My first chatbot

I built my first chatbot convinced it was an afternoon's work. The technical part almost was; what came afterwards, when real people started writing to it, was not. Here's how I did it, what went wrong and, above all, what it says when it doesn't know something.

You've surely seen them: "build your AI chatbot in 10 minutes and serve your customers 24/7". The presenter uploads a PDF, types a question, the bot answers beautifully and the video ends. Nobody shows the next day.

My first usable version took me about 14 hours spread over two weekends, and "usable" is generous. The job came from a client with a small irrigation supplies shop that sells online and by phone. His problem was very specific: he got about 30 enquiries a day and reckoned more than half were the same ones: delivery times, how to return a product, whether a given part fits a given model, and opening hours. He spent half the morning answering the same things.

The pretty demo exists, I won't deny it. The trouble starts the day a customer who hasn't seen the video writes in.

A chatbot is a program that automatically converses with whoever writes to it, at any hour. So far, easy. What confuses people is that two very different creatures live under that name.

The classic ones work with buttons and fixed answers: "press 1 for shipping, 2 for returns". They're predictable, which is their great virtue, but as soon as someone asks something off-script they get lost and send you back to the menu like an answering machine.

The ones with a language model behind them (the "brain" that understands and writes, the same kind of program behind ChatGPT or Claude) understand the question the way people actually type it. That's their advantage. Their hidden cost is that they can make up an answer with a confidence many salespeople would envy. That's why building a chatbot with artificial intelligence isn't just plugging in a model and calling it done: you have to decide where it gets what it says.

An example with the same question: "Will my order arrive before Friday if I place it today?". The button bot replies: "Choose an option: 1. Shipping, 2. Returns". The AI one replies: "Deliveries take between 24 and 48 hours, so if you order today it should reach you before Friday". Much better... as long as that really is the delivery time.

I started where nobody wants to: the list of frequently asked questions. I asked the client to forward me a month's worth of emails and messages, and out came 42 different questions, which I grouped into 18. That alone took me about 5 hours, and it was the most useful part of the whole project.

Then I chose the tool. I used a no-code chatbot platform that lets you upload documents and pick the model, with Claude underneath, because I work mostly with Claude, I know how it behaves with long instructions and I wanted it to respect the rules I was going to give it. For the channel, I started with a chat bubble on the shop's website, which is the simplest option and the one with the fewest rules.

There were three knowledge sources: the FAQ document, the shipping and returns conditions, and a catalogue table with part compatibilities. That table caused the most trouble: it lived in a spreadsheet the client kept by hand, full of abbreviations only he understood.

And here's the lesson no video tells you: what took the longest wasn't the technical part, it was preparing the content the bot would use. The rough breakdown was:

Almost everyone asks me the same thing: "can't this go on WhatsApp?". It can, but you'd better know what you're getting into before you put a chatbot on WhatsApp.

The WhatsApp Business app, the one you download on your phone, only allows simple automatic replies: the greeting message, the away message and quick replies. A real chatbot needs Meta's official platform, the WhatsApp Business API (an API is the door through which one program talks to another), and you usually get in through a provider that gives you a panel to set it up.

What catches people out are the rules. Since 15 January 2026 Meta doesn't allow general-purpose AI assistants on the WhatsApp Business API, that is, bots that chat about any topic ChatGPT-style. It does allow the ones that serve a business's customers: bookings, orders, support (checked in September 2026). A bot like the shop's, which only talks about shipping and parts, falls within what's allowed.

And a warning: there are unofficial tools promising to hook anything up to your personal WhatsApp. They can end with your number blocked, and that number is usually the one all your customers know.

During the first week I reviewed every conversation. I'll summarise three, changing the details so nobody recognises themselves.

The good one: a customer asked at eleven at night whether a solenoid valve from the shop worked with his controller from a particular brand. The bot looked it up in the table, said yes and sent him the link. He bought it that same night. First sale handled by a machine. I admit I took a screenshot.

The funny one: someone wrote "hi, anyone there?" and the bot, ever diligent, replied with the full opening hours, the returns policy and delivery times to the Canary Islands. The customer answered "ok, thanks" and left. Technically correct; humanly, a wall of text.

The one that could have been expensive: a customer asked whether shipping was free above a certain amount. The bot said yes, above €50, with total confidence. That promotion had existed, but it had ended two months earlier and was still in an old document that slipped into the sources. I spotted it reviewing the conversation that same afternoon and the client decided to honour the price. If nobody had checked, it would have been a written commitment to a customer, made by a program nobody had authorised to promise anything.

I laughed at the second one. Not so much at the third.

This is the heart of the matter, and what decides whether a customer service chatbot helps or gets in the way.

Language models have a dangerous habit: when they don't know something, they sometimes make it up with complete ease. That's called hallucinating, and in front of a customer it gets expensive.

The first step was limiting the sources: the bot can only answer from the business's documents, and I checked that none of them were out of date. The second, a clear instruction. The one I use now says roughly this: "Answer only with the information in the attached documents. If the answer isn't there, don't infer it or fill it in: say you don't have that information and offer to pass the query to a person. Never promise prices, discounts or delivery times that aren't written down".

The handover to a person works like this: the bot asks for an email or phone number and leaves the query in an inbox the client checks twice a day; the commitment is to reply within 24 working hours. It isn't instant, but it's honest.

And one thing that isn't up for negotiation: the customer has to know they're talking to a machine. The first message says so clearly. Deceiving someone never ends well, and European regulation is heading that way.

Before and after, with the free shipping question:

• Before: "Yes! Shipping is free for orders over €50".
• After: "The shipping conditions I have don't include any free shipping promotion. If you like, I can pass your query to a person to confirm it".

Less charming, a lot more reliable.

Only what shows up on invoices, as of September 2026: the platform, €29 a month; model usage, between €6 and €11 depending on the month; the web channel, no extra cost; and about 2 hours a month of my time for review and maintenance. The model cost depends on the number of conversations: right now around 400 a month. If they became 4,000, the numbers would change, and it's worth knowing that before celebrating.

Three things I'd change: review every document and remove the old ones before uploading anything; write the "I don't know" instruction from day one instead of adding it after the scare; and shorten the answers, because nobody wants to read a contract to find out opening hours.

Who is it worth it for? Businesses with lots of repeated questions and the information already written down. Who isn't it for? Anyone who gets sensitive or all-different enquiries: there a bot saves little and can annoy a lot.

Before building anything, do the paper test: for a week or two, write down the questions people ask you. If most of them repeat, there's a chatbot in there. If not, maybe a good FAQ page is enough, and that never hallucinates.