Technology
Three routes to an AI chatbot for your business: what you get, the management and maintenance involved, and the privacy questions to ask for each.
For most organisations, an existing assistant with a business subscription is the first step. If the chatbot must answer from your own knowledge, you choose between an off-the-shelf supplier solution and building it yourself. If you want an assistant to search your own documents without uploading them, a connection is the third route.
“AI chatbot” covers a lot. One person means the window where you type a question, another the helper on a website, and a third a search assistant in their own archive. This article compares the routes: what you get, what they require of your organisation and the privacy questions involved.
This article concerns the choice. The technology is covered in the articles we link to. We do not give prices. This is an explanation, not legal advice. The broad picture on chatbots and personal data appears in ChatGPT, Claude, Mistral Vibe and the GDPR.
Every AI chatbot has a language model underneath: software that predicts text. A product is built around it. ChatGPT, Claude and Mistral Vibe (formerly Le Chat) are assistants: a model with a window, conversation memory, settings and connections. What is an LLM? Large language models explained without jargon explains what a language model is and what it does and does not remember.
An application is what an organisation builds around this: a website chatbot or an assistant consulting contracts. Your own knowledge is added, along with rules about who may ask what.
Answer this first. It determines more than the choice of technology.
A chatbot on your website talks to people you do not know and cannot train. They type whatever they want, including their complaint or medical situation. The chatbot also speaks on your behalf. And the customer must know there is no human on the other side.
An internal assistant works with people you can instruct. The risk is on the other side here: what employees put into it. A colleague uploading a client file to ask one question sends the entire file to the provider. May you put client data into ChatGPT, Claude or Mistral Vibe? explains which data this may and may not be.
| Topic | Chatbot for customers | Assistant for employees |
|---|---|---|
| Who types | Unknown people, without instruction | Colleagues, with arrangements and training |
| Which data | What the customer enters; sometimes data from your customer system | What the employee enters or uploads |
| Biggest risk | A wrong answer on your organisation's behalf; unsolicited sensitive input | Too much data with the provider; personal account use |
| Disclosing that it is AI | Yes, unless obvious | Employees know; record it in your policy |
| What to arrange beforehand | Disclosure, a route to a human, an explanation of what you do with input | Subscription, agreement, settings, AI policy |
An assistant that works tomorrow: writing, summarising, translating and analysing files. A business subscription gives the organisation a workspace with administrators, and business terms apply instead of consumer terms.
Little technology, but organisation. Someone chooses the subscription, enters into the agreement, configures settings and manages accounts. The EU AI Act also requires measures to support your staff's AI literacy. That has been the wording of Article 4 since summer 2026, expressly without requiring a fixed level per employee. Read more in our article on AI literacy.
Answers by provider appear in ChatGPT alternatives for business use: options compared on privacy and, for ChatGPT subscriptions, in which subscription you may use for business.
A chatbot using your own knowledge works with RAG. In its glossary, Anthropic describes RAG as a technique combining retrieval and generation: when a question is asked, relevant pieces are first retrieved from a knowledge base and sent to the model alongside the question. The model then does not need to draw the answer from its training. The steps appear in RAG explained: how to have AI answer from your own documents.
The third route uses an assistant you already have and gives it a tool. The standard for this is the Model Context Protocol. The MCP documentation calls it an open standard for connecting AI applications to external systems, such as data sources and tools. The source stays where it is. The assistant submits a search query and receives an answer rather than a file. Find the explanation in What is an MCP server?.
You get the window employees already know, with access to a source you manage. It requires someone to manage the connection: which assistants may connect, on behalf of which user and with which permissions. And it requires a well-organised source.
| Route | What you get | What it requires | Privacy: what to consider |
|---|---|---|---|
| Existing assistant with a business subscription | A general assistant, immediately usable, with organisational administration | Choosing a subscription, entering into an agreement, configuring settings, employee arrangements and training | Training on input, agreement, data location and especially employees' input |
| Purchased chatbot using your own knowledge | Answers from your own texts, within weeks, with an administration screen | Accurate, maintained content, testing, a supplier contact | A processor chain, knowledge base content, question retention |
| Self-built chatbot using your own knowledge | Full control over source, search step, permissions and logs | Developers, hosting, security and ongoing maintenance | Per-user permissions, where the model runs, index contents |
| Assistant connected to your documents | The familiar window, accessing a source you manage | Managing the connection and keeping the source in order | What the source returns, on whose behalf, and the answer reaching the assistant provider |
Article 50 of the AI Act concerns systems interacting directly with people. Its first paragraph requires the provider to design such a system so people are informed they are interacting with an AI system. This is unnecessary if obvious to a reasonably informed and observant person. The information must be clear and provided no later than the first interaction, according to paragraph 5.
According to the European Commission's timeline, Article 50's transparency rules became applicable on 2 August 2026. Only the obligation to mark generated content in a machine-readable way has a transitional period until 2 December 2026 for existing systems. The full timeline appears in The AI Act in autumn 2026.
The law places the obligation on the provider: whoever develops or commissions an AI system and puts it into service under their own name. If you buy a chatbot from a supplier, that supplier is generally the provider and you use the system. If you build it yourself and deploy it under your own name, that role may be yours. Where the boundary lies in your case is a question for a lawyer.
In practice, it makes little difference: enable the disclosure at the top of the window. The Commission is working on guidance to explain Article 50; its page on transparency for AI content contains a draft dated 8 May 2026. We did not see a final version there on 13 October 2026.
The GDPR applies whenever the chatbot processes personal data, which it almost always does.
PrivaSearch belongs to the third route and is intended for employees, rather than a chatbot on your website. It is the source an assistant may search, with one difference from an ordinary connection: personal data has already been removed before the assistant sees anything.
Documents are analysed within PrivaSearch, without an external AI provider. Detected names, addresses, BSNs (Dutch citizen service numbers) and other data are replaced with pseudonyms that remain consistent within a project. A colleague checks the result against the original and approves it. Until approval, a document is not searchable. See How PrivaSearch works.
The connection uses the open MCP standard and works with Claude, ChatGPT Business and Mistral Vibe. For each search query, the assistant receives a limited number of released, pseudonymised passages with a short source reference, and no files. Every call is checked against the user's permissions and logged. What the assistant receives becomes part of the conversation with its provider, and anything you type yourself goes there directly. A business account with a data processing agreement therefore remains necessary, and you do not type names into the question.
If you do not want an external assistant, you can search within PrivaSearch itself. Pseudonymisation is not anonymisation: for your own organisation, the data remains personal data, so the register and DPIA remain your responsibility. Find background on The risks of AI in your client files.
To see how an assistant answers a question about a client file without receiving the file itself, book a demo.
An AI chatbot is a programme you can talk to in ordinary language, whose answers are generated by a language model. Familiar examples include ChatGPT, Claude and Mistral Vibe. An organisation can use such an assistant or build an application around it that answers from its own texts. The model predicts text and can make mistakes; check answers where something depends on them.
Yes. With a model maker's API and a search step across your own texts, a prototype can be built quickly. The work lies in what follows: per-user permissions, security, keeping the source current and measuring quality. If you do not have developers to keep doing that, buying or connecting is usually wiser.
Article 50's AI Act transparency rules have applied since 2 August 2026. A system interacting directly with people must be designed so they know they are dealing with AI, unless obvious to a reasonably observant person. The information must be there no later than the first interaction. Put the disclosure at the top of the window.
In everyday use, the words overlap. A chatbot often means a website window serving customers within a defined subject. An assistant is broader: a helper for employees that writes, summarises and searches, and can use tools through connections. For privacy questions, what matters most is who talks to it and which data goes in.
It can, if you have made the arrangements. You need a lawful basis, a data processing agreement with the supplier, an explanation telling clients what you do with their input, and a register entry. Process no more than the question requires. Do not let a chatbot make a decision significantly affecting someone without human oversight.
Ensure the assistant receives only the passage needed for the question, rather than files, and that the connection works with the user's permissions. Remove personal data beforehand where possible and have this checked. Use a business account with a data processing agreement, limit the assistant to reading and log every call.
How retrieval-augmented generation works, how it differs from uploading and fine-tuning, why it reduces errors and which privacy questions to ask.
What the Model Context Protocol is, how a connection differs from uploading a file, what an assistant receives and which risks you should check.
What a language model does, what tokens and a context window are, why a model invents things and what it does and does not remember from your input.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →