An assistant that only knows what you told it
The difference between a useful assistant and a dangerous one is where the answers come from. Ours knows nothing you did not give it, and when it does not know, it says so.
Quote Fixed, within two working days
Almost every question a small business receives is one of the same ten. What time do you open on Sunday. Do you do gluten-free. Is there parking. How much is it. Do you give a proper invoice. Do you come out as far as Badalona.
A well-built assistant answers those ten in any language at any hour and leaves you the conversations that are worth having. A badly built one invents a Sunday opening time and sends a customer to a locked door.
The difference is technical and it is called grounding: the assistant can only answer from the documents you gave it — your menu, your prices, your hours, your terms — and is forbidden from filling in what it cannot find. When it does not know, it hands over to a person. That rule is not a setting; it is the architecture.
What it includes
-
Answers only from your material
It is built on your documents. If the answer is not in them it says it does not know and offers a person. It does not fill gaps, which is exactly where the trouble starts.
-
Four languages
Catalan, Spanish, English and French, detecting the customer’s. For a business serving visitors, that alone pays for the installation.
-
Handover to a person
With the whole thread, so whoever picks it up does not have to ask the customer to start again. Over WhatsApp, email, or wherever you work.
-
What people ask, written down
A weekly summary of what customers genuinely ask. It is usually the best to-do list you will get for the website and for the business.
-
Limits you set
Subjects it does not touch, prices it does not confirm, things it always passes on. Written by you and obeyed without exception.
-
Data in the EU
Conversations are processed on European infrastructure, with a notice that this is an assistant and not a person — which is both fair dealing and what the EU AI Act requires.
How it gets built
Four to six weeks, and half the work is gathering what the assistant is allowed to say.
-
We pull out the thirty questions you get most, from your real inbox rather than a template.
-
We gather the sources: menu, prices, hours, terms, areas. Anything not written down anywhere gets written now.
-
We build it and then try to break it: trick questions, things it does not know, mixed languages, prices that do not exist.
-
Two weeks with a person reviewing every conversation before it runs unsupervised, and the weekly summary from day one.
What decides the price
The build is quoted on how much of your own material has to be put in order, which is the expensive part: the technology is cheap and gathering and writing what the assistant is allowed to say is not. Separately there is the model’s usage, which for a small business is usually a modest monthly figure paid directly to the provider. If gathering the sources shows your case cannot support a reliable assistant, we say so and stop there.
-
How much has to be written
If your menu, your prices and your terms are already written down, this is quick. If they are nowhere, they have to be written first, and that is the bulk of it.
-
How many languages
Each language is review and testing, not just translation: somebody has to check it invents nothing in any of them.
-
Where it has to live
On the website is the simple case. Connecting it to WhatsApp or email as well adds a channel, templates and a handover path for each.
-
Model usage
Paid per use and directly to the provider. It depends on how many people ask, not on us, and we work out the figure against your volume before starting.
Who it is for
Who it is not for
- If your information is not written down anywhere, it has to be written first. An assistant with no sources invents things, without exception.
- If you expect it to replace the person who serves, it will not, and we do not sell it that way.
- If you get four questions a week, answer them yourself: this does not pay for itself.
And the rules it answers
Guides on this
- AI in a small business: what actually works
Three uses that save hours, three that do not, and why grounding is the whole difference.
- WhatsApp Business, used properly
App or API, what it costs, what to automate without annoying people, and the consent rules that do apply here.
- A website in three languages, built properly
URL structure, reciprocal hreflang and x-default: what makes Google show each person the right version.
Questions about the assistant
If your question isn’t here, write to us and we’ll answer it. Contact
Can it make an answer up?
It is built specifically so that it cannot: it answers from your documents and, when the answer is not there, says so and offers a person. None of this is infallible, which is why the first fortnight is reviewed by somebody and the weekly summary keeps arriving afterwards.
Do my customers know they are talking to a machine?
Yes, it says so in the first line. Besides being the right thing, the EU AI Act requires that a person knows when they are interacting with an AI system.
Where do the conversations go?
They are processed on European infrastructure and kept for as long as you decide, with the privacy notice updated to say so. They are not used to train anything and not sold to anyone.
Can it take bookings or payments?
It can lead to the booking and fill in what it already knows, but the confirmation is made by the booking system, not the assistant. Letting a language model confirm a payment is an unnecessary and avoidable risk.
What happens when I change my prices?
You change the document and the assistant answers with the new price. There is no retraining: it reads from your sources every time, which is exactly why it does not go stale.
Show us your inbox
Send us the ten questions you get most. We will tell you which an assistant can safely answer and which it should not touch.
- Or directly
- hola@appnweb.site
- Monday to Friday, 09:00 – 18:00 (CET)