Óscar called me on a Tuesday at nine in the morning. The first thing he said was that he wanted artificial intelligence for his company. The second, that he'd read an article in a business newspaper. The third, that his brother-in-law already had it.
—And what does your brother-in-law do with it?
—He asks it things.
—What things?
—Business things.
—And does it answer well?
—He says so.
Óscar runs a stationery distribution company in Albacete, south-eastern Spain. Forty-two employees, eight delivery vans, a warehouse that runs on an ERP from 2015 and an office manager, Loli, who handles invoicing with a combination of that ERP, three Excel spreadsheets and a notebook where she writes down everything the other two systems don't capture.
What Óscar wanted was to install ChatGPT. That's how he put it. ChatGPT. The way someone says they want to install air conditioning. He'd seen it at a talk at the business park, a supplier had shown it to him on his phone and it had struck him as magic.
—You ask it something and it answers. Like a person.
—Right. And what are you going to ask it?
—Whatever I need to know.
I asked him what he needed to know. He went quiet for a moment. Then he said he needed to know which customers owed him money, which ones had stopped buying, which product sold best in which area and whether the margin on the Cuenca delivery route was worth keeping. Things like that.
What Óscar didn't know
I explained something that isn't easy to explain without sounding like you're correcting someone. Because nobody likes being corrected, least of all a man who's been running a business for twenty-two years.
ChatGPT is a language model. It understands text, generates text, converses. It's brilliant at that. But Óscar's company data isn't text. It lives in the ERP, in Loli's spreadsheets, in her notebook. ChatGPT doesn't know that the customer in Hellín is forty-seven days overdue. It can't know, because nobody has told it. And even if you told it, what Óscar needs isn't a conversation: it's a table with debtors sorted by days overdue and a red flag when they pass sixty.
That isn't a language model. That's a database query, a business rule and a dashboard to look at it. Artificial intelligence, yes. But the other kind. The kind that doesn't make the papers.
—So ChatGPT won't work for me?
—For what you need right now, no.
—And my brother-in-law?
—Your brother-in-law uses ChatGPT to draft emails. Which is perfectly fine. But it's not the same thing.
Óscar stared at his phone the way you stare at a lottery ticket that turns out to be from the wrong draw.
Loli and the red notebook
That week I went to Albacete. I sat with Loli at her desk, between the printer and a plastic plant that had been there since the company moved to the business park. Loli showed me her system. It had logic. It made sense. What it didn't have was a connection to anything else.
—And what's this here? —I asked, pointing at a column in the notebook.
—Those are the partial returns the ERP doesn't capture.
—Why doesn't it capture them?
—Because when they built the ERP, partial returns didn't exist. Then we started doing them and nobody updated it.
—Since when?
—Since eighteen.
Eight years of partial returns written down by hand in a red notebook. Data that existed but no system could see. When Óscar asked ChatGPT about the margin on the Cuenca route, ChatGPT wouldn't know that half the orders from Cuenca came back partially. It would make up a convincing answer. That's what language models do when they don't have the data: they make things up. Politely and with good grammar, but they make things up.
The right tool
What we did was something else. We connected the ERP to Loli's spreadsheets. We pulled the returns out of the red notebook and fed them into the system. We built a dashboard — not a chatbot, not a language model — that showed on one screen what Óscar needed to see: debtors, products by area, margin by route, customers who had stopped buying.
It wasn't spectacular. It didn't converse. It didn't generate text. It wouldn't have impressed anyone at a business-park talk. But within two weeks, Óscar discovered that the Cuenca route, the one he thought wasn't paying off, actually paid off well once you subtracted the returns that had been misaccounted. And that the customer in Hellín wasn't a debtor: he had a pending credit note that Loli had written in her notebook but nobody had entered into the ERP.
—So he didn't owe us anything.
—Thirty-two euros. The rest was our mistake.
Óscar called the man from Hellín that afternoon. He hadn't ordered in two months. The man said he'd switched to another supplier because he thought they'd stopped extending him credit.
Artificial intelligence — the real kind
People think about artificial intelligence the way they think about electricity: they assume it's one thing when it's actually many. A plug, a lightbulb, a transformer, a bolt of lightning — all electricity, but not the same thing. ChatGPT is a bolt of lightning. Spectacular, powerful, impressive. But if what you need is a lightbulb that switches on every morning at seven, the lightning won't do.
There's AI that classifies documents without mistakes. AI that spots patterns in sales data. AI that predicts which customer is about to leave before they leave. None of these need to talk to you. None of them need a language model with billions of parameters. They're smaller solutions, cheaper, more reliable and — above all — better suited to the problem.
The trick — if there is a trick — is not to fall in love with the tool before you understand the problem.
What matters, no fluff
A language model understands text and generates text. If your problem is data scattered across systems that don't talk to each other, what you need isn't a conversation with a machine: it's a dashboard that connects that data and puts it in front of you.
Artificial intelligence isn't just ChatGPT. It's also the query, the classification, the pattern no human would spot in a spreadsheet. And almost always, the right solution is simpler and cheaper than you've been told.
Three months later I went back to Albacete for something else. Loli was still at her desk, between the printer and the plastic plant. But the red notebook was gone. I asked her about it.
—It's in the drawer —she said—. Just in case.
She got up, went to the coffee machine, came back with two cups and set one in front of me.
—You know the best part? —she said, sitting down—. Óscar doesn't ask me any more how much the man from Hellín owes. He looks it up himself.
She stirred her coffee with a plastic spoon. Took a sip. Stared at the computer screen where, for the first time in twenty-two years, all the numbers were in the same place.
—Now that's intelligence —she said. And I'm not sure she meant the artificial kind.
Does your company need AI — or does it need to see its own data?
Sometimes the answer isn't what makes the headlines. A conversation about your business can clarify which tool actually fits.
Let's talk