Request a consultation

In-house AI chat

An AI chat that knows your documents and never leaves your premises

Your staff are already using AI – just with private accounts, on private devices, with company data in other people’s chat windows. Your own AI chat puts an end to that without taking anything away from anyone.

Book an initial consultation AI overview
Runs on your server Answers with sources Permissions per group Model can be swapped at any time

The difference

Why answers with sources are the real point

A language model without access to your documents makes up plausible answers. That is not a weakness of any particular provider but how it works: it predicts the most likely next word, not the correct one.

That is why our set-up works the other way round. The question is first searched against your approved documents, the passages found are passed to the model, and the answer refers to the file and section. Anyone who wants to check clicks through. And if nothing is found, the answer is “there is nothing on this in the documents” rather than made-up information.

This also changes how people use the tool: an answer with a source gets checked; an answer without one gets believed.

  • Your documents are searched, not trained into the model – the model itself stays unchanged
  • That is why the model can be changed at any time without losing any knowledge
  • No answer without evidence – the source is shown underneath

In everyday use

What the chat is actually used for

Not for spectacular things, but for the small interruptions that together take up half the day.

Looking things up instead of searching

“What notice period is in the framework agreement with supplier X?” – an answer with its source, instead of twenty minutes in the file share and a call to the colleague who drew up the contract back then.

First draft

Quote text, rejection letter, minutes, job advertisement. The form is there in seconds; the content remains your job. Especially for texts nobody likes writing, the blank screen is the real hurdle.

Onboarding

New employees ask the in-house documentation instead of interrupting colleagues. What cannot be answered also shows where the documentation has gaps.

First-level support

The chat answers recurring questions from your own instructions. Anything it cannot answer reliably is passed on to a person – together with the conversation so far.

Translating

In-house, without a draft contract or an HR matter going through a public translation service.

Summarising

Long minutes, tenders, audit reports – with references to the passages behind them, so the summary can still be checked.

Scope of services

What we set up

The set-up is done within a few days. After that, the time goes into the data sources and into the question of who may see what – that is where it is decided whether the tool is taken up in the business.

  • Server in a data centre or on your existing hardware
  • Selection and set-up of the language model to suit the task and the available computing power
  • Sign-in with your existing user accounts, with two-factor authentication
  • Interface with your logo and your name
  • Connection of the first knowledge sources – file shares, SharePoint, intranet, manuals
  • Permissions per group – sales sees different folders from accounting
  • Staff briefing and a guidance sheet on what belongs in the chat and what does not

Not included
Cleaning up permissions that have grown over time in existing data – that is a project in its own right and usually the bigger one. Digitising paper files. Guarantees on the accuracy of individual answers. And legal approval by a lawyer or data protection officer. This list is the honest part of the offer: leaving it out only postpones the questions.

What this means for data protection

Personal data in the folders searched remains personal data. AI does not change the legal basis – access rights are what matter. That is exactly why checking permissions comes before go-live in our process, not after. The legal assessment is a matter for your data protection officer; we provide them with the technical information they need for it.

Process

One department, one knowledge source, six weeks

Starting small is not a precaution here but the faster route. After six weeks, the discussion in your organisation is no longer theoretical.

  1. Define the task and source

    Which questions should be answered, and from which documents? One well-maintained source achieves more than five that nobody has tidied up.

  2. Set up and connect

    Server, model, sign-in, permissions, first knowledge source. Usually a few days.

  3. Pilot group works with it

    One department uses the chat in day-to-day work. We look at which questions remain unanswered and add sources or permissions.

  4. Decide

    Roll it out, reshape it or drop it. The decision is based on experience in your own organisation, not on a vendor’s advertising.

Frequently asked questions

Questions about the AI chat

Which model do you use?

That depends on the task. For running in-house, open-weight models that run entirely on your server are an option. For tasks without confidential content, a vendor’s model may make more sense because it can do more. We do not commit to one, and you remain free to switch.

How fast does it respond?

On suitable hardware, within seconds. Response time depends on model size and server performance – we measure both in the feasibility test, before anything is bought.

What about employees who do not want to use it?

Then they do not use it. We have yet to see an organisation where making it compulsory improved anything. Experience shows that two or three people find it useful, tell others about it, and the rest follows by itself – or it does not, which is also an answer.

Can we start small?

That is the recommended route: one department, one knowledge source, six weeks. The set-up stays the same when more is added later – it grows instead of starting again from scratch.

Does the chat see everything in our folders?

Only what has been approved, and for each user only what their group is allowed to see anyway. For that to be right, the permissions have to be right – which is why we look at them before go-live. An AI chat exposes permissions that have grown over time faster than any audit.

What does it cost to run?

The cost consists of server capacity and support. Both depend on the number of users and on the model size, so the figure is fixed after the feasibility test, not before. What we can tell you beforehand is the order of magnitude and what makes it change.

Further reading

Related services and articles

Measure first, then buy

The feasibility test ends with a figure, not a presentation

We take one of your real knowledge sources and a list of real questions from your organisation. Afterwards you know how many of them are answered usefully – and can decide on that basis.

Send an enquiry Call +49 221 984300-0

+49 221 984300-0Switchboard and support hotline

[email protected]Reply within 4 hours on working days

Robert-Perthel-Straße 7250739 Köln – Bilderstöckchen

Mon–Fri 9 am–6 pmEmergency support outside these hours by arrangement