Discovery call →
Deutsch
On-premises AI systems run entirely on your own servers or machines. Your data never leaves your company at any point – unlike cloud services, where your input is processed on someone else's servers. We advise, plan and implement AI solutions that work exactly where they belong: with you.
On-premises AI means the AI models run on hardware that belongs to you – in your server room, on your workstations, or on dedicated machines in your network. Nothing goes to OpenAI, Google, Microsoft or any other external cloud provider.
That has three immediate consequences:
Your data never leaves your company. No processing on US servers, no questions about the legal basis for data exports under Art. 6 GDPR, no dependence on the privacy policies of American corporations.
No recurring API fees. You invest once in hardware and setup, after which your systems run without monthly billing per token, per request or per user.
You decide which models run, how they are configured and who has access. No cloud provider can change features, raise prices or discontinue the service.
On-premises AI is not always the better choice. There are scenarios where cloud services make more sense. That is why we advise honestly – even when it means advising you against our own service.
| Criterion | On-premises AI | Cloud AI (e.g. ChatGPT, Gemini) |
|---|---|---|
| Data protection | Data stays inside the company | Data is processed externally |
| GDPR compliance | Fully given | Depends on provider and configuration |
| Running costs | No API costs after setup | Billed per token / per user |
| Model quality | Very good for specialised tasks | Broader, often stronger on general tasks |
| Adaptability | Full (fine-tuning, RAG, your own data) | Limited (prompt engineering, limited adaptation) |
| Hardware investment | One-off, predictable | No hardware of your own needed |
| Scaling | Bounded by your own hardware | Practically unlimited |
| Dependency | No external dependency | Provider can change prices, features, terms |
If your data is sensitive, if compliance requirements apply, or if you want to stay independent in the long run: on-premises AI. If you want to start quickly, have little IT infrastructure, or are solving general tasks: cloud AI may make more sense.
From transcription through to a complex document pipeline: every solution runs entirely on your own infrastructure.
Speech to text, without the cloud
Meetings, interviews, phone calls – transcribed entirely on your hardware. No audio leaves your company. The basis is Whisper from OpenAI, run on your premises via Ollama or a dedicated Whisper instance.
Typical applications: producing meeting minutes automatically · medical documentation · quality assurance interviews · capturing knowledge from experienced colleagues
Specialist documents, processed on your premises
Technical documentation, contracts, patents, regulatory submissions – translated with on-premises language models, tuned to your specialist terminology. No document leaves your company.
Our particular advantage: as a former translation company with over 30 years of experience across 30+ languages, we know when AI translation is good enough – and when it is not.
Typical applications: technical documentation into several languages · CE declarations of conformity · patent applications · multilingual knowledge bases
Your documents, intelligently searchable
Retrieval-augmented generation connects your existing documents to an on-premises language model. The result: an AI assistant that answers questions about your own material – precisely, with sources, and entirely on your premises.
Typical applications: knowledge management and internal search · onboarding new colleagues · technical support · compliance checks · patent research
Training AI models on your specialist knowledge
When a pre-trained model does not command your specialist terminology or does not match your particular task profile: fine-tuning on your own company data makes the model your model.
Typical applications: domain-specific text generation · specialised translation with company glossaries · document classification · industry-specific assistants
Automation without the detour through the cloud
Classification, extraction, summarisation, routing – automated entirely on your premises. Incoming documents are processed without ever touching an external system.
Typical applications: automatic processing of incoming invoices · contract management · preparing regulatory submissions · archiving with AI metadata
Before we implement anything, we understand where you stand. The assessment takes 1–3 days and produces a concrete plan.
Hardware, network, existing systems, security architecture – we record what is there.
Which use cases should AI support? Which data is involved? Which compliance requirements apply?
Is your existing hardware enough? What would need adding? We appraise the options and costs realistically.
You receive a written report with concrete recommendations, a schedule and a cost range.
What you need depends on the use case. This overview gives you a first orientation.
| Requirement | Entry level | Professional | Enterprise |
|---|---|---|---|
| Mac | M1+ with 32 GB RAM | Mac Studio with 64–128 GB RAM | Mac Studio cluster or dedicated server |
| PC | GPU with 16 GB VRAM | GPU with 24+ GB VRAM (e.g. RTX 4090) | Multi-GPU setup or dedicated server |
| Server | – | Hetzner GPU server (hosted in DE) | On-premises server with professional GPUs |
| Suitable for | Transcription, simple RAG systems, first trials | RAG over larger volumes, fine-tuning, translation | Production systems with high throughput, many concurrent users |
These figures are guide values. What you actually need depends on model size, the number of concurrent users and the use case. The assessment settles what makes sense for your project.
We do not sell on-premises AI at any price. If cloud AI makes more sense for your use case – because you want to start quickly, process no sensitive data, or because model quality is decisive – we will say so. A position, not a sale.
“Buy or build? I guide strategic AI decisions honestly – even when the answer is ‘cloud AI is enough'. A position, not a sale.”
Thirty minutes – you describe what you need, we appraise feasibility and effort. No quote, no obligation.
1–3 days of stocktaking: hardware, requirements, data protection context. Result: a written report with a concrete plan.
Choosing the right model (Whisper, Llama, Mistral, Qwen and others), system architecture, data protection concept, interface definition.
Setting up the infrastructure (Ollama, AnythingLLM, custom pipelines), configuration, integration into existing systems.
Quality checks, user testing with your real data, controlled go-live.
Full documentation, training for your people, an optional support contract for ongoing operation.
Three examples from our consulting work – all delivered without cloud dependency.
Thirty years of experience in technical documentation and specialist translation for regulated industries make the difference. We know when language has to be precise – and when AI helps with that.
Expert appraisal of AI output in more than 30 languages – we spot quality problems that purely technical tools miss.
The international standard for professional translation services – quality assurance at the highest level.
Thirty years of experience in industry, medicine, law and public administration – we know the compliance requirements that can make AI projects fail.
Stefan Weimar is a TÜV-certified AI manager – your point of contact combines technical depth with a strategic view.
On-premises AI means the AI models run on hardware that belongs to you – in your server room, on your workstations, or on dedicated machines in your network. Nothing goes to OpenAI, Google, Microsoft or any other external cloud provider. Your data never leaves your company at any point.
For many specialised tasks in regulated industries: yes. Transcription, specialised translation, RAG systems over your own documents – here on-premises models often suit your specific needs better. For general conversation and broad knowledge questions, the large cloud models still have the edge. We will tell you honestly which approach makes more sense for your particular use case.
It depends on the use case. Entry-level setups for transcription and simple RAG systems run on a current MacBook M3 or a PC with a good GPU (from 16 GB of VRAM). For production systems with high throughput and many concurrent users we recommend dedicated hardware. The IT infrastructure assessment settles what makes sense for your project – and what it would cost.
After setup: close to zero. No API fees, no monthly subscription, no charge per token or request. You pay once for consulting, setup and, where needed, new hardware – after that the system runs on your own infrastructure with no ongoing third-party costs.
A first proof of concept is often possible in 2–4 weeks – a simple transcription system, say, or a RAG chatbot over your documents. A production-ready solution with integration into existing systems, user testing and documentation takes 4–12 weeks depending on complexity. The assessment produces a concrete schedule.
On-premises AI systems can be updated with new models at any time. The open model ecosystem (Ollama, llama.cpp, LM Studio) is growing quickly – more capable models appear regularly and can be brought in without provider lock-in. You are not tied to a vendor who can raise prices or withdraw features.
No – but that is where we are strongest. Our experience in industry, healthcare, law and public administration makes us a dependable partner where data protection and compliance allow no compromises. Companies outside regulated industries benefit from running AI on their own premises too – for instance where trade secrets or client-confidential data are processed.
In the free discovery call we work out whether and how on-premises AI makes sense in your situation. Thirty minutes that give you clarity.
Equip your team to understand and apply AI systems. Practical workshops for beginners and advanced users – tailored to your industry.
To the workshops →Secure hard-won knowledge systematically before it leaves the company. AI-assisted knowledge capture for the German Mittelstand – minimally invasive, GDPR-compliant.
Discover SaPeReS →