AI for law firms that keeps your data in-house
An AI assistant that knows your firm's archive (matters, opinions, filings, contracts) and answers by always citing the source document. It runs on a server inside the firm: no document is sent to a cloud service.
A legal AI assistant that knows your firm's history
Every firm has knowledge nobody can search any more: years of opinions, filings, contracts and correspondence scattered across network drives, PDFs, Word files and scans. The usual question is "have we written something on this before?", and the answer depends on who happens to remember.
The assistant does one thing well: it searches the firm's archive and answers in plain language, pointing every statement to the document, matter and page it came from. If the answer isn't in the archive, it says so. It does not replace case-law databases: it works on your own material, which no database has.
The difference from ChatGPT or a cloud legal AI tool is where the data lives. Here the model, the index and the documents sit on hardware the firm controls (on site or on a dedicated server). No copies held by third parties, no one else's model trained on your files.
Privilege and GDPR: why the AI stays inside the firm
Professional secrecy
Lawyers are bound to keep confidential everything learned in a mandate. Uploading filings to an external service means handing them to a third party and relying on its terms of use. If the model runs in-house, the question doesn't arise.
GDPR and special-category data
Case files contain health, criminal and family data. With an on-premise system there is no transfer to AI vendors (let alone outside the EU), and the impact assessment covers a perimeter you control: servers, access, logs, backups.
Permissions like your practice management system
Not everyone at the firm sees every matter. Search filters by matter and user before the model is even queried: someone without access to a file never receives an excerpt from it.
Answers you can verify
Every answer lists its sources with a link to the document. A lawyer checks it in seconds; anything not cited is not used. That rule turns hallucinations into a manageable problem.
The pipeline: from archive to a sourced answer
Text extraction and OCR for scans
PDFs, Word files, exported email and scans are converted to text; scans go through OCR. Each document gets metadata: client, matter, date, document type. It is the dullest stage and the one that decides the quality of everything else.
Semantic indexing in a local database
Text is split into passages and turned into vectors that capture meaning, stored in a database on the firm's server. Search finds "early termination" even when the filing says "dissolution before expiry".
Search filtered by matter and permissions
For each question the most relevant passages are retrieved, already filtered to the files that user may see and, optionally, by client, period or document type.
Model answer, always with a citation
An open-weight language model with 70-120 billion parameters (or larger, if the hardware allows) reads only those passages and writes the answer, citing document and page. No outbound calls.
How big an archive really is
A realistic order of magnitude: 50 GB of documents (PDFs, Word, scans) become 5-10 GB of text once extracted, because most of the weight is images and layout. The semantic index usually takes 5-15 GB. That fits comfortably on a single server; the component that really matters is the GPU running the model, sized on how many people use it at once. I don't have a price list: hardware and scope are clear before we start, and we begin with a prototype on part of the archive.
Why the model searches your documents instead of "learning" them
To make an AI "know" an archive you use RAG (Retrieval-Augmented Generation): the model doesn't memorise documents, it looks them up at question time and answers from what it finds. A new opinion is in the index within minutes, a deleted document disappears from answers, and every sentence can be traced to its source.
Fine-tuning (retraining the model on your texts) does something else: it teaches style, format and classification, not facts. A small fine-tune makes sense if you want drafts that follow the structure and wording of the firm's filings, or to sort documents by type automatically. Using it to make the model "know" what is in your matters is a mistake: information blurs, doesn't update and can't be cited.
ChatGPT vs cloud legal AI vs private AI
| ChatGPT and generic chatbots | Cloud legal AI software | Private on-premise AI | |
|---|---|---|---|
| Where documents go | Vendor's servers | Vendor's servers | They stay inside the firm |
| Knows the firm's archive | Only files uploaded by hand | Partly, if you upload documents | Yes, the whole indexed archive |
| Cites internal sources | No | Often public databases | Always, document and page |
| Per-matter permissions | No | Depends on the product | Yes, the firm's own |
| Best for | Generic text, no client data | Research on legislation and case law | Work on case files and the firm's know-how |
Stefano Salvucci
Engineer registered with the Order of Engineers of Rome, 12+ years across software development, architecture and automation, co-founder of BusinessinCloud. I work as a fractional CTO for organisations with no tech team: I design and build the system, install it on the firm's infrastructure and leave the documentation and access with you, not me.
How we start
First a free 30-minute call to understand the archive, the number of users and the constraints. Then a prototype on a sample of the archive (one or two practice areas): if the answers aren't useful, the project ends there with a contained investment. If it works, it is extended to the whole archive and fitted to the way the firm already works.
AI for law firms: common questions
Can a law firm use ChatGPT with client documents?
With client documents it is a delicate choice: files are processed on the vendor's servers and protection depends on the contract and account settings. For generic text it is fine; for privileged filings and opinions, a system running inside the firm is the safer option.
What does on-premise AI mean?
The language model, the document index and the documents themselves sit on hardware the firm controls (on site or on a dedicated server), with no calls to external AI services. It works even without an internet connection.
Does legal AI make things up?
A language model can be wrong. That is why the system answers only from passages retrieved from the archive and always cites document and page: a sentence without a source should not be used. If the archive doesn't contain the answer, the assistant says so.
What is the difference between RAG and fine-tuning?
RAG makes the model search your documents at question time: it is how the archive becomes known, it updates immediately and it allows citations. Fine-tuning retrains the model: it is for style, filing format and classification, not for the content of your matters.
Does it work with scans and old documents?
Yes, scans go through OCR before indexing. Quality depends on the scans: skewed pages, stamps and handwriting read worse, which is one of the things the prototype checks.
What hardware is needed?
Archive and index are small: 50 GB of documents become a 5-15 GB index. The component to size is the GPU running the model (70-120 billion parameters or more), based on how many people use it at once and how fast answers need to arrive. That is settled before buying anything.
How much does AI for a law firm cost?
It depends on archive size, number of users and hardware. I don't have a price list: scope and cost of the prototype are clear before we start, and the prototype on a sample of the archive exists precisely to decide whether to go further.
Let's talk about your archive
30 minutes, free: we assess whether and how an on-premise system can help the firm.