You want AI on contracts and discovery, but you will not send client files to a public chatbot.
Precedents, opinions and past matters sit in a DMS that nobody can search by meaning.
First-pass review, clause comparison and chronologies eat the evenings before a deadline.
Banks, insurers and public bodies whose outside-counsel guidelines forbid third-party AI.
The rules on confidentiality were written before chat windows existed — and the numbers show what happens when data and tools drift apart.
ABA Formal Opinion 512 (2024): before client information goes into a self-learning AI tool, the lawyer needs the client’s informed consent — boiler-plate in the engagement letter is not enough.
of organisations say non-public company information has already been entered into public GenAI tools — and 27% have banned them at least for a while.
lawyer utilisation — about three billable hours in an eight-hour day. Search, first-pass review and retyping take the rest.
of billable hourly tasks could potentially be automated with AI; for legal assistants the figure is 81%.
of lawyers use AI widely, although 79% use it in some form — adoption mostly stops where confidential files begin.
Put your monthly volume into the calculator and see hours and dollars.
The path of one request through a model that lives in your server room. Nothing in it calls a public API; every answer carries its sources.
New or changed files in iManage, NetDocuments, SharePoint or a matter folder start their own run — permissions come along.
IngestOCR turns scans and image PDFs into text; layout, headings and clause numbers are kept.
IngestEach document is cut into clauses and passages, tagged with matter, client, date and author.
IngestPassages are embedded on the same server and stored in a private vector index — never on a third-party service.
IngestA question in the firm’s chat window or a Word add-in, with the matter selected — answers respect ethical walls.
AskHybrid search — meaning plus exact terms — pulls the clauses and precedents the lawyer is allowed to see.
AskA local model writes the answer only from the passages found, quoting the document and page for every point.
AskNo source, or sources that disagree? The model says so instead of guessing, and points to what it found.
AskA contract is checked clause by clause against the firm’s positions: missing, weaker or unusual terms are flagged.
ReviewRedlines and fallback clauses come from the firm’s own approved language, marked as drafts.
ReviewLong data-room bundles are processed document by document overnight, with a summary per file in the morning.
ReviewQuestion, sources, model version and user are logged — for supervision, billing and client audits.
ReviewA local AI is a language model that runs on the firm’s own server — in the office or in a private cloud account the firm controls — instead of a public chatbot. It reads the firm’s documents, answers questions with citations, compares contracts against the firm’s playbook and drafts first versions from approved precedents. Client files never leave the firm’s network, and no vendor trains a model on them.
That matters because the confidentiality rules did not change when AI arrived. ABA Formal Opinion 512 asks lawyers to understand where client data goes before it enters an AI tool, and to get informed consent when a tool may learn from it. A model the firm owns removes the third party from that question.
Precedent search takes seconds instead of an afternoon in the DMS. First-pass contract review arrives with the issues already marked. Long bundles are summarised overnight. Lawyers keep the judgement, the signature and the client relationship; the model keeps the reading.
Typical minutes for one commercial contract in a mid-size firm. Your own numbers go into the calculator below.
Based on 2 h 45 min by hand and 43 min with local AI per item, from the table above.
The model finds, compares and drafts. Every judgement, every signature and every word sent to a client stays with a lawyer.
Opinions and strategy are written by lawyers; the model only brings the sources.
A lawyer reads and signs off each redline and letter.
The firm decides with the client how AI is used on the matter.
Access rules come from the DMS; nobody sees matters they are walled off from.
Opposing counsel and the courtroom stay with people.
Flip a switch to hand a step to the model or take it back.
Answers grounded in the firm’s own documents, with sources a lawyer can open, from a model the firm owns — and a log that satisfies the supervision rules.
We look at the documents and the work, agree what the model may and may not do, and size the hardware.
The model on a server in your network, on your real documents, with a small group of users.
Access rules, logging and integrations, then the whole team — with the old way kept as a fallback.
Model updates in maintenance windows, monitoring and a monthly report of use and time saved.
It removes the main risk ABA Formal Opinion 512 describes: client data going to a third party and being used to train its model. The model runs on the firm’s own server or private cloud account, is never trained on your inputs by a vendor, and every request is logged. Whether to tell the client is still the firm’s decision — we help you write the disclosure.
For answering from your own documents, yes — the model is told to answer only from the passages it found and to cite them. Open models such as Llama 3.3 70B and Qwen 2.5 handle contract language well when they are grounded this way; where they are not sure, they say so.
The model does not answer from memory. If the documents do not support an answer, it reports that nothing was found. Every point links to the document and page, so a lawyer checks the source in one click.
For a firm of 10–50 lawyers, one server with a 48 GB GPU or two 24 GB cards runs a 70B model for several users at once. We size it on your document volume and can start on rented hardware in your own cloud account.
Yes. Documents are read through their APIs with the user’s permissions, so ethical walls and matter security carry over. The DMS stays the system of record.
The firm. Weights, index and logs sit on your hardware; we keep no copies and no telemetry. If we stop working together, everything keeps running.
Describe your use case in two sentences — we reply within a day with a hardware and model recommendation.
hello@wireclad.com →