How to Use AI as a Lawyer: Uses, Limits and the Risks

A brief has a name on it, and a filed citation that turns out not to exist is the kind of mistake a court notices and sanctions. That single fact should shape where AI belongs in a legal practice and where it does not. The tools are genuinely useful for drafting, summarising and organising — and genuinely dangerous the moment their output is treated as a checked fact rather than a first draft someone still has to verify.

Start with what makes this different from most office work

Two professional duties narrow the options before you even open a tool. Confidentiality: client information is privileged, and pasting it into a consumer chat tool with no enterprise data agreement puts information you have a duty to protect somewhere you no longer fully control. Candour to the court: anything filed under your name is attributed to you, including a citation you did not personally check. Neither duty is new or AI-specific — both predate these tools by decades — but a tool that produces fluent, confident text at no cost per attempt makes it unusually easy to skip the step where you would normally have caught a problem.

The practical fix is the same tiered approach that works in other regulated professions: information you would never put in a general consumer tool, information that is fine inside a vetted enterprise tool with no training on your inputs, and work that stays entirely on you regardless of which tool touched a draft first. How to use AI as a banker and how to use AI as a tax preparer apply the identical structure to client-confidentiality duties in their own fields, and the underlying logic traces back to the same place: NIST’s AI Risk Management Framework, which treats "what happens if this specific use goes wrong" as the question to answer before adoption, not after.

What it is genuinely good at

The pattern across every item below is the same: the tool produces a draft or a summary, and you supply the legal judgement that turns it into something you would actually file or send.

1. 01Summarising a long document against a specific question. "Where in this lease does it address early termination" or "what does this deposition transcript say about the date of the incident" is a search problem over text already in front of the tool, and that is the category it handles most reliably.
2. 02Drafting a first-pass version of a routine, low-stakes document — a standard NDA, a demand letter, an engagement letter — from a template and a short set of facts, so you are editing rather than starting from a blank page.
3. 03Explaining a contract clause in plain language for a client email, once you already know what the clause means and are translating it rather than discovering it for the first time.
4. 04Organising discovery material. Turning a stack of unsorted documents into a categorised index by date, party or topic, which you then review rather than trust outright.
5. 05Rehearsing the other side’s likely argument before a hearing or a negotiation, so you walk in having heard the pushback once already.

A worked example

Say you need a client-facing summary of a 40-page deposition transcript, focused specifically on anything the witness said about a disputed delivery date.

The bad version: paste the whole transcript and ask "what happened with the delivery." You get fluent prose that may quietly conflate the witness’s testimony with a related exhibit, or smooth over a contradiction in the transcript that actually matters to the case — stated with the same confidence whether it is accurate or not, because that confidence is a property of how these systems generate text, not a sign they have verified anything against the source.

The version that works: ask specifically for every passage where the witness addresses the delivery date, quoted with page and line references, with no paraphrasing or synthesis layered on top. That is the same instinct OpenAI and Anthropic both give in their own prompting guidance: name the exact task and the exact constraint, rather than asking a broad question and trusting the synthesis that comes back. You then read those passages yourself, in context, and write the actual summary from what you read — using the tool to find the needle in forty pages rather than to decide what the needle means.

The tool can find the passage faster than you can. It cannot tell you, reliably, whether the case it just cited to support your argument actually exists.

Where it will quietly get you into trouble

• —Case citations and quotations produced from memory rather than looked up. A model can generate a case name, a reporter citation and a holding that sound entirely plausible and do not exist, or exist but say something different — a widely cited survey of hallucination in large language models covers why fluent and accurate are not the same property in these systems. Treat every citation as unverified until you have opened the actual case yourself, every time, with no exceptions for ones that "sound right."
• —Client-identifying or privileged information pasted into a consumer tool with no enterprise data agreement behind it. The confidentiality risk is separate from, and in addition to, the accuracy risk, and it does not pause because the tool was convenient in the moment.
• —Legal research treated as finished rather than started. A summary of "the law on X" is a lead toward the primary sources, not a substitute for reading them, particularly because the law varies by jurisdiction in ways a general-purpose model has no reliable way to track.
• —Strategic or ethical judgement calls dressed up as factual questions. "Does this fact pattern support a fraud claim" gets you an answer with the tone of certainty and none of the accountability — that call stays with you regardless of how confidently the question gets answered.
• —Treating a drafted document as finished because the prose reads cleanly. Fluent, well-formatted text is not the same signal as legally correct text, and a client or opposing counsel will not know the difference until it matters.

Getting started without a project plan

1. 01Write down your practice’s rule on client data before using any tool on a real matter, even if it is one sentence. A short written rule beats no rule, which is what most small practices currently have.
2. 02Pick one recurring, checkable task first — document summarisation against a specific question is the usual right answer, because you can judge the output instantly against the source. Find the repetitive part is a short test if the choice is not obvious.
3. 03Run that task both ways for a few matters and keep the time actually saved, not a general impression that it helped.
4. 04Treat every citation and every generated fact the same way: as a lead you personally verify against the primary source, never as the finished answer. Checking an AI answer when you are not the expert is the general version of that habit, and it is the one habit that matters most here.
5. 05Expand to a second task only once the first is routine. Adopting several uses at once tends to end with nobody using any of them carefully.

The part that actually protects the role

The tasks these tools genuinely speed up — summarising, first-draft prose, organising discovery — are the ones furthest from judgement, and judgement is exactly what a signature and a bar licence exist to make someone accountable for. How to use AI as an auditor covers the same split for a different profession built around evidence and sign-off, and the logic holds here without much modification: a tool that cannot be held accountable for being wrong cannot do the part of the job that actually carries the risk.

The useful stance is neither refusing the tools nor deferring to them. It is being the lawyer who gets a first draft or a document summary in a fraction of the usual time, and then does the actual verification every time, on every citation and every fact that matters. AI answers questions covers which kinds of questions these tools answer reliably and which they do not, and that sorting test applies directly to anything handed to you mid-matter.

The short version

Use AI for document summarisation against a specific question, first-draft routine documents, plain-language explanations and discovery organisation, on information your practice has cleared for it. Never trust a citation, a quotation or a legal conclusion it produces without opening the primary source yourself. Write down a data rule before you start, pick one checkable task, measure the time saved, and expand only once the first one is boring.

That checking habit is also the one the labour market has started pricing in: PwC has measured a real wage premium for workers with demonstrable AI skills across the wider economy — 62%, up from 57% the year before. Coursium teaches exactly that layer — practical use of the tools plus the checking habit that has to go with it, in short lessons rather than a compliance manual. Stay ahead of AI by learning them on your phone.