The summer they overstocked

Domingo makes and sells artisan ice cream in a town on the coast of Almería, in south-eastern Spain. Not the beach cones, but tubs he sells to hotels, restaurants and a handful of local shops. He's been at it for eighteen years. And every spring, around April, he made the most important decision of the year: how much to produce for the summer.

He made it the way he always had. He looked at what he'd sold the previous summer, added a bit because each year was a little bigger, took a bit off if the year had been strange, and produced that amount. Gut, in short. Eighteen years of gut.

The gut got it more or less right. But the "more or less" cost him money on both sides. The summers he came up short, in August he had no product to serve the hotels, lost sales and, worse, looked bad to clients who remembered it afterwards. The summers he overdid it, September caught him with the cold rooms full of tubs to dump at a loss or throw away, because artisan ice cream doesn't keep from one year to the next.

The summer he told me about was one of the second kind. He produced expecting a scorcher and August came weak, with two weeks of easterly wind and bad weather. He was left, all told, with over a thousand units. A thousand tubs made, with their milk, their fruit, their cold-room electricity and their work hours, that ended up nearly given away. Domingo worked out what it had cost him and lost his appetite.

What Domingo didn't know he had

Domingo thought he was making the decision blind. He wasn't. He was making it with a single figure — last year's — while many more sat to hand without his knowing.

Because Domingo had, in his invoicing system, eighteen years of sales. Not just how much he sold each summer, but how much he sold each week, of each flavour, to each type of client. He had, without ever looking for it, the trail of how his business behaved: that the hotels spiked their orders the moment the season opened, that the restaurants ordered steadily all summer, that certain flavours rose in July and others in August, that a hot summer gave a surge and a cool one flattened it.

All of that was written down, invoice by invoice, for eighteen years. But it was written the way the things nobody reads are written: in columns, in thousands of rows, in a format built for making invoices, not for understanding the business. Domingo had eighteen years of answers and hadn't asked that data a single question.

From gut to question

He called me because a friend had told him "with artificial intelligence you can predict your sales". I told him to be careful with that phrase, because it promises to foretell the future, and nobody does that. What you can do is something else, less magical and far more useful: look properly at eighteen years of behaviour and produce a reasoned forecast, with its margins, instead of a figure pulled from the gut.

—And how is that different from what I do?
—You use one figure and remember three summers. The system uses all the data and forgets none of it. You decide with what you remember. This decides with what happened.

We took the eighteen years of invoices and set them to tell what they knew. Not to foretell, but to answer specific questions: how much each type of client tends to order each week of summer, which flavours move when, how much things shift with the weather, how much margin of error each forecast carries. The difference between "make this" and "make this, and be ready for that if the first week of July comes on strong".

A forecast with honesty

The important thing about a good forecast isn't the number. It's the honesty about what isn't known. A charlatan gives you an exact number and doesn't blink. A serious forecast tells you: "the most likely is this, but here's the range, and it depends above all on these two things, which are worth watching in June".

For Domingo we built something simple to read. For each flavour and each type of client, a forecast of how much he'd need, week by week, with its band of "at least this, at most this". And we flagged the two or three signals that really moved the needle, the ones he had to watch early in the season to adjust: how the hotels' first orders came in, what the fortnight's weather forecast said, how tourism in the area was going.

Domingo stopped producing a fixed amount in April for the whole summer. He started producing in batches, with a forecast in front of him, adjusting as things unfolded. He was still the one deciding. But now he decided looking at eighteen years, not three summers.

What changed, in tubs

The first summer with a forecast wasn't perfect, because no summer is. But by September Domingo was left with fewer than two hundred units instead of over a thousand. And he didn't come up short with any hotel in August, because the forecast had warned him to fine-tune the flavour that had run out the previous year.

He told me over a coffee, by then in autumn, and said something that struck me as fair.

—I thought experience was knowing how much to make. And it turns out experience was all stored on the computer, and I was deciding from memory when I could have decided with the whole of it.

I told him his experience still counted, and a lot, because the forecast told him what was likely, but the one who knew his clients and could smell when a summer was coming strange was him. We hadn't taken away his craft. We'd given him, at last, access to his own memory.

Because that's almost always what data analysis is in a small company. It isn't magic, nor is it foretelling the future. It's no longer deciding on one figure and three memories, when you have eighteen years of data waiting for someone to ask it a question.