Your client files do not go to a vendor.
Document search, contract review, intake and conflicts, transcription and email triage, running on a computer in your own office instead of somebody else's cloud.
The State Bar guidance is the door, not the wall
The State Bar of California publishes practical guidance on generative and agentic AI in the practice of law. It tells you to vet how an AI vendor handles client data before you use it. A system that runs on hardware you own, under your own roof, mostly removes the vendor from that question.
Several other state bars have published guidance that points the same direction: understand where client confidences go, check the provider's data handling and retention, and supervise the output. None of that is a prohibition. It is a vetting requirement, and the easiest vetting answer is that the file did not leave the building.
You still owe supervision and competence on whatever you install. No architecture removes that, and I would be suspicious of anyone who told you otherwise.
What I install first, usually
- Document search across the firm's own files
- Contract first pass, with a human doing the final read
- Intake and conflicts
- Call transcription
- Email triage
Document search, the one that pays for itself first
Most firms already have the answer to the question they are about to spend an hour researching. It is in a brief from four years ago, in a matter folder nobody remembers the name of, on a server that supports filename search and nothing else.
Search across the firm's own files means an associate can ask in plain language and get back the passages that actually address it, with the file path and the page. It reads your closed matters, your templates, your forms and your precedent bank. It does not reach outside the firm, and it does not need to.
The output is a pointer, not an authority. Anything that ends up in a filing gets read by the lawyer whose name is on it.
AI contract review and legal document review, scoped narrowly
The first pass is the useful pass. Given a contract and the firm's own position on the issues that recur, the system marks the clauses that deviate from your standard, flags what is missing, pulls out the dates, parties, amounts and termination mechanics, and produces a summary a partner can skim in two minutes instead of forty.
What it hands you is a starting point with the boring extraction already done. A lawyer decides what any of it means. On a document review or discovery set, the same approach sorts and clusters so a human reads in a sensible order rather than alphabetically.
Where the model is unsure, the workflow is built to say so rather than guess. A first pass that flags its own uncertainty is worth considerably more than one that sounds confident everywhere.
Legal document automation and intake
The engagement letter, the retainer, the standard motion, the client questionnaire, the conflicts check. These are the documents your firm produces most often and enjoys producing least. Document automation assembles them from your own templates and your own matter data, so the paralegal is reviewing a draft instead of building one from a prior file with the last client's name still in it.
Intake runs the same way: the call or the web form becomes a structured matter record, the conflicts check runs against your own database, and the file opens with the fields already filled. Nothing is accepted as a client by a machine.
The human review rule
Nothing that touches a filing, money or a client goes out without a named person at your firm reading it. That is in the contract.
This is the clause that makes the rest of it defensible. One named employee owns final review of AI output touching a filing, a payment, or a client communication, and the review is logged. When your malpractice carrier or your managing partner asks how this is supervised, that clause is the answer, in writing, before anything is installed.
Why on-premise, for a firm specifically
A law firm's confidentiality duty does not get lighter because a tool is convenient. Running the model on hardware the firm owns means there is no third party holding client confidences, no retention schedule written by somebody else, and no question about whether your matters are training anything.
It also simplifies the conversation with clients who impose their own outside counsel guidelines. If a client asks what AI tools the firm uses and where the data goes, a one-page architecture statement answers that faster than a vendor questionnaire.
How you should check me
Ask for the architecture statement. Hand it to whoever does your malpractice renewal.
Then ask the three questions you would ask any vendor: what leaves the building and under what written rule, who can access the box during maintenance and is that access logged, and what happens to all of it if we stop working together. The answers are the same whether you ask me on a call or read them in the contract, which is the point.
The honest limits
Local open-weight models are strong at search, extraction, classification, summarizing and drafting over your own documents. The hardest reasoning still tends to go to a frontier model, and when it does, it goes out on the firm's own account, with the minimum approved context for that task, under a written data classification. Privileged material is not in that category.
I am also not going to tell you this replaces an associate or promise you a number of hours back. I do not know your matter mix, and a number I invented would not survive your own accounting anyway.
How it goes
- Twenty minute call, free.
- The audit: half a day in the firm, the workflow map, the confidentiality review, and one thing installed that same week.
- The proposal, handed to you inside the audit readout.
- The install, four to six weeks, then go live.
The Operator's AI Audit: $3,500. Credited in full against the Private AI System if you go ahead within 30 days.
Who this is for
Trades and home services get their own page too.