We ask first.

This website works without cookies. If you agree, we will also use cookies for measuring visits and our advertising. You can change your mind at any time from “Cookie settings” at the bottom of every page. Cookie policy

The clear panel of an AI assistant above a limestone counter, with its sources, a guide, the terms and a booking system, laid out beside it in daylight.

ServicesAI assistants & automations

The assistant on your site answers only from what you know. The rest is lifted to a person.

AI chatbot for your website

The assistant on your site answers only from what you know. The rest is lifted to a person.

An AI chatbot for a website, an AI assistant, answers your customers day and night, in any language, on the site and in messaging apps, from your own data: the catalogue, the availability of this hour, the terms as they are written. The model does not do that on its own; the sources you give it and the limits you set do. Here is how one is built, what it never decides alone, what the law requires from 2 August 2026, and how the conversations stay private.

Tell us about yours
2.8.2026
From this day, an AI assistant says it is an AI from the first contact.AI Act
You
are responsible for what the chatbot on your site says. In 2024 Air Canada was held liable for a wrong answer from its own; the tribunal rejected the idea that the chatbot was a separate entity.Moffatt v. Air Canada, 2024
3
are the right outcomes of any question: an answer from a source, an honest "I don't have that", or a hand-over to a person. An answer "from memory" is none of the three.
No
training of models on your conversations: OpenAI, Anthropic and Google do not, by default, use data sent through their business services to train models.OpenAI, Anthropic, Google Cloud

Last updated:

It does not answer from what it knows. It answers from what you gave it.

A language model, like the ones behind ChatGPT, writes well about almost anything, which is exactly what makes it risky for a business: it will answer with confidence what it does not know. An assistant for your site is built the other way round. Before it answers, it searches your own sources, finds the passage that answers the question and writes the answer from there, with a reference. The technique is called retrieval-augmented generation (RAG): search first, then answer.

Sources come in two kinds. Texts that rarely change: frequent questions, terms, guides, descriptions of services. And data that changes every hour: price, stock, availability, free appointments, order status. The first are read from an index that is refreshed when they change; the second are asked directly, at the moment of the question, of the system you already run.

  • Texts

    Frequent questions, purchase and booking terms, the returns policy, guides, menus. What your person would say, written once.

  • Catalogue

    Products with their attributes, variants and prices, from the shop or the ERP, so it compares and recommends with facts.

  • Availability and bookings

    Rooms, places, appointments, live from the booking system or the calendar, with the booking step in that same system.

  • Orders

    Where an order or a return stands, once the customer has been identified. Without that, the assistant says nothing about anyone's account.

Whatever needs judgement is not the assistant's to decide.

Limits are written before the first conversation, not after the first mistake. The assistant knows what it may settle alone, what it must pass to a person and what it does not touch at all. And when something is not in the sources, it says so instead of filling the gap.

The companies that build the models recommend the same: allow "I don't know" explicitly, ground every answer in a quote from the source, keep the model's general knowledge out. These cut errors a great deal; they do not remove them, which is why what matters goes through a person.

  • Settles alone

    Opening hours, how to get there, what is included, which product fits, whether there is availability, a booking or an appointment within your rules.

  • Passes to a person

    A group or trade price, an exception to a term, a complaint, anything outside the rules. With a summary, so nobody has to ask twice.

  • Does not touch

    Medical, legal or financial advice, payments inside the chat, promises written in no source.

  • Says "I don't have that"

    And offers the next way forward: a phone number, a form, a time when a person will reply.

One assistant, wherever they write to you.

Customers do not only ask on the site. They write on WhatsApp, Messenger, Instagram, Viber, they send e-mail. The same assistant can answer everywhere, with the same sources and the same limits, so the answer does not change with the place it was asked.

Every channel has its own rules for automated messages, and the assistant keeps them. The hand-over to a person happens in the same channel, so the customer never has to start again.

  • The site and the shop

    A chat window that fits your page, can be read by a screen reader and works with the keyboard.

  • Messaging

    WhatsApp, Messenger, Instagram, Viber: the questions already arriving there, with the same answer the site would give.

  • E-mail

    A draft reply or a reply from the same sources, and whatever needs judgement to the right person.

One sentence at the start: you are talking with an AI.

From 2 August 2026, the AI Act asks the assistant to say it is an AI from the first contact. In practice it is one sentence in the greeting, in your own words and in the visitor's language, in every channel where it answers.

That sentence does something more: the customer knows who they are talking to, and asks more freely. When they want a person, they say so, and the assistant passes them on.

We measure what works. We do not read what was written.

A customer writing to an assistant often says more than they would in a form. That trust holds only if the conversations stay where they belong. That is why the assistants we build run on servers in Europe, and we do not read your customers' conversations.

Whether the assistant works shows in numbers, not in texts: how many questions close on their own, how many go to a person, how many stay unanswered, how many end in a booking, an appointment or a sale. When hand-overs on one subject grow, a source or a rule is missing there.

  • Servers in Europe

    Conversations and source data are processed and stored on European infrastructure.

  • No reading

    Noria does not read conversations. Only the person you name sees a conversation, when the assistant hands it over.

  • No model training

    The business services of OpenAI, Anthropic and Google do not, by default, use what passes through them to train models.

  • Measured in numbers

    What is measured is counts and rates, never the customer's words.

Before you put an AI chatbot on your site.

In the order that matters. The first three steps are worth doing even if you decide against an assistant.

  1. Write down the twenty questions you hear most often, by phone, e-mail and message, with the answer you give today.

  2. Find where each term is written (prices, cancellations, returns, delivery) and keep one version of it, at a stable URL.

  3. Name who keeps each source right when it changes.

  4. Decide what the assistant settles alone, what goes to a person and what it does not touch, and who receives each hand-over.

  5. List the systems it must ask live (catalogue, stock, bookings, calendar, orders) and the channels where it will answer.

  6. Write the greeting sentence that says it is an AI, in every language you serve.

  7. Ask where it will run and who will see the conversations, before you choose who builds it.

  8. Test it with real questions, and with hard ones, before a customer sees it, and follow its numbers in the first month.

We build the assistant from your sources. And lift it to wherever they write to you.

We design and build AI assistants that answer from your own information, on the site and in every messaging channel, with the limits written from the start and every decision that needs judgement coming to you ready. We already have an assistant running for a client. On the AI assistants page you can see a Noria study, a hotel's assistant, showing where each answer comes from.

  • The questions and the sources

    We gather what your customers ask and where the answer lives, and write what is missing.

  • Live connections to any system

    Shop, ERP, booking system, calendar, CRM or whatever else you use, through its own interface.

  • Every channel

    Site, WhatsApp, Messenger, Instagram, Viber, e-mail and any other you need, with the same sources and limits.

  • Whichever model fits

    We work with all the major models and choose per job, on servers in Europe.

  • Limits, hand-over, privacy

    What it settles alone, what goes to which person and how. Measured in numbers, without reading conversations.

Start with usPart of AI assistants & automations

The questions we hear most often.

A chatbot with set answers follows scripts: if a question fits none, it gets stuck. An AI chatbot understands the question as the customer writes it, in any language, finds the answer in your sources and writes it naturally. The limits, though, are yours to set, as with the first kind.

It can use OpenAI's models, the ones behind ChatGPT, or those of Anthropic, Google and others. The model is chosen per job; what makes an assistant right is not the model but the sources and limits you give it.

It can, as a person can, and the business that has it on its site is responsible: in 2024 Air Canada paid the fare difference for a wrong rule its chatbot gave. The risk shrinks a great deal when it answers only from sources, with a reference, when it may say "I don't have that", and when anything needing judgement goes to a person.

Yes. From 2 August 2026, the AI Act asks the assistant to say it is an AI from the first contact. One clear sentence in the greeting, in the visitor's language, is enough.

In the assistants we build, Noria does not read conversations; only your person sees one, when it is handed over. They run on servers in Europe, and the business services of OpenAI, Anthropic and Google do not, by default, use that data to train models.

Yes, the same assistant, with the same sources. Every channel has its rules: WhatsApp, for example, has not accepted general-purpose AI assistants on its business platform since January 2026, but it does accept assistants that serve a business's own customers.

Yes, when it is connected to the system that holds them and within the rules you set: availability, free slots, order status after the customer is identified. Payments it leaves to your checkout, not to the chat.

On the sources that must be written or cleaned up, on the live connections, on the languages and channels, and on the volume of conversations. Running the model is usually the smallest part; the largest is getting the sources right.