note, but I'm not sure about this quantity, please confirm it." Merche looked at the photo, confirmed in two seconds, and done. A machine that doubts and asks is worth a thousand times more than one that's nearly always right and, the few times it fails, does so in silence and puts a wrong number into the warehouse that nobody discovers until material runs short.

What changed in the back office

Two months in, the cardboard box was still on the table at the back, but nearly empty. It stayed for the four odd notes that came in on paper and that someone wanted to keep out of habit. Merche had gone from emptying the box every afternoon to reviewing, in half an hour, the notes the machine flagged for her. The warehouse balanced daily, not weekly. And the invoices went out with what had actually been delivered, without waiting for someone to finish typing.

Tomás said something to me that made me smile, because it was true.

—I thought this artificial intelligence stuff was for companies in white coats. And it turns out what's sorted out my life is that the computer reads the badly handwritten delivery notes my drivers bring in covered in dust.

I told him it's almost always like that. That the technology that most changes a company's day-to-day isn't the spectacular kind, the kind in the presentations. It's the kind that gets into the quiet work, into the invisible stretch someone spends every afternoon moving paper into a computer, and removes it. Not the whole warehouse. Not the van. Just the typing time. But that time, multiplied across every afternoon of the year, was a whole person devoted to copying what a piece of paper already said.

Because in many companies, the biggest saving isn't in doing new things. It's in no longer doing by hand what can already be read on its own.