On-premise AI for companies: your own assistants and agents, on your infrastructure

Most companies we talk to want AI on their own manuals, contracts, tickets and process data, and many cannot or will not send that data to a provider they do not control. Running the models yourself, or in a cloud account under your own contract, answers the first question your data protection officer will ask.

It is not always the right answer. Hardware has to be bought and run, open-weight models still trail the largest commercial models on some tasks, and a small group of users can cost less in a cloud region. We say which applies to you in scoping, before anything is bought.

• Your data stays in your network or your own cloud account.
• You choose the model and can change it without rebuilding the application.
• Costs follow hardware and capacity you plan, not a token bill that grows with every new user.
• Your works council and data protection officer can see what runs where.

In your data centre when data may not leave the network, or when usage is high and steady enough for own hardware to pay off. In a cloud account you control, on AWS, Azure or the provider you already use, when a quick start or changing load matters more. Both, behind one gateway, when different data needs different rules.

The setup is the same in each case: containers on Kubernetes, a LiteLLM gateway in front of the models and infrastructure as code in your repositories. Moving a workload later is a change of configuration and capacity, so it does not mean starting again.

Systems that answer from your own documents or prepare work for a person to decide. Each one starts with a use case that has a number attached: hours saved per week, answers a team no longer has to look up, or cases prepared per day.

• A service assistant that answers from manuals, spare parts lists and past tickets, for service technicians and the hotline.
• An assistant for quotes and tenders that finds the requirements, earlier offers and the right product data.
• Internal knowledge search that respects the document permissions you already have.
• An agent that prepares one step of a process, such as checking an incoming order, and hands it to a person.

We choose open-weight models by task and test the candidates on your own cases, in German where your users work in German, before one is picked. Where your rules allow it, commercial models such as Claude or GPT can sit behind the same gateway for the tasks they do better.

Hardware is sized from the number of users, the length of the documents and the answer times you need. We recommend the servers and GPUs; you buy them directly or through your usual IT supplier, or start in a cloud account and decide later.

A demo shows that a system can answer. An acceptance test shows how often it answers correctly. Before go-live we agree acceptance criteria with your domain experts, run 100 to 200 test cases from your own documents and cases, and report a pass rate with a confidence interval for each criterion. The criteria go into the contract, so acceptance is a measurement and not a matter of opinion.

The same test cases become the regression suite that runs after every change to the model, the prompts or the documents.

A model, its documents and its users change every month. After handover your team runs the system with the runbooks and training we leave, or we run it under a monthly retainer: model and infrastructure updates, regression tests on every change, monitoring of answer times, errors and cost per team, and a monthly quality report from a named engineer.

Every contract includes a data processing agreement. For systems on your own servers no data leaves your network, and we work only through the access you grant. Your works council gets a plain description of what the system does, what it logs and who can see the logs, and the rules agreed with it are in place before anything goes live.