Elena runs a logistics company in Murcia, south-eastern Spain. Thirty-eight employees, one warehouse in town, another in Cartagena, and a problem she'd been trying to solve for eighteen months: every morning, her customer-service team spent the first hour reading overnight emails, deciding whether each one was a complaint, a tracking enquiry or a new order, and forwarding it to the right department. Three people. One hour. Every single day.
A technology vendor proposed installing a chatbot. Elena asked what for. The vendor explained that the chatbot would read the emails, understand what they were asking and reply automatically.
—And if it doesn't know what to say?
—Then it escalates to a person.
—And how does it know it doesn't know?
—Well, that gets fine-tuned.
Elena wasn't naive. She knew that "that gets fine-tuned" meant they didn't know either. But she had a real problem and half a million emails a year, so she said yes.
The first month
The chatbot handled tracking enquiries well. It had access to the tracking system, cross-referenced the order number with the database and told the customer where their parcel was. Fast, accurate, no errors. Customers didn't complain. Some didn't even realise they weren't talking to a person.
Complaints were different. A woman from Lorca wrote because her order had arrived damaged. The chatbot replied that her shipment had been successfully delivered the day before. The woman replied that yes, it had arrived, but broken. The chatbot replied that it was glad she had received her order.
Elena saw that email on a Friday afternoon. She called the woman herself. It took twenty minutes to calm her down.
New orders were yet another problem. Elena's big clients didn't order through a form: they ordered by email, free text. "I need two hundred boxes of model KR-40 by Monday" mixed with "by the way, you overcharged me on invoice 2847 last month." One email, two different matters, each for a different department. The chatbot couldn't separate them. It treated the email as one thing and replied to whichever subject seemed clearest, which was usually the order. The complaint got lost.
Three problems, three solutions
I told Elena something that isn't easy to hear when you've already invested in a solution: what she had wasn't wrong; it was wrongly applied. The chatbot worked well for what it worked well for. For the rest, something else was needed.
Elena had three problems. Not one.
The first was answering tracking enquiries. For that, the chatbot was fine. A language model connected to the tracking system: question, lookup, answer. Clean.
The second was classifying incoming emails. Separating complaints from orders from enquiries. For that, she didn't need a chatbot that conversed. She needed a classifier. A smaller, faster, cheaper system that reads each email, identifies what kind it is and sends it to the right department. Without replying to anything. Without making anything up. Just classify. It's artificial intelligence, yes, but it doesn't need to generate text or hold a conversation. It needs to read, decide and route.
The third was handling complaints. And for that, for now, she needed a person. Not because the technology couldn't do it, but because Elena's customers expected empathy, not efficiency. A complaint well handled by a person saves a customer. A complaint handled by a machine that says "glad you received your order" loses one.
—So I need three different things?
—You need two systems and a team that does what it's always done, but without wasting the first hour of the day reading emails that a machine can sort on its own.
The economics of using just enough
There's something nobody mentions at innovation talks: large language models are expensive. Not just to buy; to run. Every time you ask a large model to process an email, you pay for the computing power it needs to understand the text and generate a reply. If you have five hundred emails a day, it adds up.
A classifier doesn't need that power. It doesn't generate text. It reads, decides and routes. The cost per email is a fraction. For Elena, the difference between running all five hundred daily emails through the chatbot or running them through a classifier and sending only the tracking queries to the chatbot was a seventy per cent saving in processing costs.
—Wait. Are you telling me I save money by using less artificial intelligence?
—I'm telling you that you save money by using the right amount of artificial intelligence for each thing.
Elena leaned back in her chair and stared at the ceiling. She had that look people get when something clicks and they wonder why nobody had explained it before.
What worked
We built the two pieces. The classifier at the front: it reads each email, decides whether it's a tracking enquiry, a new order or a complaint, and sends it where it belongs. Tracking queries go to the chatbot, which replies automatically. Orders go to the sales team. Complaints go to customer service, which now starts the day with a clean inbox and only the emails that need a human.
The three people who used to spend the first hour reading and forwarding emails now spend that hour handling complaints. Nobody was let go. An hour of work that added no value was replaced by an hour of work that did.
A month later, Elena sent me a message. It said: "The woman from Lorca ordered again. Sandra replied. By hand."
She didn't need to say anything more.
The toolbox
Artificial intelligence isn't a product. It's a toolbox. Inside there are screwdrivers, wrenches, hammers and saws. They all work. None of them works for everything. A language model is a power saw: impressive, powerful, expensive. But if what you need is to tighten a screw, what you need is a screwdriver.
Sometimes the answer is a chatbot. Sometimes it's a classifier. Sometimes it's a business rule that fits in three lines of code. Sometimes it's all three chained together, each doing what it does best. What you always need is someone who understands the problem before opening the toolbox.
Are you using the right tool for your problem?
Sometimes what holds a company back isn't the technology — it's having chosen it without understanding the problem first. We can look at it together.
Let's talk