Local AI
Local AI for SMEs.
Your data stays in-house. Your work gets done faster.
The situation
Lots of manual work, little control.
This is what many offices look like today.
Manual routines
Looking up values in Excel and typing them in elsewhere, sifting through tenders, copying deadlines and required documents.
Shadow AI
Staff have long been using ChatGPT and other services, often with private accounts and without approval.
Data with US providers
Whatever is uploaded sits with the provider and falls under the CLOUD Act, even when the servers are in the EU.
AI literacy is required
Since February 2025, companies that use AI must ensure their staff have sufficient AI literacy (AI Act, Art. 4).
The idea
AI that stays private.
The AI model runs on your own hardware in your server room. No text, no document and no calculation leaves the building.
Everything on your own hardware
Open models such as Qwen, DeepSeek or Gemma run on a GPU server of your own. You decide what runs and who has access.
Your data stays in-house
Contracts, drawings, calculations: no uploads, nothing passed on to third parties, no training on your data.
Independent of any vendor
No per-seat fees, no price increases by e-mail. Models can be swapped at any time.
In perspective: In 2026, open models are on a par with the cloud for routine work. On the hardest expert questions the cloud is ahead; but most office work is correspondence, summaries and cross-checking.
In practice
What it can do in the office today.
Ask your documents
“How high is the contractual penalty?” The answer comes with its source.
Evaluate tenders
Identify the relevant lot, extract deadlines and required documents.
Excel transfers
Match lists, fill gaps and check values for plausibility.
Bills of quantities and GAEB
Summarise and compare bills of quantities and GAEB files.
Correspondence
Draft cover letters, minutes and reports from templates.
AI workspace with SSO
Chat in the browser for all staff, sign-in with the company account.
People stay responsible. The AI delivers drafts.
Automation
The AI reads and understands. The code writes. A person signs off.
Whatever follows rules is handled by conventional code. The language model only steps in where structure is missing: free text, PDFs, inconsistent spreadsheets. Results go to a separate output folder, never straight into your original files.
Laufzettel – the workflow
A fixed sequence of steps with permitted tools and folders.
Gegenzeichnung – the sign-off
No result is accepted until a person approves it.
Kontorbuch – the log
A complete record of every run, traceable for data protection and the AI Act.
Performance
One graphics card with 48 GB.
That is enough for an office with 20–30 workstations, because never does everyone generate an answer at the same time.
Measured on an RTX PRO 5000 (48 GB) with Qwen 3.8 27B.
Approach
How we work together.
Workshop.
Half a day on site with your real files. Result: a prioritised list of use cases.
Pilot.
4–6 weeks at a fixed price: one automation and the AI workspace for 5–10 testers. It ends with a go or no-go based on measured figures.
Rollout.
At a fixed project price: set up the server, SSO sign-in, and bring the knowledge base and the automation into production.
Operation.
Updates, model changes and monitoring for a monthly fee. Or a handover to your IT.
Every stage has a clear result and a fixed price. You decide again after each stage.
Next step
Half a day on your real workflows.
We test 2–3 workflows directly on your files, live and locally. You get a prioritised list and a fixed-price offer for the pilot.
