Getting the business off the ground

The videos say an AI business can be set up in a weekend. It took me three weeks, I spent €149 and I had three scares, one of them involving data protection law. Here's the full diary, with what the machine did and what I had to do myself.

In the first part I went through what the internet promises and chose a business using four criteria: investment, time, competition and what I know how to do. The winner was a small, very specific service: building FAQ assistants for local shops, with a fixed price and monthly maintenance. In other words, a chat on the shop's website or phone that answers the usual questions (opening hours, prices, whether they have such-and-such, how to book) without the owner having to pick up the phone every ten minutes.

The goal: €1,000 net a month within six months of launch. The rule, the usual one: real numbers, however they turn out.

I started this phase with a €500 budget for the whole challenge and about 10 hours a week.

There's a whole genre of videos that starts the same way: "On Friday I had nothing. By Sunday my AI agency was already billing". In between, a template, a couple of requests to the assistant and lots of epic music.

I didn't entirely believe it, but I'm not immune either. My initial plan, jotted in a notebook with a naivety I now find touching: Saturday, the offer and the website; Sunday, the first test assistant; Monday, out visiting shops. Planned launch: 8 June.

Reality had other plans. I hit the street on 22 June, two weeks later than planned. And it wasn't laziness: a service that talks to someone else's customers, on someone else's behalf, has a lot more pieces than fit in a twelve-minute video.

Week 1 (1 to 7 June). I set out to nail down the offer, the name and the website. With Claude I defined what the service includes, what it doesn't and how much to charge: €150 for setup and €25 a month for maintenance, launch price. I bought the domain, set up a small server to host the assistants and, in a burst of optimism, a paid website template. 14 hours. Memorable detail: I disliked the template the moment I installed it and ended up building the website from scratch with Claude's help. In the end the business was called Mostrador Digital.

Week 2 (8 to 14 June). The prototype. A friend with a bakery agreed to be a pilot client in exchange for not paying anything. I asked him for the twenty questions he gets asked most and used them to build the knowledge base, which is the body of information the assistant can draw its answers from. I also made the logo (I cover that separately in My business's images, made with artificial intelligence). 13 hours, about 5 of them on the logo. Memorable detail: the first scare, which I describe below.

Week 3 (15 to 21 June). Website finished, legal texts, installation at the bakery and the service photos, about 4 hours of the week's total. Claude drafted the privacy policy and the terms of service, which I took to an accounting and legal advisory office for review. 13 hours. Memorable detail: on its first day, the bakery's customers asked things neither my friend nor I had anticipated: whether there was gluten-free bread, whether you could park out front and whether the dog could come in.

If I split the work into three piles, it looks like this.

What artificial intelligence got right first time: the website copy, the structure of the offer and the chat code that gets embedded in the shop's page. I barely touched any of it, and it saved me about 8 hours.

What helped but needed a lot of correcting: the bakery assistant's answers, the draft legal texts, the logo and the images. The model brought the volume and I brought the judgement. This pile took about 17 hours, and a good part of it was reviewing.

What only I could do: deciding the price, convincing my friend, sitting down with him to understand his business, setting up the server, paying, talking to the advisory office and testing everything as if I were a customer in a hurry. About 23 hours. No model has a coffee with the baker to understand why he gets asked so much about the empanadas.

This fits with what I said when discussing whether you can make money with artificial intelligence: the tool speeds up work, it doesn't replace it. It took hours of writing and coding off my hands, but all the decisions, the paperwork and dealing with people were still mine. And the hours spent reviewing what the machine does count too, even if they never show up in the videos.

A note that stings: I set myself ten hours a week and did between 13 and 14. I was already breaking my own rule.

I only include what I have an invoice or receipt for. The Claude subscription isn't included because I was already paying for it before the series; if this business ever had to pay for it alone, it would have to be added.

The API, for beginners, is the door through which a program uses the model directly, without going through the chat, and you pay for what you consume. It's what makes the shops' assistants work.

The expense that surprised me was the advisory office, not because of the amount but because of what they found (I explain it in the next section). It's by far the best-spent money.

The one I could have saved is the template: €39 for something I didn't use. And the image tool, which I used heavily for one week and didn't cancel in time, so it'll keep showing up in the part 4 accounts.

Three scares, told without make-up.

The first, the bakery opened on Sundays (11 June). Testing the assistant, I asked about weekend opening hours and it calmly replied that on Sunday they opened from nine to two. They don't. The model filled the gap with whatever seemed likely. The blame was shared: its for inventing and mine for not telling it that, if something isn't in the shop's information, it should say so and suggest calling. I fixed it with that instruction and a battery of 40 trick questions that I now run on every assistant before installing it. 3 hours.

The second, the ten euros that vanished in one night (13 June). I left the prototype public to show my friend and some automated program started firing questions at it non-stop. By morning the credit was at zero. One hundred percent my fault: I hadn't set any usage limit. Now every assistant has a cap on messages per hour and a monthly spending limit. 2 hours and €10 of lesson.

The third, phone numbers stored without warning (17 June). The bakery's customers sometimes left their phone number in the chat to place orders, and my system stored every conversation. The advisory office asked me where people were told about that and who could see it. Nowhere, and me. I had to add a visible privacy notice in the chat and automatic deletion of conversations after 30 days. The machine wasn't to blame here; it simply didn't occur to me to ask it, and it didn't warn me.

If there's a common lesson, it's that the big failures came from rushing. The machine only made one, and it was the most dangerous because it sounded right.

As of 21 June, ready to show: the Mostrador Digital website with the offer and prices, an assistant running at the pilot bakery, reviewed legal texts and a contact form. Half-done: a demo I could show on my phone without depending on the shop's connection.

Accumulated spend so far: €149. Accumulated hours: 40. Income: zero, because the pilot is free. Those three figures are the starting point for the accounts in part 4.

Having a service set up that nobody knows about is like opening a bar in a dead-end alley. In part 3 I tell how I got the word out: I tried in-person visits to local shops, flyers, some ads and social media, with €200 at most.

I'll confess the vertigo I felt: building I can do; what came next was walking into a hardware shop and explaining to the owner why he needs a chat. If they threw me out, at least I'd have an anecdote for part 4.