Document and contract analysis
Finding clauses, risks and differences between contract versions. Every answer links to the exact passage in the files.
Law firms
AI for law firms that doesn't take data outside the office
Legal professional privilege rules out sending case files to public chatbots. That is why we deploy closed language models running on a server in the law firm: they analyse documents, prepare reports and track legal changes, and the data goes nowhere.
// In scope
- 01Document and contract analysis
- 02Reports and case notes
- 03Legal monitoring
- 04Client chatbot
// Solutions
What AI does in a law firm
An assistant that knows the case files and current law.
Reports and case notes
File summaries, timelines, evidence lists – in minutes instead of hours.
Legal monitoring
A daily review of the Journal of Laws, Monitor Polski and draft bills from official sources, flagging which cases a change affects.
Client chatbot
Answers organisational questions, collects case details and books consultations – with no access to the files.
// Case study
Local AI for the litigation practice of a commercial law firm
The firm handles court and administrative disputes for large companies. We don’t disclose the client’s name – we are bound by a non-disclosure agreement.
366screens in the clickable prototype
0documents in the cloud
100%answers with source and page
1Mac Studio in the firm’s office
The challenge
Hours of routine
Analysing scans, extracting facts and deadlines, comparing documents and drafting repetitive parts of pleadings.
No cloud at all
Client documents cannot go to ChatGPT or any external API – professional secrecy and contracts require it.
Vendor independence
The knowledge base, prompts, templates and configuration belong to the firm and can be handed over to another contractor.
Verifiability
Every claim the system makes must point to the document and page it comes from.
// What we did
From analysing lawyers’ work to an agent for the whole team
- 01Analysis and architecture. A multi-step process instead of “the whole case into the model at once”.
- 02Case map. Every fact links to its source: document, page, scan fragment. No source, no fact.
- 03Prototype – 366 screens. Cases and documents, discrepancies, deadline calculator, pleading builder, legislation and case-law base, permissions.
- 04AI chat agent. Answers questions about the firm’s entire base in plain language and always cites the document and page.
- 05Pilot plan. Comparing models on anonymised real cases – before the firm buys hardware.
Firm assistant local · offline
Which deadlines expire this week?
Searched: your cases · deadlines · pleadings
Three deadlines expire this week:
- Wed – statement of defence, case X GC 112/26 claim.pdf, p. 1
- Thu – complaint against the decision decision.pdf, p. 2
- Fri – remarks on the hearing minutes minutes.pdf, p. 7
Ask about a case, deadline or ruling…
The agent sees only what a given lawyer has access to. Permissions are checked before the search, so the agent never even learns that someone else’s case exists.
// Prototype
What the system looks like
A clickable prototype on synthetic data – 366 screens. Click a screenshot to see it full size.
// Key decisions
Why the system is built this way
Every decision serves the same goal: data stays in the office and every conclusion can be checked.
RAG and templates, not fine-tuning
Knowledge comes from an external base with citations – easy to update and verify. Style adaptation (LoRA) optionally later.
Everything local
A Mac Studio in the office, no internet access. Remote access via the firm’s own VPN (Headscale / WireGuard) with 2FA, no SaaS middlemen.
Permissions at search level
Every fragment is tagged with client and case. The filter runs before vector search, so another case’s fragment physically can’t reach an answer.
Law by date
Each provision is stored with its period of validity, and the system applies the wording in force on the given day.
Open components only
Licences that allow commercial use. Everything can be exported – no lock-in to us.
Tech stack
Runtime: MLX on Apple Silicon.
Models (pilot): Qwen3.6 27B, Mistral Small 3.x and 4, gpt-oss-120b, Bielik 11B v3 for Polish.
OCR: docling + multimodal Mistral Small.
Search: bge-m3, bge-reranker-v2-m3, BM25 + vectors in Qdrant / pgvector.
Documents and law: docxtpl, Polish Sejm ELI API, court rulings database.
// Result
What the firm has today
A clear next path: pilot → model and hardware choice → staged rollout. The interface can be explored on synthetic data.
- One place for the firm’s knowledgeA chat agent instead of digging through folders and cases.
- A working prototypeTo discuss workflows and priorities with the lawyers.
- Data never leaves the officeAn architecture with no cloud and no external APIs.
- A transparent systemEvery conclusion has a source, and the whole thing can be handed to another contractor.
// CRM and ERP
Your own CRM or ERP, or an integration with what you already use
We can build a system tailored to your company or connect AI to the system you already use. We integrate with the most popular solutions in Poland:
Kancelaria Prawna OnlineLEX (Wolters Kluwer)Legalis (C.H. Beck)Comarch ERP OptimaComarch ERP XLSymfoniaenova365InsERT Subiekt nexoSAP Business OneMicrosoft Dynamics 365OdooLivespacePipedriveHubSpotSalesforceBitrix24Firmao
Your system isn't on the list? If it has an API, file export or a database, we'll write a connector. See also: What is RAG, AI implementation.
FAQ
Questions
Can't find your question? Write to us – we reply within one business day.
Does the model work offline?
Yes. The model and the document base run on a server in the firm. A connection is only needed to fetch legal acts from official sources and can be limited to those addresses.
Will AI replace lawyers?
No. It prepares material: summaries, lists, draft letters. The decision and signature always stay with the lawyer.
What hardware is needed?
Usually a Mac Studio or a GPU server is enough. We size it to the number of users and documents.
// Let's talk
AI for law firms that doesn't take data outside the office
Free online consultation. We'll show you which processes in your industry AI can take over fastest and what it will cost.
