Service · AI Integration & Data Sovereignty
AI for your company. Your data stays in-house.
Your team wants to work with AI, and you want to stay on the safe side legally. We solve this dilemma with an AI assistant that runs on a server inside your company and knows your documents and systems. Confidential data never leaves the building.
In brief
Your team probably uses AI already. The question is where your data ends up.
In many companies, staff already work with ChatGPT, Claude or other AI services, often with private accounts and without management knowing. This is called shadow AI. Whatever is entered ends up on the providers’ servers, mostly in the US. That includes customer data and contracts.
That quickly breaches the GDPR and the data protection commitments in your customer contracts. A ban does little, because the benefit of AI is too great. The solution is an AI that helps just as much but stays inside your company.
Cloud AI
e.g. ChatGPT, Claude, Copilot
- Runs on the provider’s servers, mostly in the US
- Every input leaves your company
- Control over data and model lies with the provider
In-house AI
open language model on your server
- Runs in your server room or data centre
- No input and no document leaves your network
- You keep full control
Open language models can run on your own servers and are hardly inferior to the well-known cloud services for most office tasks. We choose the right model for your tasks, install it and connect it to your systems.
On the safe side, technically and legally.
We take care of both, technology that demonstrably keeps your data in-house, and the documentation your data protection officer and your customers need. So you use AI without legal risk.
-
Data stays in-house
Requests and documents never leave your network. That meets strict requirements such as professional secrecy and the GDPR.
-
No per-user costs
Once installed, your whole team works with it, without a monthly licence per head.
-
Connected to your systems
The assistant knows your documents and templates and answers with sources.
Case study
Brunnhuber Steuerberatung GmbH
A tax advisory firm in Upper Austria: bookkeeping, payroll, annual accounts and advice for businesses and private clients. A successful team that could take on more clients if day-to-day work took less time.
- employees
- 32
- employees
- clients
- 1,100
- clients
- documents in the archive
- 180,000
- documents in the archive
- data allowed to leave the building
- 0
- data allowed to leave the building
The company is fictional and the figures are illustrative. The approach is exactly how we deliver in-house AI projects.
The result up front
- more time for clients every week
- 82 h
- more time for clients every week
- of the team works with the assistant daily
- 91 %
- of the team works with the assistant daily
- client data with cloud providers
- 0
- client data with cloud providers
- until the investment has paid for itself
- 4 months
- until the investment has paid for itself
The path to the result
Eight steps to your own AI in the company.
Step 01 / 08 · First conversation
The team already uses AI. Nobody knows where the data goes.
The owner of Brunnhuber Steuerberatung GmbH calls after learning something unsettling. Several staff members already use ChatGPT for their work, on their own initiative and with private accounts. Nobody knows exactly what has been entered. Yet client data is subject to professional secrecy. She does not want to ban AI; her team should benefit from it. She needs a solution that is secure and under her own control.
What we do
- First conversation on site or by video call
- Orientation: what can in-house AI do today?
- An honest assessment of benefit and effort
From practice
Without a secure AI tool for the team, shadow AI moves in. With one, you gain productivity and take back control of your data.
Call notes
First call · 45 min
“I found out that some of the team already use ChatGPT for client enquiries. I don’t want to take AI away from them, but it can’t stay like this.”
Our notes
- Several staff use AI privately for work
- A lot of time spent searching files and previous years
- Letters to the tax office are written from scratch every time
- Wish: an AI tool that never leaves the firm
Next step: analyse one week of day-to-day work
Step 02 / 08 · Analysis
Find out where the time goes.
For one week, the staff record how they spend their time. On top of that, we talk to every team. The picture is clear. Almost 150 hours a week go into tasks where an AI assistant can help enormously. First and foremost, searching files and previous years.
What we do
- Analysis of day-to-day work across all teams
- Assessment of which tasks suit AI
- Prioritisation by time saved and benefit
From practice
The biggest lever rarely lies in one spectacular use case. It lies in many small tasks that cost time every day.
- Searching files, previous years and documents
- 46 h
- Answering client enquiries by email
- 38 h
- Drafting letters to the tax office
- 29 h
- Checking and summarising tax assessments
- 21 h
- Researching legislation and circulars
- 14 h
- Total
- 148 h
Basis: one week of time tracking, conversations with all teams.
Step 03 / 08 · Data protection & requirements
Clear rules before the first question is asked.
Together with the owner, we define what the AI may and may not do. The most important requirement is simple. No request and no document leaves the firm’s network. In addition, every person should only find files they already have access to today.
What we do
- Requirements for data protection and professional secrecy
- Adoption of the existing access rights
- Documentation for data protection and clients
From practice
With AI, data protection is a question of architecture. With in-house AI, it becomes a selling point for your clients.
For Brunnhuber Steuerberatung GmbH, only in-house AI is an option.
Step 04 / 08 · Concept
Four tasks where AI helps right away.
The analysis yields four concrete use cases. The same principle applies to all of them. The AI delivers a proposal with sources, the expert checks and decides. Work gets faster, and responsibility stays where it belongs.
What we do
- Use cases selected together with the team
- Templates and instructions for recurring tasks
- Business case with expected time savings
From practice
Successful AI rollouts start with a few clear use cases. Once the team experiences the benefit, further ideas come by themselves.
-
01
Knowledge search in the archive
Questions in plain language, answers from files and previous years, with sources.
-
02
Replies to clients
A draft reply email based on the file and in the firm’s tone.
-
03
Letters to the tax office
Responses and appeals built from templates and the facts of the case.
-
04
Summarising assessments
The key points and deviations at a glance.
The AI proposes, the expert checks and decides.
Step 05 / 08 · Technology & model choice
Your own AI server, sized exactly right.
The AI server sits in the firm’s server room and is connected to the document archive and the practice software. The team reaches the assistant in the browser and directly from Outlook. We do not pick the language model by gut feeling. We test different open models with real tasks from the firm’s daily work and size the hardware to match.
What we do
- Selection and testing of open language models with your real tasks
- Sizing and procurement of the hardware
- Connection to document archive and practice software
From practice
Which model and which hardware fit depends on your tasks and the number of users. We know the common open models from practice and know which one is good enough for your tasks, so you do not pay for capacity you do not need.
The firm’s network
The firm’s team
In the browser and directly in Outlook
AI server
Language model and document search
-
Document archive
Files, previous years, assessments
-
Practice software
Clients and deadlines
-
Templates & knowledge
Samples, circulars, guidelines
No connection to cloud AI
Key facts
- Location
- The firm’s server room
- Language model
- Open model, chosen by testing
- Concurrent users
- The whole team
- Access rights
- Adopted from the existing archive
Step 06 / 08 · Installation & integration
Four weeks from server to finished assistant.
We install the AI server, import the document archive and adopt the existing access rights along the way. Then we set up templates and instructions for the four use cases. The assistant answers in the firm’s language and cites the source for every statement, so any answer can be checked with one click.
What we do
- Installation and setup of the AI server
- Import of the document archive with existing rights
- Templates, instructions and tests with real tasks
From practice
Sources are the key to trust. When every answer points to the original document, the team uses the assistant from day one.
Four weeks to launch
-
Week 1
Hardware, installation, model test
-
Weeks 2–3
Import archive, adopt rights, integration
-
Week 4
Templates, instructions, tests with real tasks
Step 07 / 08 · Training & rollout
Enable the team, not just hand over a tool.
First, four AI ambassadors from different teams work with the assistant for two weeks. Then come short training sessions for everyone and a guideline for working with AI in the firm. The ambassadors remain the first point of contact in daily work.
What we do
- Pilot with AI ambassadors from the team
- Training and guidelines for working with AI
- Support in the first weeks, a fixed point of contact
From practice
Teams that quickly work confidently with AI have one thing in common, namely good examples from their own daily work. That is why we train with real tasks instead of theory.
-
Weeks 1–2
Pilot
Four AI ambassadors work with the assistant every day.
-
Week 3
Training
Two short sessions for everyone, with tasks from the firm’s daily work.
-
Weeks 3–4
Support
We are available daily, the ambassadors help within the team.
-
From week 5
Regular operation
The whole team works with the assistant.
Adopted from the pilot
- A collection of proven questions for every team
- A direct button for email drafts in Outlook
- Dedicated templates for frequent letters to the tax office
Step 08 / 08 · Measurement & result
What it achieved, in numbers.
Three months after launch, the team tracks its time for another week. The five analysed tasks now take less than half the time. Clients usually get their answer the same day, and the time gained goes into new clients and advice.
What we do
- Measurement against the baseline from the analysis
- Operation, maintenance and model updates
- Expansion with further use cases
From practice
The investment in in-house AI is predictable, with no per-user costs and no surprises on the bill. The more the team uses it, the better it pays off.
| Metric | Before | After |
|---|---|---|
| Time for the five tasks per week | 148 h | 66 h |
| Searching files and previous years per week | 46 h | 12 h |
| Reply to client enquiries | avg. 2 days | same day |
| Team works with AI daily | – | 91 % |
Economics per year
- Time gained (82 h × 46 weeks × €55)
- + €207,500
- Operation, maintenance, updates, power
- − €9,600
- Result per year
- + €197,900
- One-off investment incl. hardware
- €70,000
- Payback
- about 4 months
- Time gained valued at €55 per hour.
- Not included: additional clients and happier clients thanks to faster replies.
Transferable
Not just for tax firms.
Wherever confidential data meets a lot of knowledge work, in-house AI saves time without taking a risk. Typical areas:
-
Law firms
Search files and prepare pleadings.
-
Medical practices and healthcare
Summarise findings and draft doctor’s letters.
-
Industry and engineering
Make manuals and design documents searchable without giving away know-how.
-
Finance and insurance
Review contracts and answer enquiries faster.
-
HR departments
Answer questions on policies and contracts, prepare documents.
-
Public administration
Pre-check applications and draft decisions.
Is in-house AI right for you?
If you nod at two or more of these, a conversation is worth it.
- Your team works daily with documents and knowledge from files.
- Confidential data must not go to cloud providers.
- Your staff already use AI without you knowing which data leaves the building.
- A lot of time goes into searching and drafting.