Private and local enterprise AIUse AI without sensitive data leaving the building

For many companies, cloud AI is off the table because of data protection, trade secrets or internal policies. Local AI runs on your hardware, with open models, under your control. I show you when that is the right path and how it starts.

Book a first call Test the AI Appliance

4 data classes: decide the operating path
3 paths: own server, appliance, EU cloud
4 weeks: pilot with real data
0 €: licence costs for the core components

Stacked servers with glowing circuit traces, symbol of local AI on your own hardware

The problem

Pressure to innovate meets data protection

Departments want to use AI, IT and data protection put on the brakes for good reason. Both are right. Four tensions I find in almost every company.

Confidential data must not go to foreign clouds

Contracts, engineering data, HR records, customer correspondence. For much of it there is no legal basis to transfer it to a provider outside the EU. For the rest, there is simply no trust.

Shadow AI is already there

Where no approved tool exists, employees use private accounts. The ban does not work, and IT has no overview of which data goes where.

Costs run away with usage

Tokens, user licences, provider price changes. Anyone who builds AI into processes ties themselves to a cost model they do not control.

Dependence becomes a strategic risk

Models get discontinued, features change, interfaces are rebuilt. Anyone who relies on one provider builds their processes on foreign ground.

Typical situation

Does this sound familiar?

How companies recognise that a local or private enterprise AI is the right answer.

A server inside a glass house with a glowing lock, a blurred cloud behind it — AI-generated

The data protection officer has prohibited or heavily restricted the use of ChatGPT and similar services.
Customers or group policies require data to stay in Germany or the EU.
Employees use AI anyway, just privately and without rules.
A quote for a cloud solution is on the table, but nobody can estimate the running costs over three years.
There are documents, drawings or contracts that must not leave the building even after anonymisation.
IT wants to offer AI, but in a controlled way, with login via the company directory and clear data rules.

The solution

Private enterprise AI: open models on your infrastructure

Local AI means an open language model runs on a server in your network, in your data centre or on a compact appliance. Employees work with it through a web interface, ask questions about documents, draft texts or summarise cases. No request leaves your company.

There are three paths, which can also be combined. Your own server with a GPU when many users or large models are needed. An AI appliance as a pre-configured device when you want to test quickly without a procurement project. A European cloud with a data processing agreement when you want to rent compute but keep data sovereignty. Which path holds up is decided by your data classification, not by the vendor brochure.

I build these environments from open components: Open WebUI as the interface, Ollama as the model server, open language models of European and international origin, ChromaDB for company knowledge. No licence costs for the core components and no lock-in to a manufacturer. If you want to switch later, you can.

Operating path

Which data may go where?

Data classification decides the operating path, not the vendor brochure. Four classes, three paths, one clear rule.

Tap a data class: which paths are permissible?

PermittedOwn serverin your network or data centre

PermittedAI Appliancepre-configured, on the network the same day

Not permittedEU cloudrented compute, data-processing agreement

Typical content

Customer correspondence, quotes, ticket histories, inspection logs

Stays within the company, often not cleanly separable even after anonymisation.

How it works

Four steps to your own enterprise AI

Classify data

Which data may go where? A moderated half-day sorts your information into four classes. The result justifies every later decision to data protection and the works council.

Choose model and operating path

Data class, number of users and use case determine model size, hardware and the operating path: own server, appliance or EU cloud.

Pilot with real data

A defined pilot with your documents and your questions. Four weeks, clear criteria, honest assessment. Afterwards you know whether and how it holds up.

Roll out and operate

Login via your directory, permissions, data sources, training and operations. The competence stays with you, I support as long as needed.

Business benefit

What local AI brings to your company

Full data control
Data stays in-house. That simplifies GDPR, processing agreements and customer audits considerably because there is no transfer to third parties.
Predictable costs
Hardware and rollout instead of tokens and user licences. The effort is one-off and manageable, usage afterwards is unlimited.
No vendor lock-in
Open models and open components. You can swap models, change the operating path and keep your data and your knowledge.
Shadow AI becomes a company offering
Employees get an approved tool they are allowed to use. Usage becomes visible and compliant.
Confidential use cases become possible
Contract review, HR data, engineering documents: cases that would never have been approved in the cloud.
Competence in your own house
Your IT and your departments learn to run and extend the system. The knowledge stays with you, not with the provider.

Recognise your situation? In a free first call we check whether this route fits you.

Book a first call

Examples

Three applications that only work locally

Anonymised scenarios from the mid-market. The benefit is described qualitatively; numbers emerge in the pilot.

Procurement and legal

Analyse contracts and tenders confidentially

Starting point: Supply contracts, framework agreements and tender documents contain prices and terms that do not belong in a cloud AI. Still, a quick summary and clause check would help.
Solution: A locally run language model summarises documents and flags deadlines, liability clauses and deviations from standard terms. The documents never leave the network.
Possible benefit: A faster overview of extensive documents with full data control. The legal assessment stays with the experts.
Possible technology: Local language model with large context window, Open WebUI, document pipeline

IT and management

Replace shadow AI with an approved internal offering

Starting point: Employees use private AI accounts for translations, texts and analyses. IT knows but cannot control it. A ban does not work in practice.
Solution: An internal AI portal with a local model, login via the company directory and clear rules on which data may go in. Plus a data classification that justifies the approval.
Possible benefit: Usage becomes visible and compliant. Employees get a tool they are actually allowed to use.
Possible technology: Open WebUI with directory integration, Ollama, local language model

Customer service

Process customer correspondence containing personal data

Starting point: Service wants AI to prepare answers to customer enquiries. The enquiries contain names, addresses and contract data that must not go to external services without a processing agreement under GDPR.
Solution: Answer drafts are created with a local model that draws on text blocks and product knowledge. No data transfer to third parties, no processing agreement needed.
Possible benefit: Shorter handling time per enquiry, a consistent tone and a solution the data protection officer can support.
Possible technology: Local language model, Open WebUI, RAG on product knowledge

See all use cases

Technology

What I build local AI with

A transparent stack of open components. Which models are suitable depends on your data class and hardware and is a topic for the first call.

OllamaModel server: loads and runs open language models on your hardware.
Open WebUIInterface for employees: chat, document questions, permissions, login via your directory.
Open language modelsModels from Mistral, Meta, Qwen and others, quantised for local hardware. Selected by data class and task.
ChromaDBVector database for company knowledge (RAG) so that answers are based on your documents.
n8nWorkflow automation to connect the AI with ERP, CRM, email and file storage.
HardwareFrom a compact device with AI accelerator to a server with GPUs. Sized by users and model size.

I build, test and implement. The technology is the last step, not the first.

Comparison

Local AI or cloud and pay-per-use solution

Both paths lead to AI. The difference lies in where your data resides, how costs arise and how independent you remain.

Criterion	Local enterprise AI	Cloud or paid solution
Data sovereignty	Your data stays entirely in-house	Your data is transferred to external providers
Cost model	One-off hardware and rollout, then unlimited use	Running costs per use, token or user licence
Operations	Works entirely without an internet connection	Requires a permanent cloud connection
Compliance	Eases GDPR and EU AI Act through local processing	Requires review of provider, location and contract
Dependence	No vendor lock-in, open components	Tied to one provider including price and feature changes
Model quality	On par for everyday tasks, slightly behind the largest cloud models for complex reasoning	Access to the largest models, without control over changes
Competence	Competence building stays in your company	Knowledge stays mostly with the provider

The cost and benefit advantage

Costs remain predictable and limited instead of rising with usage, tokens or licences. The money spent goes into lasting internal competence instead of permanent external costs.

Frequently asked

What companies ask about local AI

Can you run ChatGPT locally in a company?

Not ChatGPT itself, that is a cloud service. But open language models with comparable usability run locally on your own hardware, with Open WebUI as the interface. For most everyday business tasks they are sufficient; for very complex tasks the largest cloud models are ahead.

How good are local models compared to ChatGPT or Claude?

For summarising, structuring, translating, asking questions about your own documents and classifying, good open models are on par in daily use. For very long, multi-step reasoning tasks the large cloud models are stronger. For confidential data the difference is secondary because the cloud is not an option there anyway.

What does a local enterprise AI cost?

Costs consist of hardware, rollout and operations. A test device for one team is in the range of a notebook, a server for many users in the range of a smaller IT investment. There are no running licence costs for the core components. Concrete figures follow from number of users and model size, which I estimate with you in the first call.

Which hardware does a local language model need?

GPU memory is decisive. Small models run on a compact device with an AI accelerator or a current Mac, medium models need a graphics card with 24 GB or more, large models a server with several cards. Much of this can be tested beforehand with an appliance.

Is local AI GDPR-compliant?

Local processing avoids transfers to third parties and with them the hardest GDPR questions. But the system only becomes compliant through a permissions concept, purpose limitation and documentation. Both are part of my rollout.

Do we need our own IT specialists for operations?

The existing IT is usually enough for operations if it looks after servers or virtual machines. I set up the system, document it and train your administrators. For model selection and extensions I remain available.

What about updates and new models?

Open models appear at short intervals. The model server lets you load new models and keep old ones. I recommend a quarterly review in which we assess whether a switch is worthwhile.

Local or European cloud, which is better?

That depends on data class, number of users and IT capacity. Highly confidential data speaks for local. Many users with little infrastructure of their own can do well with an EU cloud and a processing agreement. A combination is often sensible: local for the confidential, rented for the rest.

Get to know local AI

In a free first call we clarify your data situation and whether local, appliance or EU cloud fits you. If you prefer hands-on, start with a four-week appliance test.

Book a first call See the AI Appliance

Private and local enterprise AIUse AI without sensitive data leaving the building

Pressure to innovate meets data protection