„The common reflex of postponing AI projects because of data protection law and the AI Act leads in the wrong direction. The legal requirements are not an obstacle but a design principle. Take them seriously and you build systems that hold up technically as well as in terms of liability, and stay predictable."
On-premise AI suite with DATEV connection for GHJ
An interdisciplinary professional services firm in Offenburg runs openly available AI models on its own hardware. The AI has access to the DATEV accounting world and to the firm's own document base, without client data ever leaving the building.
GHJ, Offenburg
Tax advisors, auditors, lawyers
- 41 Colleagues with productive access to AI
- 0 Client records in external clouds
Starting point
Three years of an AI boom, three years of business as usual in professional services.
Ever since AI became widely available, the profession has worked under a hard line: client data simply must not flow into public AI services. Professional conduct rules, confidentiality, data protection law and the EU AI Act draw red lines that no firm can step behind. That applies to GHJ just as it does to every other firm in Germany.
While other industries were talking about productivity gains, nothing was possible in the offices of tax advisors, auditors and lawyers. Partners and staff alike experienced this as a widening gap between what was technically possible and what was legally allowed.
The decision at GHJ was to read the legal requirements not as a prohibition but as a blueprint. If client data cannot leave the firm, then the AI has to come into the firm.
Approach
From a workshop in Offenburg into production, and beyond.
We did not start with a model but with the actual work being done in the firm. Out of daily operation, further building blocks then grew that make a real difference in practice.
- 01
Workshop in Offenburg
Two days in the firm with the partners and staff from tax, audit and law. We mapped the work actually being done, captured the legal guardrails and collected the use cases that tie up hours today and should look different tomorrow. What came out was a prioritised list, not a diagram.
- 02
Architecture and choice of models
We settled on a local inference platform based on an NVIDIA DGX Spark. For the different tasks in the firm, several openly available models are kept available, each used where it fits best. In parallel we designed the architecture for the DATEV connection and the document base, including the permissions and audit model.
- 03
Connecting DATEV and the document base
The DATEV connection brings structured client and accounting data into the AI suite in a controlled way, without duplicating it or sending it outside. On top of that, the firm's own document world is indexed: pleadings, opinions, templates, contract models and internal knowledge documents. That way the system answers not only from what the model knows, but from the lived knowledge of the firm.
- 04
Rollout to the staff
In supported sessions the teams were introduced to the AI suite: what the system can do, where its limits are, and how the client stays protected. Instead of a single training event, an ongoing exchange emerged that surfaces new use cases and feeds them back into daily operation.
- 05
Grown out of daily use: the Excel add-in
In practice it became clear that a considerable part of the work happens in spreadsheets, and that colleagues were switching back and forth between the AI suite and Excel. That became the next step: an Excel add-in that brings the local AI suite directly into the spreadsheet. The AI reads the table and writes into it as well: explaining and adjusting formulas, producing analyses from raw data, checking plausibility, all directly in Excel.
Impact
From watching on the sidelines to using it in production.
The question has changed from "when will we finally be allowed to" into "what do we do with it tomorrow". The firm now uses AI as naturally as industries where this has been possible for longer. The difference is that the legal basis carries the whole thing rather than being added afterwards.
Recurring tasks in tax, audit and law run faster, research effort drops, first drafts take minutes instead of hours. Human judgement remains the place where responsibility sits. The system takes load off; it does not decide.
The approach has proved durable beyond the firm itself: GHJ now also supports other companies in the digital transformation of their finance function and passes on its own experience with local, legally sound AI.
Would you use AI in-house but cannot put data in the cloud?
Local AI that works within the law does not only exist in Offenburg. We build the same setup for other professional services firms and for companies under comparable regulatory requirements: local models on your own hardware, controlled connections to your specialist and document systems, and data protection documented along professional conduct rules and the AI Act. In a first conversation we can see whether this also covers your obligations.