An AI chatbot that answers from your real content, works on your site, on WhatsApp and on Instagram, and hands the conversation to a person the moment it needs to. Declared as AI, as the European regulation requires.
We work with
And how it differs from the button bot we have all suffered.
The classic chatbot walks a tree: press 1 for orders, press 2 for returns. It works as long as the question fits a branch. The moment it does not, the customer gets stuck and goes looking for a phone number.
An AI one understands the question as the person writes it, typos and half sentences included. The customer no longer has to guess which box we filed their problem under.
What matters is where its answers come from. We build it against your documents, and the model writes with those in front of it. That is what stops it making things up, which is the real risk with this technology.
A customer of a chandlery, on a Tuesday at eleven at night.
Shop chat · 23:04
You are talking to an assistant that runs on artificial intelligence. If you would rather speak to a person, say so and I will pass you over.
hi, do you have something like the buoyancy aid i saw the other day but childrens size and ok for kayaking
We do. For kayaking in children's sizes we stock two certified aids, 30 to 40 kg and 40 to 50 kg. Both have a crotch strap, which is what the standard requires for children.Source: product page
which one would you recommend for a 9 year old
That depends on weight rather than age, and I would rather not guess. I can pass you to someone first thing tomorrow, or if you tell me the weight I will tell you which of the two fits.Not found in the documentation
→ Case opened for the team, with the whole conversation attached
Three things make that exchange work: the notice at the top, the source under each fact, and a last answer that admits it does not know.
Six concrete jobs. The sixth is the one that makes the difference.
Opening times, address, payment methods, delivery windows, what the warranty covers. These are the questions that eat half a working day a month and almost always arrive when the office is closed.
The bot answers from your real documents, not from what the model thinks it remembers. Change a condition on your site and the answer changes with it.
Wired into your catalogue, the bot searches by what the customer describes, not by the exact product name. "Something similar but in blue and for outdoor use" is a perfectly normal query. Where live stock exists, it checks before promising anything. And when there is none, it says so.
The customer books without leaving the chat. The bot shows the real openings in the calendar, confirms, writes it down and sends the reminder. If it has to be cancelled, the slot frees itself.
It is the same engine as the booking automation, with a conversation in front of it.
Not every enquiry is worth the same. The bot asks the three or four questions your sales team always asks, and delivers the contact to the CRM already sorted, with what the customer said and a tag for what they are after.
The team walks into the call already knowing what it is about.
"Where is my order?" is, in almost every shop, the most repeated question of the month. The bot answers it from the order number and the carrier, without anyone opening the shipping dashboard. It also takes in problems and asks for a photo where that helps, then leaves the case open for a person to close.
This is the part that decides whether a chatbot helps or annoys.
The bot steps aside when the customer asks, when it detects anger or a complaint, and when it finds no answer in its documentation.
On handover it passes the whole conversation across, so the customer does not explain the problem twice.
The route a question takes from being typed to coming back answered.
When step two comes back empty, the bot does not fill the gap itself. It says so and hands over. That one rule is what separates a chatbot that helps from one that invents delivery terms nobody ever wrote.
The button one is not always worse. It depends on what your customers ask.
| Criterion | Button bot | AI bot |
|---|---|---|
| Expected questions | Handles them well and fast | Handles them just as well |
| Unexpected questions | Gets stuck | Answers if it is in the documentation |
| Maintenance | The tree has to be redrawn | The document gets updated |
| Risk of a false answer | None | Real, which is why it is fenced in |
| Cost per conversation | Effectively zero | Small, but not zero |
| Fits when | You get five questions and always the same ones | You have a catalogue, terms or varied cases |
In plenty of businesses the best result comes from mixing the two: buttons for the usual, AI for the rest.
Five phases. The first one decides whether the bot is any use.
We read your last few months of emails and messages. What people actually ask is rarely what you think.
Catalogue, terms and answers gathered into documents the bot can consult. Anything missing gets written.
Sign-offTone, languages, forbidden topics, handover rules and the AI disclosure. Wired into your CRM and calendar.
We break it on purpose: awkward questions, complaints, things out of scope. It gets tuned until it fails gracefully.
Sign-offIt goes live with conversations in plain view. Whatever it cannot answer becomes new documentation.
It comes from the EU Artificial Intelligence Act, so it is not optional.
If your business answers through an AI system, the person has a right to know. In practice one line at the start of the conversation covers it, but it has to be there and it has to be understandable.
Every chatbot we build declares it. We also leave a route to a person, information about what happens to the conversation data, and a log of what the bot answered so it can be reviewed.
There is more detail on the blog: the notice an AI chatbot has to carry.
We do not publish rates. This is what moves the figure.
A button bot only knows how to walk the tree someone drew. If the question does not fit, it gets stuck. An AI one understands the question as the customer actually writes it and answers from the material we gave it: your catalogue, your terms, your opening hours.
That is the real risk of building an AI chatbot carelessly. It is reduced three ways: the bot answers only from your documents, not from what the model thinks it knows; some topics are off limits, such as firm prices or legal deadlines; and when it finds no answer it hands over to a person instead of improvising.
Yes. The EU Artificial Intelligence Act requires people to know they are talking to an AI system. Every chatbot we build says so at the start of the conversation.
It hands over. The bot alerts your team on whichever channel you use, passes the whole conversation so nobody has to repeat themselves, and steps aside. It also hands over on its own when it detects anger, a complaint or a topic beyond its scope.
It is handled like any customer data: legal basis, notice to the user and a retention period. The accounts are in your name and the conversations are yours. Where the sector requires it, we can work with models hosted in the EU or on your own server.
Seeing the questions your business gets is enough for us to tell you whether an AI chatbot is worth it, or whether something simpler will do.
Talk to usWe also build automations, AI agents and websites.