Private AI for companies: assistants built on your own documents

At Lagunbot, private AI means an artificial intelligence assistant that answers using your organisation's documentation (manuals, procedures, technical data sheets, internal rules or FAQs) and does not use that data to train any model. In the industry this is known as RAG.

What private AI means in practice

• It answers with your documents: it searches the knowledge base we prepare with you, not the internet.
• Your data does not train the AI: the language model is used with zero data retention, so nothing sent to it is stored.
• You decide which documents go in, who uses the assistant (your team, your customers or residents) and what it must not talk about.
• The chat channel and the conversations can be hosted in the European Union when the project requires it.

We do not install AI models on your company's servers ("on-premise"). In most cases there is no need: what protects the information is deciding which documents the assistant sees, who uses it and what the providers do with the data.

What RAG is, without the jargon

RAG stands for "retrieval-augmented generation". It works like a new, careful employee: before answering, it opens the right folder and replies with what it says there.

1. We prepare your documents and organise them into a knowledge base.
2. When someone asks a question, the system looks for the passages that deal with that topic.
3. The language model writes the answer from those passages.

If a procedure or an opening time changes, the document is updated in the knowledge base and the assistant answers with the new version, with no retraining. That said, the answers are only as good as the documents: that is why part of the work is putting them in order, and why we test the assistant in Basque and Spanish with real questions before it goes live.

What it is used for, by sector

Industry: the knowledge of veteran operators

Much of what an operator with years on the shop floor knows is not written in any manual. An internal assistant gathers that knowledge together with the machine manuals, procedures and technical data sheets, and puts it within reach of whoever has just joined. It also works outwards, as a product assistant that answers customers. We explain it in AI for industry and in the article on generational handover in Basque industry (in Spanish).

Town councils: up-to-date municipal information

Here the information is public, but the principle is the same: the assistant answers with what the town council publishes (procedures, opening hours, services, events) and is updated when something changes. This is what we do for the town councils of Arrasate-Mondragón and Beasain. The full service is described in chatbot for town councils.

Clinics

Internal protocols, instructions for preparing for a test or answers to patients' usual questions, with information validated by the clinic. The assistant does not diagnose or replace healthcare staff. More in AI for clinics.

Academies and training centres

Programmes, calendars, centre rules and enrolment conditions, without waiting for the office to open. More in AI for academies.

How we handle data

Before we start, we set out which provider is involved in each part of the assistant and in which country it is hosted, so that your data protection officer can review it. In the architecture we usually use:

• The chat channel and the conversations can be hosted on servers in the European Union, in Frankfurt, when the project requires it.
• The platform that orchestrates the assistant and stores the knowledge base holds SOC 2 Type II and ISO 27001 certifications.
• The language model is used with zero data retention: it does not keep what is written or use it for training.
• If any part processes data outside the European Union, the transfer is covered by the standard contractual clauses and the EU-US Data Privacy Framework in the providers' contracts.

We also recommend not including personal data about customers, patients or employees in the knowledge base: an assistant that answers questions about procedures almost never needs it. The rest is in our privacy policy.

How it is rolled out

Like the rest of our projects, which we describe in how we work: a free 30-minute assessment, selecting documents with someone on your team who knows the subject, preparing the knowledge base, testing in Basque and Spanish, and going live, as an internal assistant or on your website and WhatsApp. After that, we review every month what people are asking.

Frequently asked questions

Are my documents used to train the AI?

No. The language model works with zero data retention: it reads the passages it needs for each answer and neither stores them nor learns from them.

Do we need to install a model on our servers?

No, and it is not a service we offer. For most companies and institutions it is enough to control which documents the assistant sees, who uses it and where the conversations are hosted.

Does it answer in Basque?

Yes. We set it up to answer in Basque and Spanish, and we test it in both languages before it goes live.

What happens if the answer is not in the documents?

We configure it to say so and refer the person to a member of staff, rather than improvising.

Shall we try it with your documents?

Book a 30-minute meeting or write to us from the contact page.