Marga runs an electrical installations company in Elche, on Spain's south-eastern coast. Twenty-two employees, twelve vans, and a process that repeated every day and that nobody wanted to do.
The process went like this. Every morning, between fifteen and thirty job sheets came in. The electricians filled them out by hand, in the van, after finishing each job: what they'd done, what materials they'd used, how many hours. In the afternoon, someone in the office — Rebeca — picked up those job sheets, one by one, and did three things with each. First, she entered the data into the invoicing system. Second, she deducted the materials used from the warehouse stock. Third, if the sheet mentioned anything pending — "missing a box of fittings", "client wants a quote for the second bathroom" — she opened a task for someone to handle it.
Three steps. Thirty sheets. Every single day. It took Rebeca the whole afternoon. And when Rebeca was on holiday, or off sick, the sheets piled up in a tray, and the following Monday someone had to do a hundred and twenty in one go.
Marga called me because she'd heard about artificial intelligence and wanted "a chatbot or something like that". I asked what for. She explained the job sheets. I told her that what she needed wasn't a chatbot.
—It isn't? But everyone's talking about chatbots.
—A chatbot answers questions. You don't need anyone to answer questions. You need someone to do a job.
The difference nobody explains well
There's a confusion that comes up again and again, and I understand it, because the words are used loosely everywhere. A chatbot is a thing that talks. You ask it something and it answers. It's a conversation. It's good for helping a customer, for clearing up a doubt, for guiding someone through a menu. But it ends where the sentence ends. When it stops talking, nothing has been done.
An agent is a different thing. An agent doesn't converse: it executes. You give it a job with several steps, explain what to do at each one, and give it access to the tools it needs — the invoicing system, the warehouse, the task manager — and it does it. On its own. From start to finish. It doesn't hand you back a nicely worded reply: it hands you back the finished job.
I explained it to Marga with an image she grasped at once.
—A chatbot is like a colleague you ask where the stapler is, and they tell you. An agent is like Rebeca: she takes the job sheet, reads it, invoices, deducts the stock, and opens the task. The difference is that one knows, and the other does.
—Then I want Rebeca.
—You want someone to do what Rebeca does with the job sheets. Rebeca, meanwhile, is going to do things a program can't.
What an agent is, with no mystery
A custom agent isn't a product you buy off a shelf. It's something built for one specific company, because it does that specific company's work. Marga's had to understand Marga's job sheets, know her materials catalogue, connect to her invoicing system and her warehouse. Another company's would be different, because the work would be different.
Underneath, it's no more than this: a language model — one of those that understands text written by people — given three things. One, clear instructions on what to do, step by step. Two, access to the company's tools, just as you give a new employee a username and password for the programs. And three, rules about when to stop and tell a person. That third one is the important one. A good agent isn't the one that does everything alone; it's the one that knows what it shouldn't do alone.
We gave Marga's a simple rule: if a job sheet was clear, it processed the whole thing without asking. If something was odd — a material not in the catalogue, a strange quantity, a note it didn't understand — it didn't make things up. It set the sheet aside and sent Rebeca a message: "Three sheets to review." Rebeca looked at them in five minutes. The other twenty-seven were already done.
The first month
The first month was one of distrust, as it should be. Rebeca reviewed everything the agent did, sheet by sheet, the way you check a new intern. In the first two weeks she found things. A material the agent deducted wrongly because two reference codes looked alike. A client whose name was written three different ways in the system. We fixed it. Every time Rebeca corrected something, the agent learned it as a rule, and never got it wrong again.
By the third week, Rebeca stopped reviewing everything. She started reviewing only what the agent set aside, which was the only doubtful part. The afternoon opened up. And here's what Marga didn't expect: Rebeca wasn't left without work. Marga gave her the work that had gone undone for two years because nobody had time. Calling clients from six months ago to offer them the annual inspection. Negotiating better prices with two suppliers. Actually tidying up the warehouse.
—I have the same person —Marga told me— but now she does what makes money, not what wears you out.
What the agent didn't do
It's worth saying what Marga's agent didn't do too, because selling smoke is easy and doesn't last.
It made no business decisions. It didn't set prices, didn't approve quotes, didn't talk to clients. It did a three-step administrative job, repetitive and well defined, that a person used to do with their mind elsewhere because it was work that required no thought, only consistency.
Nor was it built in a day. It took five weeks: understanding the process properly, connecting the tools, writing the rules, testing with real job sheets, correcting. Most of the work wasn't the artificial intelligence. It was understanding exactly how Rebeca worked, step by step, so we could teach it to a machine. That's almost always the hard part, and the part skipped by anyone who promises magic agents in an afternoon.
What really changed
Six months later, Marga told me a small thing that sums it up well. Over Easter week, Rebeca took a week's holiday. For the first time in five years, the job sheets didn't pile up. The agent kept doing them every afternoon, setting the doubtful ones aside in a folder. When Rebeca came back on Monday, instead of a hundred and twenty sheets to catch up on, she found nine to review.
—It's like having one more employee —Marga told me—. One that doesn't get sick, doesn't go on holiday, and doesn't get bored doing the same thing over and over. Even if it isn't on the payroll.
I told her that was exactly the idea. That artificial intelligence used well isn't the kind that replaces people. It's the kind that takes the work that bores people, so people can do the work only they can do. Marga's agent will never call a client to sell them an inspection. Rebeca does that, now that she has her afternoons back. And she's good at it, because it's one of those things a machine can't do and a person can.
Chatbot or agent: how to tell which you need
If what you want is for someone to answer — a customer, a doubt, a frequent question — what you need is a chatbot. It begins and ends in the conversation.
If what you want is for someone to do a multi-step job that today eats up a person's hours — read something, enter it into a system, move data from one place to another, open tasks — what you need is an agent. And a custom agent isn't bought ready-made: it's built by first understanding, calmly, how the person who does that job today actually does it.
Is there a multi-step job that repeats every day in your company?
If there's an administrative process someone does over and over, it can almost always become an agent that does it alone and flags whatever doesn't fit. We can look at it together.
Let's talk