AI training for law firms: what to teach first

What should AI training for law firms actually cover?

Most AI training starts with a demo of what a chatbot can do. For a law firm that is the wrong starting point, because the first real question a partner asks is not how the tool works, it is what happens to the matter details once they are typed into it. The ABA's only formal ethics opinion on generative AI, Formal Opinion 512, ties this directly to existing duties: competence under Rule 1.1 requires a reasonable understanding of what a tool does with the input, and confidentiality under Rule 1.6 requires a client's informed consent before anything risking disclosure of their information goes into it. Training that works teaches that boundary in the first fifteen minutes, then moves to hands-on practice on two or three real tasks using a tool the firm has actually vetted.

Which legal tasks are worth training on first?

Start with tasks that are frequent and where a wrong first draft is caught before it becomes a filed document or a client-facing message.

• First-pass legal research, with every citation checked against the primary source before it is used
• Flagging and summarizing documents in discovery or due diligence, as a starting point for attorney review
• Drafting routine correspondence, intake summaries, and status updates to clients
• Summarizing long contracts or agreements ahead of a senior lawyer's read-through

Why does confidentiality decide what a firm can train on at all?

A law firm cannot treat this the way a marketing team treats a new tool, because the content itself is usually privileged or confidential, and that changes what is trainable and what is off-limits before a single prompt is written. Formal Opinion 512 says a lawyer must assess, tool by tool, whether using it risks disclosing client information, and get informed consent before inputting matter details where that risk exists. In practice, that means training has to separate tools covered by the firm's own data processing agreement, where inputs are excluded from model training and access is controlled, from consumer-grade tools with no such guarantee. Teaching which bucket a given tool falls into, before teaching how to prompt it, is the training step most firms skip.

Why is the adoption gap bigger than the training gap right now?

Individual lawyers are moving faster than their firms are managing them. Adoption at the individual level has more than doubled in a year, but most firms still have not caught up with a policy or a training program, which is exactly the gap structured training exists to close. Separately, solo and small firms have reached majority AI use but are not converting it into revenue the way larger firms are, which points at a process problem training alone will not fix: the habit exists, the governance and the workflow around it often does not.

What should a firm avoid when training lawyers on AI?

Two mistakes cause most of the wasted budget and the real risk. The first is buying a legal-specific AI platform before proving, on a general tool the firm already has a vetted licence for, that a task like document summarization or first-pass research actually saves time. The second, and the more serious one, is letting AI output reach a court filing or a client-facing opinion without a lawyer checking every citation and every factual claim against the source. Fabricated case citations from unchecked AI drafts have already drawn court sanctions in multiple jurisdictions. AI can produce a useful first draft of a brief, a summary, or a research memo in minutes. It should never be the last set of eyes on anything signed, filed, or sent as final advice.