AI assistant for law firms: find clauses, precedents and exhibits across your case files

In short

A useful AI assistant for a law firm is above all an intelligent search engine over your own case files and templates: it finds a clause, an internal precedent or an exhibit in seconds, and it cites its source with every answer. It is not a robot that issues legal opinions.

• Finding a precise clause in a file with several hundred exhibits easily takes 15 to 30 minutes; a well-built RAG assistant brings that down to under a minute, source verification included.
• For a firm of 5 lawyers each saving 2 hours of research per week, that is roughly 440 hours recovered per year.
• On budget: custom development at an agency runs between 15,000 and 50,000 euros; my base, already proven in production, starts at 2 640 € excl. VAT, adapted to your documents.
• Non-negotiable conditions: systematic source citation, EU hosting (or on your premises), no data sent to model training, access rights per case file.
• Acknowledged limit: the assistant finds and cites, it does not analyze and does not advise. Validation stays human, and that is exactly what professional ethics require.

The problem, in concrete terms

A mid-sized litigation file quickly amounts to several hundred exhibits: successive pleadings, opposing party exhibits, correspondence, expert reports, internal notes. Add to that the firm's template library (standard contracts, clauses negotiated over the years, pleading templates) and the archived files of the last ten years. The information almost always exists. The problem is finding it.

In practice, the search relies on three crutches: the partner's memory ("we did an earn-out clause like that in 2023, who was it for again?"), the server's folder tree (folders named differently depending on who created them) and the file explorer's full-text search, which reads neither scanned PDFs nor nuances of wording. The result: an ordinary document search takes between 10 and 30 minutes when you roughly know where to look, and far longer when the person "who knows where it is" is in a hearing or on leave.

That time has a direct cost. For an associate, every hour spent digging through the server is an hour that is neither billed nor spent on substance. And it has an indirect cost: clauses already negotiated get reinvented, an outdated template is reused, an internal precedent that would have saved half a day is missed.

What the assistant does (and what it does not)

Technically, this is a RAG assistant (Retrieval-Augmented Generation): your documents are indexed, and when someone asks a question in natural language ("which asset and liability guarantee clause did we use on the business goodwill transfers in 2024?"), the system finds the relevant passages in YOUR documents, then formulates an answer based on those passages only, citing the source document and page. It is the same mechanism I detail on the RAG chatbot page.

Three rules make the tool usable in a legal context:

• Systematic citation. Every statement points to an identified exhibit. The lawyer verifies in one click. Without this traceability, the tool is a non-starter: a plausible but unverifiable answer has no value in a case file.
• No source, no answer. If the information is not in the corpus, the assistant says so, instead of improvising. This is the central safeguard against hallucinations; I explain the mechanisms in RAG: preventing the AI from making things up.
• No legal advice. The assistant finds, compares, cites. It does not qualify matters legally, does not strategize, does not advise the client. The analysis stays with the lawyer. This is not a marketing precaution: it is the boundary that makes the tool ethically defensible, and it is also where it is genuinely good.

This framing sets it apart from a general-purpose ChatGPT, which answers everything with confidence, including by inventing case law. The examples of lawyers sanctioned for citing decisions fabricated by a generative AI have circulated throughout the profession; a RAG bounded to your documents, by construction, can only cite what exists in your corpus.

Before / after: the Verdier & Associés firm

Take a fictional but realistic case: Verdier & Associés, a business law firm of 5 lawyers (2 partners, 3 associates) in Nantes, around 300 active files and some fifteen years of archives.

Before. An associate is preparing a share transfer and wants to rely on a price adjustment clause negotiated two years earlier. She queries her memory, then the partner's (unavailable, in a meeting), then the server: 40 minutes to lay her hands on the right protocol, in a file filed under the holding company's name rather than the target company's. Multiply this scenario by everyone's daily requests: where is the latest signed version, what did the expert reply on this point, what deadline did we obtain in the similar file.

After. The same question put to the assistant ("price adjustment clause on a share transfer, 2023-2024 files") surfaces in about thirty seconds the three relevant protocols, with an excerpt from each clause and the link to the document. The associate opens the right exhibit, checks the context, adapts it. Total time: 3 minutes. The assistant did not draft anything in her place; it removed the digging phase.

For this firm profile, a conservative estimate of 2 hours of research avoided per week per lawyer gives 10 hours per week, or roughly 440 hours per year over 44 working weeks. That is the equivalent of a quarter of a full-time position recovered, without hiring.

Confidentiality and professional secrecy

This is the number one objection, and it is legitimate: we are talking about exhibits covered by professional secrecy. Four requirements structure the answer:

• Controlled hosting. The documents and the index stay on European infrastructure, or on a firm server for the most sensitive files. I have detailed the options in sovereign and self-hosted RAG.
• No training data. The language model is used in inference only, with contractual conditions that exclude any reuse of your content to train anything whatsoever. This is a contract point, not a verbal promise.
• Per-file isolation. The firm's access rights are replicated in the assistant: an associate queries only the files they have access to, and a file under a Chinese wall can be excluded from the index. The GDPR detail is in RAG and GDPR.
• Logging. Who asked what, when, with which sources returned: essential to audit usage and to reassure both the bar president and the professional liability insurer.

The Conseil national des barreaux has published a guide on ethics and artificial intelligence, and the CNIL has issued recommendations on AI systems: both converge on what I have just described, data control, transparency, human oversight. Better to build the tool within that framework from the start.

What it costs (and what it returns)

On the market, three families of pricing. Subscription legaltech solutions are generally billed from a few tens to a few hundred euros per user per month, with a standardized scope and your data held by the vendor. Custom development at an agency commonly runs between 15,000 and 50,000 euros, sometimes more, for a project that starts from scratch. Between the two, my approach: a RAG base already developed and proven in production (it is the engine running on beforbuild.com, my B2B SaaS, where the assistant answers on business documentation with systematic citation), which I adapt to your documents and your access rules. Fixed package starting at 2 640 € excl. VAT, customization quoted before starting, in service in 2 to 4 weeks. The monthly operating cost (model API, hosting) generally stays under a few tens of euros.

The return calculation is simple and verifiable at your firm. Take the firm of 5 lawyers again: 2 hours saved per week per person, 44 weeks, 440 hours per year. Valued at the employer cost of an associate, around 50 to 60 euros per loaded hour, that is 22,000 to 26,000 euros of time per year. If only a third of those hours converts into billable time at 150 or 200 euros per hour, we are still talking about more than 20,000 euros of annual billing. Even halving these assumptions, the initial investment is covered in the first year. Do the calculation with your own figures: number of lawyers, actual research hours, hourly rate; that is precisely what we work out together during the audit.

Where to start

Three steps, with no commitment at the outset:

• 1. A free 30-minute audit. We look at where research time goes in your firm, which corpora to index first (templates? active files? archives?) and what level of confidentiality is required. You leave with an honest opinion, including if the answer is "it is not worth it for you."
• 2. A pilot on a measurable scope. For example the template library and the files of a single practice group. We measure the rate of correctly sourced answers and the time actually saved, on questions your teams really ask.
• 3. Gradual extension. We expand to the other files, connect full access rights, train the team. The tool is only worth it if it is used every day.

Sources

• CNIL · Artificial intelligence and personal data
• Conseil national des barreaux · Guide on ethics and artificial intelligence
• Regulation (EU) 2016/679 (GDPR), official text on EUR-Lex