ChatGPT is worth $20 a month to a lawyer for a handful of specific jobs: custom GPTs for team knowledge, voice mode, reasoning models for analysis, image generation, and deep research. For legal drafting itself, Claude beats it.
Let me be honest about where ChatGPT sits in my stack, because most articles on this topic are written by people selling you something.
My daily driver for legal drafting is Claude. Demand letters, client emails, document summaries: Claude writes them in a voice that doesn’t need heavy editing. But ChatGPT never left my toolkit, and it never will, because it does five or six specific jobs better than anything else. Most serious AI users I know run both. Two subscriptions, $20 each, about $40/month total. That’s the cost of doing this properly.
Here’s what ChatGPT is actually good at, and the two things it should never do for you.
| Task | Better tool |
|---|---|
| Legal drafting (letters, emails, briefs) | Claude |
| Custom GPTs for firm knowledge | ChatGPT |
| Voice dictation and thinking out loud | ChatGPT |
| Long documents (200+ pages) | Claude |
| Deep research on opposing parties | ChatGPT |
| Case citations | Neither, without manual verification |
Teach it who you are, once
Before anything else, set your custom instructions. This is the settings box where you tell ChatGPT your role, your firm, and how you want answers formatted. Set it once and every future conversation starts from context instead of zero.
ChatGPT also has memory, which picks up details across conversations over time. Between the two, it stops treating you like a random stranger and starts acting like it works at your firm. Skipping this step is why most lawyers get generic output. A sample block to steal is at the bottom of this page.
Custom GPTs: your firm’s knowledge, packaged
This is the feature Claude has no real equivalent for, and it’s the one I use hardest.
A custom GPT is ChatGPT preloaded with your documents and instructions. I built a Law Firm Intake Specialist GPT for My Legal Academy that lawyers use to qualify intake leads: it knows what a good case looks like, what disqualifies one, and what questions to ask in what order. Intake staff who have never read a statute get consistent, correct qualification guidance. (Automating the intake conversation itself, phones and follow-up included, is a different layer: see AI intake for law firms.)
You can build one for your firm in about 30 minutes with zero code. Upload your FAQ, your fee structure, your intake criteria. Write two paragraphs of instructions. Done. Now every person at your firm asks the same brain the same questions and gets the same answers, instead of five staff members improvising five different versions of your fee explanation.
Voice mode: the commute is billable thinking time now
ChatGPT’s voice mode is a real conversation, not dictation software. I know lawyers who dictate case notes while driving between hearings and arrive with a structured summary waiting in their account. Others talk through an argument out loud on the way to court: state the position, let ChatGPT push back, sharpen the response.
Talking through a problem is how most litigators think anyway. Now the wall talks back.
Reasoning models for analysis, not drafting
ChatGPT’s reasoning models (the thinking modes, the o-series) work through problems step by step instead of answering instantly. Slower, and worth it for analysis tasks.
Two places they earn their keep: comparing settlement options, where you feed in three offers with different structures and tax treatments and get a systematic breakdown of tradeoffs, and timeline reconstruction, where you paste a messy pile of dates from records and correspondence and get back a clean chronology with the gaps flagged. At my lemon law firm, repair timelines decide cases. A model that catches “the third repair attempt predates the notice letter” is doing real work.
For drafting prose, skip the reasoning models. Regular mode is faster and the writing is the same or better.
Image generation for firm marketing
ChatGPT generates decent social graphics and ad concepts. Quote cards, post visuals, rough layouts for your designer to react to instead of starting cold. My marketing team uses AI images as concept drafts constantly.
One caveat, and it matters: keep it tasteful and keep it true. No fake courtroom victories, no fabricated settlement numbers, nothing that implies results you didn’t get. Bar advertising rules did not get suspended because the image came from a machine.
Deep research on the other side
Deep research mode browses dozens of sources for 10 to 20 minutes and returns a cited report. Before a negotiation with a manufacturer, I can get a rundown of their recent recalls, pending litigation, and public financial pressure. Same tool works on opposing parties, expert witnesses, or an industry you’re touching for the first time.
It replaces the first hour of background research, not the lawyering. But that first hour used to cost an associate’s afternoon. For case law itself, use AI legal research tools built on real databases; deep research is for facts, not citations.
The two things it must never do
First: case citations. ChatGPT invents cases. Confidently, with realistic reporter numbers and convincing quotes. Lawyers have been sanctioned in federal court for filing briefs built on citations that did not exist. This is not a rare glitch, it’s how the technology works. Use ChatGPT to find leads, then verify every single citation in Westlaw or Lexis before it appears in anything with your bar number on it.
Second: final documents without review. Everything AI produces is a draft until a lawyer reads it. Your name goes on the filing, not OpenAI’s.
Getting started today
Get the paid account at chatgpt.com. The free tier trains on your data and locks you out of the features this whole article is about.
Then paste something like this into custom instructions, edited for your practice:
I am a lemon law attorney at a consumer protection firm handling cases in all 50 states. Assume I know the law; skip basic legal explanations. Be direct and concise. When drafting, write in plain English at a level a client can understand. Flag anything that needs verification, especially case citations and statutory deadlines. Never present a case citation as verified.
Then run one real task before the day ends. Dictate your case notes through voice mode after your next client call, or start building an intake GPT from your existing FAQ. Thirty minutes, either one.
If you want the prompts that make daily use pay off, the prompt library is next. And if you landed here without reading the adoption order, start with the roadmap. The order matters more than the tool.