Custom GPTs & chatbots – AI with your own company knowledge

LLMs know a lot — but not what applies in your company.

ChatGPT, Claude and others have vast general knowledge – but no knowledge of your products, prices, processes or customer data. And, depending on the model, its knowledge may be a year old. Anyone who wants to change that must expand the model with their own knowledge.

The good news: this is possible. The bad news: there are many ways to do it – and most lead in the wrong direction.

Book a free initial consultation

Start an online audit →

The Problem with Chatbot Tools

Yes, there are many quick tools to connect your own content or upload a few PDFs. It is fast, looks impressive – and when you test it properly, you realise how inaccurate these systems are.

Can you afford a bot wrongly telling an employee how many days of holiday they still have? Or quoting a customer the wrong prices? No.

The reason: simple RAG systems are, at heart, a primitive database of text fragments. The bot looks for the most similar text and returns it – without understanding whether the information is correct, up to date or in the right context.

How we get it right

The most common mistake: waiting for the full system. The team never learns, the solution covers only part of the use cases, and by the time it is finally ready, the technology has already changed twice.

The right approach starts from the bottom up:

Concept first
What should the bot be able to do — and what should it not do? Which questions should it answer, and which should it pass on? A clear concept prevents the system from being expanded in every direction later until it becomes unusable.

This also includes: how often is content updated, who is responsible for it — and how do you keep training, deployment and testing under control without starting from scratch with every change?

Define content clearly
Not too much, not too little. Only content that is truly relevant, clearly structured and indexable – no messy PDFs, no contradictory documents.

Clean pipeline, not simple RAG
Content is linked and placed in context. That needs a clear pipeline:

What are core documents – the foundation that always applies?
What are supporting documents – current, but not always relevant?
And how do control mechanisms for corrections work
Who may change what, and how is it ensured that outdated or conflicting content is identified and cleaned up?

This structure is the difference between a bot that hallucinates and one that gives reliable answers.

Integrate structured data
Price databases, course times, stock levels, employee data – structured data is retrieved directly and precisely, not interpreted. That way, the bot gives the right figure – not a hallucinated one.

Choosing the right model
Depending on requirements, data security, digital sovereignty, cost and performance, there are several options:

Cloud APIs from OpenAI, Claude or Gemini for most applications.
European alternatives such as Mistral for privacy-sensitive sectors.
Or local, on-premises open-source LLMs that run entirely on your own infrastructure. No data leaves the company.

We help you choose the right architecture – not the most expensive, but the most suitable.

Finetuning and Testing
A system is only ready once it has faced real questions, real data and real errors. That takes time – but that is how you build something that truly works.

Access where it is needed
Ideally where staff already work — not in a new tool that first has to be learned. Chat on the website, directly in Teams, Slack, Outlook or Word. Or as a custom GPT in the ChatGPT Store, directly as a Claude connector, or integrated into existing systems via API. We look at what makes the most sense — for internal and external use cases.

Internal or external

Internal applications HR FAQs, onboarding, knowledge management, internal policies, product knowledge for sales, content creation with fact-checks – wherever employees ask the same questions again and again.

External applications Customer service, product advice, support – with clear boundaries, a human handover point and transparent labelling as a bot.