Private AI for law firms: search across case files and templates that never leaves the firm

The problem

A law firm lives on what it has already written. The contract template that survived hard negotiations three years ago, a partner's opinion on limitation periods, a pleading from a similar matter: all of it exists, but it is scattered across folders, e-mail and the document management system. When a trainee looks for "that limitation of liability clause we used in the IT contract", it takes half an hour or a question to someone who is busy.

General-purpose chat assistants would help here, but a firm cannot paste a case file into them. Under the Slovenian Attorneys Act, a lawyer must keep confidential whatever the client has entrusted to them, and the duty also binds other people working in the firm. Once a document goes to a foreign provider's cloud service, a third party processes it on servers you do not control. Most firms therefore decide not to use AI at all, which leaves the firm's knowledge where it always was: in people's heads and in folders nobody searches.

What the private AI does

We set up a language model that runs on the firm's own server and connect it to your documents through RAG (retrieval augmented generation). When a lawyer asks a question, the system first finds the relevant paragraphs, then builds the answer from them and lists the sources next to it. The model does not answer from memory; it answers from your material.

A sample question and answer (a made-up template, shown for format):

Question: Do our templates include a limitation of liability clause for software maintenance contracts, and how is it worded?

Answer: Yes. The software maintenance contract template has the clause in Article 11: the contractor's liability is capped at the annual fee, except for intent and gross negligence. A broader version that also excludes indirect damage is in the 2024 licence agreement template.

Source: Templates/IT/Maintenance_contract_v3.docx, Article 11; Templates/IT/Licence_agreement_2024.docx, Article 9

Typical questions firms ask during a pilot:

• Where have we argued that a claim was time-barred because the limitation period was interrupted, and what did the court decide?
• Summarise the key deadlines in this contract and sort them by date.
• Which employment templates contain a non-compete clause, and how long does it run?
• What did we advise client X about documentation for a transfer of a business share?

The answer is always a starting point for a lawyer, who opens the source and decides. The system saves the search; it does not make the call.

Data sources

We always start a pilot with one document set, so answer quality can be measured on real questions. For law firms the most common choices are:

• Contract templates and precedents: the biggest gain on the least sensitive material, a good first step.
• Internal opinions and memos: knowledge partners write once and then look for for years.
• Collected case law: decisions you have already saved and annotated in your matters.
• Closed matter files: pleadings, letters and decisions, with the same access rights as in the document system.

We support the formats firms actually hold: Word, PDF (including scans, with text recognition), e-mail and exports from the document management system. If documents are in several languages, the model answers in the language of the question and cites the original.

Architecture

The whole chain runs on a server the firm controls:

1. Indexing: documents are split into paragraphs, turned into vectors and stored in a local search index. New documents are added automatically.
2. Retrieval: for each question the system finds the most relevant paragraphs and respects the user's permissions, so restricted files return nothing.
3. Model: an open language model (we choose between the Llama, Mistral, Qwen and Gemma families depending on language needs and hardware) writes the answer from the retrieved paragraphs only.
4. Interface: a browser chat, or an OpenAI-compatible API if you want to connect existing tools.
5. Oversight: the log of questions and answers stays with you, so you can review who asked what.

The firm buys the hardware. We size it to the number of users and the volume of documents, and we do not resell it. Smaller firms can often use an existing server with an added graphics card; details are on the private AI page.

Security and regulation

A private deployment removes the main objection: documents do not leave the firm's network and do not go to an outside model provider. That does not mean compliance takes care of itself. During setup we agree on:

• who has access to the system and to which collections,
• how long the question logs are kept,
• how a file is removed from the index when its retention period ends,
• who is responsible for security updates on the server.

Because the system processes clients' personal data, the General Data Protection Regulation (GDPR) applies to it just as it does to your document system. What the EU AI Act means for a firm that uses AI is covered in our article The EU AI Act for SMEs.

The client relationship matters too. Some firms add a sentence about local document search tools to their engagement terms. That is the firm's judgement to make, not ours.

Timeline and price

• Pilot (from 2,900 €, 2–3 weeks): one document set, such as contract templates, a local deployment and an evaluation on 30 to 50 real questions you prepare.
• Production "Company knowledge" (6,000–15,000 €): several collections, permissions carried over from the document system, automatic ingestion of new documents, monitoring and team guidance.
• Support (290–790 € per month): model updates, security patches and quality monitoring.

The pilot runs in three steps. In the first week we choose the document set together, you prepare a list of questions whose answers you already know, and we install the model and index on your server. In the second week we index the documents and test the system on your questions, scoring each answer on whether it is correct, whether the source is right and whether a lawyer would use it. In the third week you get a report with the results, the list of questions where the system went wrong, and a recommendation on whether a production deployment is worth it. If it is not, we say so in writing. The decision to continue is yours.

Prices are final; the provider is not VAT-registered. The full price list and what each tier covers is on the private AI for business page. Firms that want the model to draft in the firm's own style (letters, for example) should also look at the custom fine-tuned model.

What we do not promise

• Correct answers without checking. The model can be wrong or join two paragraphs incorrectly. That is why it always cites the source and why the answer remains a starting point for the lawyer.
• The quality of the largest cloud models on every task. Open models on your own server are very good at document search and summarising, and weaker than the largest cloud models at demanding legal reasoning. We show the difference on your questions during the pilot.
• Automatic compliance. A local deployment reduces risk; responsibility for processing client data stays with the firm.
• Search across public databases. The system only knows the material we include.