ChatGPT or a Legal AI Tool: Which Should a Small Firm Use?

Use a business ChatGPT plan for drafting, summarising and admin on material you supply, and a legal AI tool for anything that depends on authority: research, citations and jurisdiction-specific drafting. Many small firms need both, with ChatGPT Business for everyone at $20–$25 a seat and a research tool only for the lawyers who litigate or advise on contested law.

The dividing line is where the answer comes from. When you paste in a contract, a transcript or your own notes, a general model works from the text in front of it, and its output can be checked against that text. When you ask it what the law is, it answers from patterns in its training data and whatever the web search turns up, which is where invented cases and wrong-jurisdiction answers come from.

Follow me on Instagram@sagnikteaches

Where each one gets its answers

ChatGPT, Claude and Gemini are general models. They write fluently in legal register and follow drafting instructions well, but they have no built-in access to an authoritative, up-to-date database of cases and legislation, and no citator telling them whether a decision has been overturned. Web search helps with recent news and official pages; it doesn't replace a research database.

Connect on LinkedInSagnik Bhattacharya

Legal AI tools put a model on top of a legal database. CoCounsel sits on Thomson Reuters' Westlaw content, Lexis+ with Protégé on LexisNexis content, and Clio's Vincent and Clio Work on vLex, which Clio bought in 2025; Clio Work became available as a standalone product for solo and small firms in April 2026. When these tools cite a case, it's drawn from the database and linked. They still make mistakes (summarising a case wrongly, missing a later decision), so every authority still gets opened and read. But the starting point is real material rather than a plausible guess.

Subscribe on YouTube@codingliquids

A third kind sits inside practice-management software. Clio's built-in assistant, now called Manage AI (formerly Clio Duo), works on the firm's own matter data in Clio Manage: pulling dates out of court documents into the calendar, drafting client messages and preparing draft invoices, with each suggested action waiting for someone to approve it. It isn't a research tool, but for a firm already on Clio it can take some of the admin and drafting work you might otherwise give to ChatGPT, without copying matter details into a second system.

Side by side on what a small firm is judged on

How the legal side is priced, and when a seat is worth it, is covered in per-seat or pay-as-you-go legal AI pricing.

A research question put to a general model

This is the test worth running yourself before you decide. An associate pasted a court decision into ChatGPT and asked two things. The first request worked well:

Summarise the attached decision in 300 words for a partner.
Cover: the facts in two sentences, the issue, the court's answer,
and the reasoning in three bullet points. Quote the paragraph
numbers you rely on.

The summary was accurate, and because it quoted paragraph numbers the associate could check each point in a couple of minutes. The second request went wrong:

Which later cases have applied or distinguished this decision?
Give citations.

The reply (illustrative) listed three later decisions with names, years, courts and a one-line summary of each, in perfect citation format. The first was real and correctly described. The second was real but concerned a different point entirely. The third didn't exist; its name combined parts of two real cases. Nothing in the tone of the answer distinguished the invented case from the real ones. Courts in several countries have criticised and sanctioned lawyers for filing citations of exactly this kind, and a simple routine for catching them is set out in how to stop AI inventing case law.

The lesson isn't that ChatGPT is useless for legal work. It's that the first task (working from supplied text) is safe to check, and the second (recalling authority) isn't.

Tasks a general model handles well in a law office

• First drafts of routine letters from your notes and a precedent: client updates, chasers to the other side, letters of engagement for standard work.
• Summaries and chronologies of documents you supply: transcripts, bundles, correspondence. Summarising bundles and transcripts safely covers the checks.
• Plain-English explanations of a process for a client, written from your own points.
• Admin and marketing: meeting agendas, internal notes, website copy, job adverts.
• Comparing two versions of a clause or document and listing the differences.

Version comparison shows both the strength and the limit of working from supplied text. An associate pastes the firm's draft of a settlement agreement and the other side's mark-up, and asks for every change as a table: clause number, old wording, new wording, and whether the change favours our client, the other side or neither. The illustrative reply lists eleven changes correctly, including a confidentiality clause widened to cover "the terms and existence of this agreement" and a payment deadline moved from 14 to 28 days. What it misses is the twelfth: the other side edited the definition of "Termination Date" in the definitions clause, which shifts the meaning of three later clauses whose wording didn't change at all. The model compared clause by clause and reported only visible edits. The fix is a standing line in the prompt, "list any changed definitions first, then every clause that uses a changed defined term", and a run of the firm's ordinary comparison software alongside, because a redline is still the record of what changed.

Plain-English explanations for clients are the other quick win. A first draft built from a fee earner's bullet points might turn "the respondent has filed a response resisting the claim in full and the matter will be listed for a preliminary hearing" into "your former employer has formally disagreed with every part of your claim, so the next step is a short hearing where the judge sets out the timetable for the case." That's the right register. The check is the same as for any letter: does it say anything the notes didn't? In this case it's clean, but a draft that adds "this is common and shouldn't worry you" has crossed from explaining into reassurance the lawyer never gave.

Here's the kind of before-and-after that makes the case for a general plan (illustrative). The fee earner's notes after a call:

spoke to other side's lawyer. they want 2 more weeks for their
disclosure, say client ill. i said we'd take instructions. need
client ok by thurs. if no agreement we apply to court, costs risk
both ways, explain. hearing date unaffected for now

And the draft ChatGPT produced from those notes and the firm's update-letter precedent:

Dear [client],

I spoke to the other side's lawyer today. They have asked for a
further two weeks to provide their documents, explaining that their
client has been unwell.

We need your instructions on whether to agree. If we don't agree,
we could ask the court to decide. That would involve some cost, and
there is a risk the court orders us to pay part of the other side's
costs if it thinks the request was reasonable.

At the moment this does not affect the hearing date.

Could you let me know by Thursday how you would like to proceed?

One fix: the notes said "costs risk both ways", and the draft only described the downside. The fee earner added a sentence explaining that the other side could also be ordered to pay costs if their request were refused as unreasonable. Otherwise it went out as drafted, saving about 15 minutes. Drafting routines like this are covered in more depth in how small law firms use AI to draft letters.

Tasks that justify a legal tool

• Finding authority on a point, and checking it's still good law.
• Research memos that will be relied on in advice or submissions.
• Jurisdiction-specific drafting where the tool's templates reflect current law and practice notes.
• Checking a draft's citations against the database.

If a lawyer does these more than a few times a week, a research-grade seat usually pays for itself in time saved searching. If they do them a few times a month, the firm's existing research subscription and a careful routine may be enough.

Choosing for a three-lawyer firm

Take a firm of three lawyers and two support staff, doing employment work and some commercial litigation. One partner is in tribunal or court most weeks; the other two mostly advise and draft.

The firm chose the third option, with a three-month review. The deciding evidence was a two-week log: the litigating partner asked research questions most days, while the other two lawyers asked four between them. If the advisory partners' research use rises, a second seat can be added. Claude Team at the same seat price would have done the general job equally well; the choice between them came down to which the staff preferred after a week's trial on the same five tasks.

The general plan's cost is easy to justify on letters alone. If the five staff produce about 40 routine letters and emails a month and each saves 15 minutes, as the update letter above did, that's 10 hours a month for $100 of seats. The research seat has to be judged differently, on how much faster the litigating partner finds and checks authority, which is why the firm trialled it on her real matters rather than on a demo.

To stop the split blurring, the firm put a one-page routing card on the shared drive. Filled in, it reads like this (illustrative):

The last row came from an edge case in the firm's first month. An employment client bringing a whistleblowing claim sent documents naming colleagues and describing their health. Nothing in the policy strictly banned summarising them in ChatGPT Business, but the partner decided the client would reasonably expect that material to stay in the firm's document system. A flag on the matter, and a row on the card, meant nobody had to make that call alone at 6pm.

The card matters more than it looks. Most AI mistakes in small firms happen when someone uses the right tool for the wrong job because it was already open. A card that names the tool for each task, and the check that goes with it, turns a policy into a habit.

Setting up ChatGPT so it's fit for client work

1. Business plan only for client material. Consumer accounts, even paid ones, stay off client work. Business content isn't used for training by default.
2. Admin controls on. In ChatGPT Business, connectors are now called apps and are enabled by an admin. Switch on only the ones you've assessed, such as your file storage, and leave the rest off.
3. Shared projects by work type, not by client. A project for "employment settlement letters" holding precedents and instructions is reusable and easy to keep clean. OpenAI is retiring custom GPTs, which stop running on 11 December 2026; its migration route turns an existing GPT into a plugin. Don't build anything new on a GPT, and move any the firm already uses into a project before that date.
4. A standing instruction in every legal project: "Do not state the law or cite authority unless it appears in the documents provided. If asked, say that research must be done in the firm's research tool." Then test it. Ask, inside the project, "What's the time limit for bringing an employment claim?" An illustrative first reply: "I can't give legal research here; please use the firm's research tool. Generally, though, such claims must be brought within a few months." The second sentence is the instruction leaking. Tighten it ("do not give general statements of the law either, even hedged ones") and ask again until the reply stops at the first sentence.
5. A confidentiality line in engagement letters about using AI tools within the firm's systems, checked against your regulator's guidance. Whether solicitors can use ChatGPT without breaching confidentiality goes through the questions.

Ways a small firm gets this choice wrong

• Buying the research tool for everyone. Seats for staff who draft letters and never research are the commonest overspend.
• Using web search as research. A realistic mistake: a trainee asks ChatGPT, with search on, for the time limit to bring a particular claim. The answer is clear and sourced, from a government page in a different jurisdiction with a different limit. The trainee's supervisor catches it only because the number looked unfamiliar. Always name the jurisdiction, and never rely on the answer without a primary source.
• Trusting the legal tool's summaries unread. Linked citations reduce invention; they don't guarantee the summary of a case is right.
• Letting two tools drift into the same job. If the legal tool also drafts letters well, decide which one the firm uses for letters, so precedents and instructions live in one place.

Further reads

• AI for Small Law Firms: What to Automate First — Which of these tasks to tackle first in a small firm.
• Which Legal Tasks Should a Small Firm Never Hand to AI? — The work that shouldn't go to either kind of tool.
• AI Contract Review for Small Firms: What It Catches and Misses — What contract-review tools catch and miss in practice.
• How Small Law Firms Use AI to Answer Client Status Questions — Using AI to answer 'where is my matter?' questions.
• Does ChatGPT Train on Client Data? Business vs Free Plans — Training defaults on free and business ChatGPT plans.
• Can Lawyers Use AI Note Takers in Client Meetings? — The rules if you want AI in client meetings too.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A day-by-day 90-day AI plan for a six-person law firm: policy first, two scored pilots, shadow testing, an error log and a go/no-go decision.
• How Conveyancers Use AI to Cut Admin on Each Transaction — Stage-by-stage admin savings for a conveyancing file, a title-summary prompt with sample output, chaser wording and the tasks AI must never touch.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI ChatGPT Business pricing page and help pages on apps and Projects; Anthropic Claude Team pricing page; Thomson Reuters CoCounsel Legal plans page; Clio announcements on Clio Work and the vLex acquisition; Clio help article on Manage AI (the evolution of Clio Duo).