What the assistant does in a day
An ordinary day of a lawyer with an AI assistant looks like this:
1. In the morning: a deadline summary for the week from case files and the calendar, with the cases whose deadlines expire first at the top.
2. An incoming contract: a comparison with the firm’s template, a list of deviations with clause numbers and comments on risky terms.
3. A colleague asks, “Do we have a position on penalties in supply contracts?”: an answer from the firm’s archive with references to opinions and court decisions.
4. A claim or a letter to a client: a draft on the firm’s template with facts from the case materials.
5. Before a hearing: a timeline of events and a list of evidence from the case materials.
The lawyer checks and edits the result instead of starting from a blank page. Savings appear where AI is built into the firm’s process, not opened in a separate tab: according to Thomson Reuters, 22% of organizations have a visible AI strategy, and they are twice as likely to see revenue growth from AI.
How it works with the firm’s documents
The assistant answers not from the model’s general knowledge but from the firm’s knowledge base: contract templates, case positions, legal opinions, internal rules and case correspondence.
Documents are split into fragments and indexed inside the firm’s network. For a question, the assistant finds the right fragments, writes an answer and lists below it the documents it took the text from.
An example: a lawyer asks how the firm calculated the limitation period in disputes over work contracts. The assistant finds opinions on several cases in the archive, writes a two-sentence conclusion and links to each opinion. If the opinions differ, it shows both instead of choosing by itself.
The model runs on the firm’s own server. Text does not go abroad or to services that have no data processing agreement with the firm.
You can replace the model without redoing the rules and connections: the templates and the knowledge base stay, only the engine changes. After rollout, the system, the data and the access rights stay with the firm.
Checking every reference
Models invent case law even with access to a legal database. In May 2024, Stanford RegLab and HAI counted errors in more than 17% of answers from legal AI services and in more than 34% for Westlaw AI-Assisted Research.
So the assistant has checks built into the process:
1. Every reference to a legal rule, a court decision or a firm document comes with its source.
2. The decision number and the quote are compared with the text in the database.
3. The assistant marks a reference that is not in the database as unverified instead of silently removing it.
4. The lawyer sees a document with an unverified reference flagged before it goes to court or to the client.
The lawyer remains responsible for each reference. The assistant shrinks the check from looking up every decision to reviewing the flags.
Access rights and the action log
The assistant has a separate account in the firm’s systems and rights only to its own process. A contract assistant only needs to read templates and incoming contracts; it does not need the archive of all cases or admin rights.
We set up access to cases by the firm’s roles: in answers, a lawyer sees only the cases they already have access to.
All of the assistant’s actions go into a log: who asked a question, which documents the assistant opened and what it answered. Every change to the rules is recorded in the support report.
The log is not only for control: it shows which questions lawyers ask most often, and the knowledge base grows on exactly those.
The assistant does not follow commands from incoming emails and documents. If an email asks it to “forget the rules” or send a client list, the request goes to a person.
People make the decisions: the assistant does not sign documents, file them in court or negotiate.
How to start with one process
We start with a pilot: one AI agent on one process from $1,499 in 14 days. We choose the process in a free 30-minute review: checking incoming contracts, answers from the knowledge base or deadline control.
One process, not everything at once, because on one process the metric is visible already in the pilot. The firm decides on the next step by numbers, not by impressions.
What the firm provides in the first days: access to templates and the archive for the chosen process, examples of typical tasks over a month, the rules your lawyers work by and one responsible person for half an hour a day for disputed cases.
We write the success metric into the contract, for example a lawyer’s hours per task or the share of drafts that need no rework. After the pilot you have a report on the metric, an agent with a perpetual license, the data and the access rights.
The cost of the pilot counts toward the price of an AI department if the firm decides to add agents for other processes: 3–5 AI agents from $4,999 in 4–6 weeks.