The attorneys have had the demo. They asked good questions about hallucinations and confidentiality, nodded, and went back to their desks. Meanwhile the paralegals, who will actually run the tool every day, got a link to a video. Six months later, usage is low, and the one paralegal who figured it out on her own is the only person the firm cannot afford to lose.
Staff training is where AI adoption succeeds or fails in a law firm, because staff do the repetitive, document-heavy work that these tools are good at. This page describes the training we run and why it is built the way it is.
Why staff training matters more than attorney training
Attorneys use AI a few times a week for judgment-adjacent work. Paralegals, legal assistants, intake coordinators, and billing staff use it many times a day for extraction, summarization, drafting, and reformatting. The volume is on the staff side, and so is the risk, because an unreviewed summary in a case file is an error that surfaces months later.
Training staff also changes the attorneys' experience. When a paralegal hands over a deposition timeline with page references and a list of open questions, the attorney sees the tool working before ever touching it. Adoption follows the staff, not the other way around.
A curriculum by role: intake, litigation support, billing
We do not run one course. We run short sessions by role, each built around the tasks that person already does.
Intake staff learn to turn call notes into a structured summary, spot missing information, and draft the follow-up email. The AI client intake summaries page describes the template they work from. They also learn what not to do: no conflict decisions, no advice, no deadlines calculated by the tool without a person checking.
Litigation support learns document work: summarizing transcripts into timelines, extracting every date and dollar figure from a records set, comparing two drafts and listing the changes, and preparing a first-pass privilege log description. Every exercise ends with a verification step against the source.
Billing and administrative staff learn narrative cleanup (turning terse time entries into compliant descriptions without changing the substance), plain-language client letters, and internal process documents. They also learn to recognize when a task involves confidential information that the policy keeps out of the tool.
The prompt patterns staff actually use
Most training fails because it teaches "prompt engineering" as a skill rather than giving people a handful of patterns. We teach four.
The first is role, task, source, format: tell the tool what it is doing, what document it is working from, and exactly what the output should look like, with an example. The second is extraction into a table, with a rule that every row cites a page or paragraph. The third is compare and list, used for drafts, contracts, and medical records against a summary. The fourth is rewrite for audience, which handles client letters and internal explanations.
Each pattern is saved as a Claude Project with instructions already written, so the paralegal opens the Project, drops in the document, and asks the question. The training then becomes "which Project do I use for this," which people remember.
Practice files: training on your own documents safely
Generic training examples do not stick. We build the exercises from the firm's own document types, using closed files that have been redacted, or synthetic files built to look like them. Staff see the tool handle a transcript formatted the way your court reporters format them, and an intake note written the way your intake coordinator writes them.
This is also where the firm's AI policy gets taught in practice rather than read. The exercise includes a document that the policy says stays out of the tool, and the right answer is to recognize it. Rules land when they are attached to a file someone has in front of them.
Measuring whether the training stuck
Three measures tell you whether the training worked. First, usage by Project: if the intake summary Project is used daily and the transcript Project is never touched, you know which session to rerun. Second, review findings: supervising attorneys note how often an AI-assisted work product needed substantive correction, and whether that rate falls over the first quarter. Third, staff requests: people who have been trained well ask for new Projects and new use cases within a month. Silence means the training did not take.
Refresher sessions every quarter are short, forty-five minutes, and built around what changed: a new Project, a new policy rule, a mistake that was caught and what it taught. Firms whose intake staff are trained this way are usually ready for the next step, which is client intake automation that feeds the summaries straight into the case management system. All of this is part of our Claude training for lawyers program, which despite the name is mostly about the people who are not lawyers.
Questions we get
How long does staff training take?
A role-specific session runs about ninety minutes, with a follow-up thirty days later. Most staff are productive with two or three Projects within a week. The follow-up is where the real questions come out, after people have tried the tool on live work.
Do paralegals need to understand how the models work?
They need to understand that the tool predicts plausible text, that plausible is not the same as true, and that anything it says about the law or about a document must be checked. That takes ten minutes to explain and a practice exercise to make real. Deeper technical background is not required.
What if a staff member refuses to use it?
That is fine, provided they still follow the policy and the firm's review requirements when reviewing others' AI-assisted work. Forced adoption breeds sloppy use. Most reluctance fades once a colleague shows a task that used to take an afternoon taking twenty minutes plus a careful check.
If you want a training plan built around your staff's actual roles and documents, tell us what you are working with.