Published on 29 October 2026
What is an internal company chatbot?
An internal company chatbot is a conversational assistant reserved for the people in the company, answering their questions from the company's own documents, procedures and software. You talk to it like a colleague, and it answers by citing where what it says comes from.
The difference with a public assistant comes down to two things. A public assistant answers using what it has learned from the whole web and knows nothing about your company: it knows neither your year-end closing procedure nor the file of the client who has just called. An internal chatbot works on your documents, and it respects each person's access rights: whoever asks the question only gets what they are allowed to read.
The word "chatbot" brings to mind the bots on shopping sites, working their way down a tree of pre-written answers. That is not what we are talking about here. An internal chatbot reads your sources and writes an answer to each question; it is also known as a company AI assistant.
What it needs to read
An internal chatbot answers with what it has been given to read, and nothing else. So its quality depends first on the sources plugged in, before the choice of model. That is the limit we write on every one of our pages: our assistant answers correctly when the sources are there, and it says so when they are missing.
The sources worth connecting, in order of how much they earn their keep. Internal procedures and reference documents, instructions, templates, records of decisions: this is what people ask a colleague ten times a day. The documentation of your business software: for an accounting firm, the online help from the accounting vendor and the payroll software. The document management system, shared drives and email, through a connector that only indexes what has been ticked. And answers that have already been approved: a question asked once, with an answer that has been reviewed, should never have to be asked twice.
What sets two chatbots reading the same sources apart: how documents are split up, the search that finds the right passage, and the re-ranking of extracts before drafting. With us, the splitting, the indexing and the re-ranking all happen on your machine; the drafting model receives the question and the useful extracts, never the knowledge base nor a whole document. The page on the company knowledge base describes how this works.
Who sees what
An internal chatbot that reads everything must not tell everything to everyone. A client file, payslips, a contract under negotiation: these documents already have access rights inside the company, and the chatbot has to respect them to the letter.
The point to check before buying: where are access rights enforced? If they sit in the instruction given to the model, "do not discuss files that do not belong to this person", they will not hold: a model can be talked round. They must be enforced server-side, before the model sees anything at all. The search only returns the passages the person has access to, and the model cannot cite what it never received.
That is how ours works. A team member never sees a client file that is not theirs, and neither does the assistant: the scope is checked server-side, within the permissions by profile or by person that you set in the admin area. The firm's clients talk to an assistant limited to their own file, in an area kept separate from the internal space. The details are on the data security page.
It cites its sources, and it says when it does not know
An answer without a source is an answer to check, so an answer that has saved nothing. An internal chatbot must cite the original passage, and that passage must open in one click: the documentation article, the procedure page, the document in the EDM system.
It also has to know when to stay quiet. When the answer exists nowhere in its sources, a good chatbot says so rather than inventing a plausible-looking procedure. An invented procedure costs more time than it saves, and it destroys trust in a single answer.
With us, a question with no answer does not die. If you decide so, it goes to your in-house expert, to your IT person or to the software vendor's support, and it becomes a tracked request. When the answer comes back and an authorised person validates it, it joins the knowledge base. The next day, the same question finds its answer, with its source. That loop is what grows what the company knows, without anyone having to document anything.
One last point: when the chatbot answers a company client, through a client area or a website, its answer is signed as automated. With us, that is a product principle, and it cannot be turned off.
The options on the market, compared tool by tool
Four ways to get an internal chatbot in 2026, plus a fifth, ours. The facts below were taken from the vendors' own pages on 29 October 2026; they change often, so check them before you sign. Each of these tools answers a real need, and the right choice depends on what you already have.
| The tool | What it reads | Price as listed on 29/10/2026 | Where the data lives |
|---|---|---|---|
| Microsoft Copilot Chat and Microsoft Copilot | Copilot Chat: the web, not your documents. Microsoft Copilot, paid: your files, your emails and your Microsoft 365 conversations, within the permissions already in place. | Copilot Chat at no extra cost with a Microsoft Entra account and an eligible Microsoft 365 subscription. Microsoft Copilot: $30 per user per month on an annual commitment, on top of the Microsoft 365 licence. | At Microsoft, in your Microsoft 365 space. Your queries and the answers are not used to train the models. |
| ChatGPT Business | The documents you upload and the connectors to your tools that the administrator switches on. | $20 per user per month on an annual commitment, $25 monthly, two seats minimum. | At OpenAI. Your data is not used for training by default. |
| Mistral's Vibe (formerly Le Chat) | Uploaded documents and around a hundred connectors: email, calendar, Slack, GitHub and others. | Team: $24.99 per user per month. Enterprise on quote. | On Mistral's cloud; the administrator can turn off training on your conversations for the whole organisation. The Enterprise offer installs on your premises or in your private cloud. |
| Dust, a no-code tool | Your tools connected through connectors: Slack, Notion, GitHub, Drive and around twenty others, or any tool via MCP. Assistants are built without writing code. | Pro: €24 per seat per month billed annually, €30 billed monthly, for teams of up to 100 people. Enterprise on quotation. | At Dust, in the United States or in Europe, your choice. |
| IntraIA | Your procedures and your documents, your EDM system, Microsoft 365 (SharePoint, OneDrive, Exchange), the documentation for your business software, and the answers validated by your teams. | €2,250 excl. VAT per year per site until 31 December 2026, with a 24-month commitment; €2,790 excl. VAT after that. Never per user. AI usage included up to €100 excl. VAT per year. | On your own instance: an isolated server we provide at Hetzner, in Germany or Finland, or your own machine. Indexing stays with you; the writing model receives the question and the relevant extracts. |
How to read this table. A per-user price looks small and then multiplies: at $30 per person, a team of twenty costs $7,200 a year, not counting the base licences. A platform installed for the company is paid for once, however many people use it.
What consumer offers do not do, and what we built first: reading your business software through its documentation, relaying a question to vendor support when the documentation is not enough, and having an answer validated by someone at your firm before it enters the knowledge base. The AI agents extend the assistant into their own areas: the software, the figures, payroll.
Two clarifications about our row in the table. The Microsoft 365 connector is developed and tested by us, but not yet seen running at a client; Google Workspace will follow. And the product can also talk to a model installed on your premises: for drafting, that calls for a more powerful machine than the one we supply. The page on sovereign AI says what we deliver and what we do not deliver yet.
What the policy must cover
An internal chatbot does not remove the need for an AI charter, it gives that charter a point: the rule "do not paste client data into a public assistant" becomes workable the day an internal alternative exists. The charter says which tools are allowed, which data never leaves, what a human reviews before it goes out, and who to report a wrong answer to. A template in eight clauses is on the company AI charter page.
The questions we get asked
- How much does an internal company chatbot cost?
- Somewhere between a per-user licence, $20 to $30 per person per month for the offers surveyed on 29 October 2026, and a platform installed for the company. With us: €2,250 excl. VAT per year per site during the launch period, then €2,790 excl. VAT, whatever the number of users, with AI usage included up to €100 excl. VAT per year. The real cost is mostly the time spent choosing the sources and validating the first answers: allow a few hours in the first week.
- Does the data leave the company?
- It depends on the tool, and that is the first question to ask. With a public assistant used without a company account, yes, and they can be used for training. With professional offers, training is excluded by contract, but the documents are indexed at the provider's end. With us, indexing happens on your instance; to draft an answer, the model receives the question and the useful extracts, never the knowledge base nor a whole document. The data security page spells this out in black and white, including what we do not have yet.
- Do you need an IT person to set up an internal chatbot?
- Not to install it, if the vendor handles that: we deliver your instance within 7 days, in your own branding. What you do need is someone on your side who knows which documents are authoritative, which software matters and who is allowed to see what. That is a manager's job, not a technician's. If you have an IT person, in-house or external, they have a role to play on access rights, and we are glad to have them along at go-live.
- Can an internal chatbot run on a local model?
- Yes, and it is a direction we are committed to. Our product can talk to a model installed on your own premises, Ollama, vLLM or any service that speaks the OpenAI dialect, and the provider can be switched in the admin settings. The limit is speed: on a server with no graphics card, a 7-billion-parameter model produces 3 to 6 words per second, too slow for chat, and needs twice the memory of the machine we supply (measured on 7 September 2026). So our advice is to go local later, once prices and capacity make it sensible. The chunking and indexing of your documents, on the other hand, already run on your side.
- What to do when the chatbot gets it wrong
- Look at the source it cites: most of the time the error comes from an out-of-date document or a duplicated procedure, and it is the document that needs fixing. Then flag the answer. With us, an answer only joins the knowledge base once an authorised person has approved it, through a single "To process" queue; a wrong answer that has got in is corrected in that same place, and the correction applies to every question that follows. Your internal policy should say who to flag it to, and promise that no one is penalised for doing so.
To try it
The demo for an accounting firm is open, no sign-up needed. Ask it about a procedure, look at the source it cites, then ask it something it has no answer to: try the demo, or first read what a firm gets out of it. For payroll and HR, see the HR and payroll page. And when you have made up your mind, the buying page states the price with nothing hidden.
