The LLM your privileged data never has to leave for.
An on-premises Business Unit that drafts, summarizes, and searches across your matter files. Runs locally on a Mac Studio in your server room. Nothing crosses your firewall — including the prompt.
Why is German law the hardest market for cloud AI?
It's not a technology problem. It's that the productivity gain and the regulatory risk arrive at the same address.
Privilege risk
Mandate scope
Insurance & audit
The 18-month internal argument every German firm is having.
A first-party view from running this diagnostic in mid-sized German practices in 2025–26.
Partners want the productivity. The compliance lead says no to cloud LLMs. The conversation stalls for 18 months while juniors quietly use ChatGPT on their personal phones — which is the worst possible outcome on every dimension: privilege, audit, and partner control.
The honest answer in 2026 is that a Mac Studio with 192 GB unified memory can run open-weight models that are good enough for the 80% of legal work that's drafting, summarizing, comparing, and searching. Not yet good enough to write a §823 BGB brief unsupervised — but more than good enough to give the associate a 70% draft to refine.
Once the hardware sits in the firm's server room, the BRAO problem disappears. So does the per-seat AI subscription. So does the partner argument about who owns the prompts and the outputs. The compliance lead becomes the demo-runner for prospective clients.
The hardware location is the regulatory answer. The model quality is the productivity answer. In 2026, both arrive in the same Mac.
Two versions of the same Tuesday morning.
Same fee earner, same matter, two operating models.
Cloud LLM ban + shadow ChatGPT
- Compliance bans cloud AI by partner resolution.
- Junior pastes anonymized clauses into ChatGPT on phone.
- Output is reviewed by no one; no audit trail exists.
- Partner doesn't see the productivity gain — but absorbs the risk.
- Result: maximum risk, minimum oversight, half the productivity.
Mac Studio in the server room
- Fee earner queries the local model from their browser, no install.
- Every prompt + output logged to firm-controlled audit DB.
- 70B open-weight model produces drafting on par with cloud frontier.
- BRAO conversation closed; PI insurer satisfied with architecture doc.
- Result: 6–12 hours / fee earner / week back, fully governed.
Four capabilities, one local engine, one audit log.
Built around the four operations that consume 60% of a fee earner's non-billable time.
Index
Reads your DMS (DATEV, RA-MICRO, Advoware, AnNoText, NetDocuments, file shares) into a local searchable vector store. Re-indexed nightly.
Draft
Generates first-draft letters, briefs, NDAs, and client emails in your firm's voice. Trained on your sample documents during configuration.
Summarize
Compresses 200-page bundles into structured matter summaries with paragraph-level citations back to the source file.
Search
Ask a question across every matter the firm has ever touched. Answers cite the source file and paragraph. Permission-aware.
Local only — by design.
We don't offer a cloud version of this Unit. Here's the honest comparison so you can see why we made that choice.
| Local · Mac Studio | Cloud LLMs | |
|---|---|---|
| §43a BRAO / privilege | Aligned — data never leaves controller | Contested — depends on contract & jurisdiction |
| Hardware | Mac Studio M4 Ultra (192 GB), included | None on your side |
| Model class | Open-weight, ~70B parameters | Frontier (GPT-4 / Claude class) |
| Data residency | Your office, your jurisdiction | Vendor cloud (US / EU mix) |
| Per-seat fee | €0 | €20–€60 / user / month |
| Audit log location | Your server, queryable | Vendor cloud, exportable |
| Internet outage | Keeps working | Stops working |
| Prompt confidentiality | Never leaves the LAN | Vendor sees prompt |
| Model update cadence | Twice yearly, you control timing | Vendor pushes at will |
What's recovered fee-earner time worth?
Sliders default to median values from our pilot firms (mid-sized German corporate boutiques). Adjust to your practice.
Talk to Secure Practice AI.
A sandbox version of the Matter Assistant. The production Unit runs on hardware in your office; nothing in this demo leaves your browser session.
What lands in your server room.
The local-deployment package, itemized for your IT lead.
Hardware: Mac Studio M4 Ultra (192 GB)
Sized for 70B-class open-weight inference at usable speed for 5–60 concurrent fee earners. Sits in your existing server rack or under a partner's desk. Needs 1× ethernet, 1× outlet (~120W idle). Fully air-cooled, near-silent.
- Apple M4 Ultra · 32-core CPU, 80-core GPU
- 192 GB unified memory
- 2 TB SSD (encrypted, FileVault)
- 70B open-weight model preloaded
- Browser-based UI, no client install
- Audit log: queryable Postgres on-device
Week 5–6 schedule
- W5 MonMac Studio delivered & racked
- W5 WedDMS connector live, indexing starts
- W5 FriFirst fee-earner pilot session
- W6 TueFirmwide onboarding workshop
- W6 FriBRAO + audit doc handover
Documentation handed to your compliance lead
- Deployment architecture diagram
- Data flow document (annotated)
- Standard DPA (GDPR Art. 28)
- Audit log schema + retention policy
- Open-weight model provenance
- BRAO §43a alignment memo
Three ways to acquire the Unit.
One Unit. Same scope, support, and outcome target across all three.
One-time
- Mac Studio M4 Ultra (192 GB)
- Open-weight 70B model
- DMS integration
- Lawyer onboarding workshop
- 12 months support
12-month installments
- Same scope as one-time
- No interest, no penalties
- Yours after month 12
- 12 months support
Rental
- Hardware loan
- Full support included
- Hardware refresh every 24 mo
- Convert to ownership anytime
Frequently asked questions
Claim the Secure Practice AI Unit.
A 15-minute call confirms regulatory and IT fit. We've answered the BRAO question dozens of times.
About 1 in 4 calls ends with us recommending you don't buy anything.