AI assistants & LLM integration
AI assistant development for your business
AI assistant development starts with your data, because an assistant is only as good as what it can draw on. We build assistants and chatbots that work on your own documents and systems, back every answer with cited sources and run in the EU or on your own servers – documented and handed over in full.
Perspective
When an AI assistant is the right call for your business – and when it is not
A language model is not an end in itself. We settle whether it solves your problem before any proposal is written.
An AI assistant of your own is the right answer when your organisation holds a lot of knowledge in written form and the same questions about it come up again and again: manuals, policies, product data, contracts, tickets or technical documentation. A chatbot on your own data finds the relevant passages, writes an answer from them and shows which document it came from. That suits internal support, sales and service, and teams who need to get to grips with extensive material quickly.
The trigger is often similar: knowledge is scattered across shared drives, wikis, PDF folders and inboxes. Search finds files but not answers, and experienced colleagues field the same questions every day. At the same time, some staff already paste internal material into public chatbots – with no overview of data protection or of where that information ends up. An assistant of your own gives all of this a controlled route.
The other side matters just as much: where an answer has to be exact and binding – a calculation, a stock level, the status of a contract – a conventional query or a well-designed form is usually a better fit than a language model. If the underlying documents are out of date or contradict each other, an assistant will not fix them. And if a ready-made product, such as an AI feature already built into your ticketing or document system, covers what you need, we recommend it – and tell you so in the first call.
Services
What we build in this field
AI assistant development is rarely just about a chat window: it covers connecting the sources, respecting permissions and checking every answer.
Chatbots on your own data
An assistant that answers questions from your manuals, policies and specialist documents. With retrieval-augmented generation (RAG) it first looks up the relevant passages and only then writes an answer – instead of guessing from a model's general knowledge.
Cited answers
Every answer points to the documents and sections it is based on, so your team can check it straight away. If the assistant finds no reliable source, it is designed to say so openly rather than invent a plausible-sounding answer.
Connecting your knowledge base
We open up the sources where your knowledge actually lives: file shares, wikis, ticketing systems, databases or line-of-business applications. New and changed documents are picked up regularly, and permissions carry over – people only get answers from documents they are allowed to open.
LLM integration in existing software
Language models as a component of your applications: summarising text, triaging requests, extracting details from documents or preparing draft replies for a person to approve. Results flow into your systems through documented APIs, not into yet another chat window.
Tools and actions via MCP
Where an assistant should also act – looking up an order status or opening a ticket, for example – we connect your systems through tightly scoped tools, such as the Model Context Protocol. You decide what it may do; critical actions can require confirmation.
Evaluation and traces
Every request can be traced: which sources were found, what the model answered and what it cost. With a fixed set of test questions we measure whether changes to the model, the prompts or the data make the answers better or worse.
Approach
How your AI assistant comes together
The same four steps as in every project we run – focused here on what matters for AI assistants.
Analysis
Over one to two weeks, working from real example questions, we establish what the assistant should answer, where the answers live and how well those sources are maintained. It ends in a written assessment with a price range – including whether a language model is the right tool at all.
Plan
In one week we settle the data sources, permissions, model and where it runs: a provider in the EU or self-hosted. The example questions become a test set against which quality can be measured later. You learn which use case goes live first.
Delivery
We deliver in short cycles and check every change against the test set rather than judging by feel. Your specialists test early with real questions and flag wrong or incomplete answers. A first productive version is typically live after six to twelve weeks, depending on scope.
Handover
Source code, credentials, prompts, the test set, documentation and runbooks are handed over in full. Your team learns how to connect new sources and read the traces. We keep operating it if you want us to, monitoring both answer quality and cost.
Example
A typical scenario: an assistant for the internal service team
An assumed case that shows how a project like this can unfold.
Not a client project
Suppose a machinery manufacturer with around 150 staff has spent years filing its knowledge in manuals, maintenance guides, service tickets and an internal wiki. New service technicians take a long time to find their way around it and phone experienced colleagues with their questions, pulling them away from their own work again and again. Some have started pasting extracts from manuals into public chatbots – which, quite rightly, worries the head of IT.
In the analysis we would typically start by collecting real questions from day-to-day service work and checking which documents hold the answers. That usually also reveals which material is out of date or contradicts itself. The first productive version would be deliberately narrow: an assistant covering manuals and maintenance guides, with cited sources, run by a provider in the EU and available only to the service team.
Building on that, the service tickets could follow as a further source, along with a tool that lets the assistant look up a machine's maintenance history. With every extension, the test set from the analysis would show whether the answers improve. How much relief this brings in a given case cannot be stated honestly in advance – the traces show it afterwards: which questions were asked, which sources helped and where answers were missing.
Technology
A replaceable model, with your data under your control
At the core there is usually a RAG pipeline: documents are prepared, split into sections and indexed for vector search, for example directly in PostgreSQL. We call the language model through an OpenAI-compatible API, so it can be swapped later without rewriting the application. The application itself is built in TypeScript or Python, with an interface in React and Next.js where one is needed.
For operation you have a choice: a model provider that processes data in the EU, or an open model running on your own servers or with a hosting provider of your choice. The application runs in containers, and we capture metrics and traces with OpenTelemetry. Which content is logged, and for how long it is kept, is something we decide together with you.
- LLM
- RAG
- Vector search
- MCP
- PostgreSQL
- Python
- TypeScript
- OpenTelemetry
FAQ
Frequently asked questions about AI assistants
What people typically ask before deciding on an AI assistant of their own.
How much does it cost to build a custom AI chatbot for a business?
It depends mainly on scope: the number and condition of the data sources, permissions, connections to existing systems and where the model runs. Smaller tools start in the four-figure range, from around €1,000. Full applications typically land in the five figures, larger platforms above that. Running costs for the model or servers depend on usage and are visible in the traces. After the analysis you receive an estimate with a price range.
Can an AI assistant be run in line with the GDPR?
We lay the technical groundwork: the model runs with a provider that processes data in the EU or on your own servers, access follows your permission model, and we agree together what gets logged. With external providers we choose terms that exclude training on your data, where the contract allows it. The legal assessment stays with your data protection officer or legal counsel – we supply the technical details they need.
How do you stop an AI assistant from making up answers?
With language models it cannot be ruled out entirely, so we work on several levels: the assistant answers from the passages it has found, backs every answer with cited sources and is designed to state openly when information is missing. A test set of real questions shows with every change how reliable the answers are. That way your team can check statements instead of trusting them blindly.
Which data and systems can be connected?
In principle anything that can be read by machine: PDF and Office documents, wikis, file shares, ticketing systems, databases and line-of-business applications with an API. What matters is less the format than the condition, because outdated or contradictory documents lead to poor answers. So we review the sources in the analysis and recommend where to begin – usually with material that is well maintained and often needed.
Can we host the AI assistant ourselves?
Yes. Instead of a cloud provider, an open language model can run on your own servers or with a hosting provider of your choice, so your data does not have to leave your infrastructure. That calls for suitable hardware, usually with GPUs, and open models differ noticeably in quality. In the analysis we compare both routes on your example questions, so you can decide on the basis of real results.
More services
What often goes with it
Web applications & business tools
Internal tools and customer portals that mirror how the work is actually done, instead of complicating it.
- TypeScript
- React
- Next.js
- PostgreSQL
APIs & system integration
Interfaces that connect existing systems cleanly – specified, versioned and documented.
- OpenAPI
- REST
- Webhooks
- OAuth 2.0
Process automation
Recurring work that runs without manual steps in between – with explicit failure handling.
- Workflows
- Queues
- Event-driven
- Cron
Thinking about an AI assistant for your business?
Tell us briefly which questions it should answer and where that knowledge lives today. You get an honest assessment – including when a language model is not the best solution.