Updated October 14, 2026: added ABA Formal Opinion 512, the sanctions cases, the Heppner privilege ruling, clearer account-tier data rules, and Anthropic's legal plugins.
ChatGPT can save a law firm real hours on client emails, intake replies, document summaries, and first-draft templates. It cannot do legal research you can trust, and it must never see privileged client information you haven't cleared for a third-party tool. Most guides to AI for law firms get one of those halves right and skip the other. This one covers both: the operational work AI genuinely handles well in a solo or small-firm practice, the setup that makes its output usable instead of generic, and the two hard lines — fabricated citations and client confidentiality — that decide whether AI is an asset or a malpractice risk. No technical background required.
TL;DR
ChatGPT, and AI for law firms generally, works best on the operational layer of a practice: client-update emails, intake responses, plain-English explanations, engagement-letter drafts, and internal procedures. It fails badly as a legal research tool — it can invent cases and citations with total confidence, and lawyers have been sanctioned for filing them unchecked, starting with Mata v. Avianca in 2023. Confidentiality is the other hard line: don't paste privileged client information into a consumer AI account without understanding your ethics obligations and the tool's data policy. Give the AI standing context about your practice areas, jurisdictions, tone, and fees, and the drafts go from generic to genuinely usable. The free Claude Cowork course teaches this approach step by step.
The problem isn't ChatGPT. It's that ChatGPT doesn't know your firm
Open a fresh ChatGPT session and ask it to "write a client email about a delayed filing" and you get something that reads like it came from a firm that doesn't exist: wrong tone, hedged everywhere, generic sign-off, no sense of what you actually told the client last week. That's not a model problem. The AI has no idea what you practice, where you're licensed, how you talk to clients, or what your standard terms are.
The fix is context. An assistant that knows you handle estate planning and small-business formation in Illinois, write in a plain and reassuring voice, charge a flat fee for a simple will, and never opine outside your licensed states produces drafts a careful attorney can edit and send. The same assistant with none of that produces filler you rewrite from scratch, which is slower than not using AI at all.
You build that context once, as a short standing brief: practice areas, jurisdictions, fee structure, tone rules, and the things you never do (give advice to non-clients, quote outcomes, discuss other matters). Paste it at the top of a chat, save it in a project, or set it as custom instructions. Two related guides walk through the mechanics: how to make AI remember your business and how to write a system prompt for your business. The rest of this page assumes you've done that ten-minute step, because every workflow below improves with it.
Client emails and intake replies: the fastest win in the firm
Most of the messages a small firm sends follow a pattern. A client wants a status update. A new inquiry wants to know whether you handle their issue and what a consultation costs. An opposing party's counsel needs a routine scheduling reply. Each one takes five to ten minutes to write well, and there are dozens of them in a week.
AI drafts the polished version in seconds from a one-sentence brief:
Write a short, warm email letting a client know we filed their LLC formation
documents today, that the state typically takes 2 to 3 weeks to process, and
that we'll send the stamped certificate as soon as it arrives. Plain language,
reassuring, no legal jargon, no promises about timing beyond the state's estimate.For intake, paste the inquiry itself and ask for a reply that screens without advising:
Below is a new inquiry from our website contact form. Draft a reply that:
- confirms whether this falls within our practice areas (estate planning,
small-business formation, Illinois only)
- explains our $150 consultation fee and how to book
- asks our three standard screening questions: county of residence, whether
any deadline or court date is pending, and whether another attorney is involved
- gives no legal advice and makes no assessment of their case
[paste inquiry]You read the draft, adjust two lines, and send. For the broader email playbook — templates, follow-up sequences, tone control — see our guide to using Claude for work. The pattern is identical whichever assistant you use.
Document summaries and plain-English explanations
AI is genuinely good at compressing and rephrasing text you hand it, and a lot of a firm's reading time is exactly that. Useful patterns:
- First-pass triage. Paste a contract or filing and ask for the key terms, each side's obligations, every date and deadline, and anything unusual. You still read what matters; the AI tells you where to look first.
- Client translations. "Explain this indemnification clause in plain English for a client with no legal training, in under 150 words, without softening what it obligates them to do." Clients understand their own matters better, and you spend less call time re-explaining.
- Note cleanup. Dictate or paste messy meeting notes and ask for a clean issue list, open questions, and next actions by owner.
The rule that keeps this safe: the AI summarizes what is in front of it. It does not tell you what the law is, and it does not decide what's significant in the matter. It saves you the first read, not the judgment.
Templates, engagement letters, and the SOPs nobody has written down
Engagement letters, standard clauses, and internal procedures are mostly structure, and AI fills structure fast once it knows your defaults.
Engagement letters and standard documents. Ask for a first-draft flat-fee engagement letter with your scope, fee, and standard terms, and you start from an 80% draft instead of a blank page. Ask it to draft a clause "in the style of" ones you paste in, and you edit from a near-match. Every one of these is a starting point for attorney review, never a finished legal document.
Internal SOPs are the easy win, because they're not client-facing and carry near-zero risk. Your intake process, your conflicts-check checklist, your new-client onboarding sequence — in most small firms these live in one person's head. Describe the process out loud, have AI structure it into a checklist, correct what it got wrong, and you have documentation a new hire or a covering attorney can actually follow.
Marketing content works the same way: practice-area pages, a client-facing FAQ, a short newsletter. Feed it your standing brief so it writes in your voice, and review for anything that reads like a guarantee or crosses your bar's advertising rules.
The two hard lines: confidentiality and fabricated citations
This is the section that decides whether AI helps your firm or hurts it. Read it before you paste anything.
Client confidentiality and privilege come first
Your duty of confidentiality governs what you may put into any third-party tool, and a consumer ChatGPT account is a third-party tool. Before client information touches a prompt:
- Know where the data goes. Read the data-handling and retention policy for the specific account tier you're on. Consumer tiers and business tiers treat your inputs differently. OpenAI says ChatGPT Business, Enterprise, and API data isn't used for training by default, and Anthropic says the same for Claude Team and Enterprise under its commercial terms. On consumer accounts (free or individual paid plans), whether your chats can train future models depends on a privacy setting you have to check yourself. On Claude's consumer plans, that choice also changes retention: 5 years if you allow training, 30 days if you don't.
- Use the right account and settings. If you're going to work with matter-related material at all, use a business-tier account, where training is off by default, not a free personal login. "No training" is not the same as "no retention," so still read how long inputs are kept.
- When in doubt, strip it out. Most drafting workflows work fine with names, account numbers, and identifying details removed or swapped for placeholders. "Client A, a 62-year-old widow in Cook County" drafts just as well as the real name.
- Start from ABA Formal Opinion 512, then check your state. The ABA's first ethics opinion on generative AI (July 29, 2024) is the national baseline, though it isn't binding. It says lawyers need a reasonable understanding of a tool's capabilities and limits, must verify output rather than rely on it uncritically, and need the client's informed consent before putting information about the representation into a self-learning tool. Boilerplate in an engagement letter doesn't count as informed consent. It also covers fees: bill the time you actually spent, and don't bill clients for learning a tool you use across your practice. State bars have added their own guidance, including California's practical guidance and Florida Bar Ethics Opinion 24-1. Your jurisdiction's rules of professional conduct, not a blog post, set your obligations.
Privilege is a separate risk. In United States v. Heppner (S.D.N.Y., February 2026), Judge Rakoff held that about 31 documents a defendant created on his own with a consumer AI tool (Claude, in that case) were not protected by attorney-client privilege or work product. The AI isn't a lawyer, the consumer platform owed him no duty of confidentiality, and sharing the output with counsel later didn't make it privileged. It's one district court decision, and another February 2026 ruling took a different view on waiver, but the practical lesson holds: tell clients not to work through their legal problem in a personal chatbot before they talk to you.
None of this makes AI unusable for a firm. It makes it usable the same way any vendor is: deliberately, under a policy, with someone accountable. Accounting practices face the same line with client financial data; ChatGPT for accountants shows how a firm sets up its assistant around it.
Fabricated citations are not a rare glitch
ChatGPT can produce a case name, a reporter citation, a pin cite, and a persuasive quote from a holding — for a case that does not exist. It does this fluently and confidently, and it is exactly the failure mode that has already produced sanctions orders against lawyers who filed AI-drafted briefs without checking. Statutes get misquoted, holdings get inverted, and jurisdiction-specific rules get silently skipped.
The court record is long by now:
- Mata v. Avianca (S.D.N.Y., June 2023): two attorneys and their firm were fined $5,000 after filing a brief built on cases ChatGPT made up.
- Wadsworth v. Walmart (D. Wyo., February 2025): Morgan & Morgan lawyers cited eight cases that don't exist. The drafting attorney was fined $3,000 and had their pro hac vice admission revoked.
- Noland v. Land of the Free (Cal. Ct. App., September 2025): a $10,000 sanction and a State Bar referral after 21 of 23 quotations in the briefs turned out to be fabricated. The court said no filing should contain citations the attorney hasn't personally read and verified.
- Couvrette v. Wisnovsky (D. Or., December 2025): sanctions and fees reported at roughly $110,000, and the case dismissed with prejudice.
Trackers that log these incidents list well over a thousand AI-hallucination cases worldwide as of mid-2026. This is not a rare glitch that happens to other people.
So the rule is absolute: every case, statute, citation, and substantive legal claim an AI produces is unverified until a licensed attorney confirms it against primary sources. Not "usually check." Always. Treat AI as a drafting and summarizing assistant that happens to be dangerously good at imitating legal research, and never as legal research itself. The verification habit, built from day one, is what separates firms that use AI well from the cautionary tales.
ChatGPT or Claude for a law firm?
Both handle the workflows above. For legal work specifically, Claude has two practical advantages worth knowing about. Its long context window means you can paste an entire contract or a full deposition transcript and question it in one pass, and its writing tends to need less editing to sound like a careful professional rather than a chirpy assistant. If your practice runs on shared folders of matter documents, Claude Cowork adds something ChatGPT's chat window doesn't: it works directly with files, so it can organize, rename, and index a matter folder rather than only discussing text you paste in.
Anthropic has also built legal-specific tooling on top of Cowork. A Legal plugin arrived on January 30, 2026, with commands for contract review against your playbook and NDA triage. On May 12, 2026, Anthropic followed with Claude for Legal: 12 practice-area plugins and 20+ connectors to tools firms already use, including Thomson Reuters CoCounsel Legal, Harvey, iManage, NetDocuments, and Docusign. Anthropic's own disclaimer applies to all of it: the output is a draft for attorney review, not legal advice.
That said, the tool matters less than the setup. A well-briefed ChatGPT beats an unbriefed Claude and vice versa. If you want the fuller head-to-head, see Claude vs ChatGPT for business. For a neutral look at the wider field of AI for law firms, including legal-specific platforms like CoCounsel, Lexis+ AI, Harvey, and Spellbook, see the best AI for lawyers. And for more ways to put AI to work across your practice, see how to use AI for your business.
Using Claude instead?
This page stays focused on ChatGPT. If your firm runs on Claude, or is deciding whether to switch, two guides go deeper:
- Claude for lawyers covers how to set Claude up for a legal practice.
- The Claude legal plugin walks through installing Anthropic's legal plugins in Cowork and what they do.
The hard lines below the tool choice stay the same: confidentiality, privilege, and verified citations.
A sane rollout for a small firm
- Write the standing brief — practice areas, jurisdictions, tone, fees, never-do rules. Ten minutes.
- Set a data rule everyone follows — which account, which settings, what never gets pasted.
- Start with the lowest-risk workflow — internal SOPs or client-update emails, not research, not filings.
- Build the verification habit immediately — every citation and legal claim is unconfirmed until checked.
- Add one workflow at a time — intake replies next, then summaries, then templates.
You don't need new legal software for any of this. It runs on a general assistant plus the documents and processes you already have.
Two related resources for the people your firm serves: if clients or prospects ask which AI tools are safe to use on their own, point them to legal AI for non-lawyers. If you want a structured starting prompt and clause checklist for common agreements, try the free AI contract generator. Like everything else here, its output is a first draft for attorney review.
Frequently asked questions
Can law firms use ChatGPT?
Yes, for the operational work around the practice of law: client emails, intake replies, document summaries, plain-English client explanations, engagement-letter first drafts, internal SOPs, and marketing content. Firms should not use it as a source of legal authority, and they need a clear policy on what client information may enter the tool, aligned with their bar's guidance on confidentiality.
Is ChatGPT confidential enough for client information?
Not by default. Any ChatGPT account is a third-party tool, and your duty of confidentiality governs what goes into it. ChatGPT Business and Enterprise don't use your data for training by default; on consumer accounts you have to check the setting yourself. Use a business-tier account, strip identifying client details from prompts where possible, and follow ABA Formal Opinion 512 and your state bar's guidance, which call for informed client consent before representation information goes into a self-learning tool.
Can ChatGPT do legal research?
No, not reliably. ChatGPT can fabricate case names, citations, and quotes from holdings that do not exist, and lawyers have been sanctioned for filing AI-generated citations without checking them, from Mata v. Avianca in 2023 to a $10,000 California sanction in 2025. Every case, statute, and legal claim it produces must be verified against primary sources by a licensed attorney before it is used in any filing or advice.
How are small law firms using AI?
Client communication and intake. Status-update emails and inquiry replies follow predictable patterns, carry low risk when reviewed before sending, and consume hours every week. Once the assistant knows your practice areas, tone, and screening questions, it drafts replies you lightly edit rather than write from scratch. Internal SOPs are a close second because they are not client-facing at all.
What is the best AI for law firms?
It depends on the work. General assistants like ChatGPT and Claude handle drafting, intake replies, summaries, and internal procedures well at a low monthly cost per seat. Legal research and citation checking belong in legal-specific tools grounded in real case law, such as CoCounsel, Lexis+ AI, or Westlaw, and larger firms add platforms like Harvey. Whatever you choose, use a business-tier account and verify every citation before it reaches a client or a court.
Is Claude or ChatGPT better for lawyers?
Both cover the core workflows. Claude's long context window handles full contracts and transcripts in one pass, its drafts tend to need less tone editing for professional work, and Claude Cowork can work directly with folders of matter documents rather than pasted text. The bigger factor is setup: an assistant briefed on your firm's context outperforms an unbriefed one regardless of brand.
The firm that briefs its AI wins
The difference between a firm that gets real hours back from AI and one that gave up after a week is rarely the tool. It's the ten-minute standing brief, the data rule, and the verification habit. Set those three things and the drafting, intake, and summary work starts compounding immediately.
The free Claude Cowork course teaches the whole approach with real prompts and workflows, built for operators and small firms rather than developers. It's a free download.