ChatGPT for lawyers: where it ships, where it sinks

Most lawyers asking "what's the best ChatGPT for me?" are buying a chat window when the firm needs an operating system.

This is the article we'd write for a law-firm partner about to spend money on AI who wants to spend it well. It covers the cost math, what fits which kind of firm, the ethics-rule constraints, the alternative tools, and when building your own system beats operating someone else's. The legal facts here are verified against the public record; the build experience comes from systems shipped inside live mid-market companies, and the pattern transfers to a firm cleanly.

It's not the chat window, it's the build

A law firm is a system. ChatGPT is one tool you could point at it. Most ChatGPT-for-lawyers articles read like a snake-oil prescription because they answer the wrong question. They ask whether ChatGPT is good for lawyers. The partner paying for it is asking something else: what do I buy or build, and who do I hire, so the firm runs on AI instead of around it?

Run the 1995 comparison. Companies that treated the internet as a thing to buy, a website and an email server, got a small bump. The ones that rebuilt the business around it got everything else. Same shape now, just faster. ChatGPT is the 1995 website. Everything the winners built after that is the intelligence layer.

Everyone can buy the chat window. The newer thing, and the part most coverage misses, is that AI now lets a firm build custom software around its own work, fast and cheap enough to finally make sense for a single practice. Five years ago, custom software for a 12-person firm was absurd. Too slow, too expensive, out of date before anyone used it. AI collapsed that cost, and because the software gets built and maintained with AI, it keeps improving instead of freezing the day it ships. That's the actual product. ChatGPT is just the raw material going into it.

Most ChatGPT advice for lawyers operates two layers below the real question. Prompt tips and use-case grids. Tool-handling content, useful the first month and irrelevant the second. The strategy question is upstream of the tool question, and the playbook for working that out doesn't change much between a software company and a law firm.

Where ChatGPT works (and where it gets you sanctioned)

Three things ChatGPT does well in a law firm. First-pass drafting: correspondence, internal memos, the shell of a demand letter. Document summarization on depositions, discovery files, long contracts. And the non-billable admin nobody wants, intake routing and scheduling logic and internal training material. The failure mode on all three is a wasted hour of attorney time. None of it reaches malpractice.

It fails at three others, sometimes catastrophically. It invents cases, so no citation it produces can be trusted on sight. The public version stores prompts, which becomes a confidentiality problem the moment client data goes in. And there's no audit trail, so you can't reconstruct who reviewed what. Those are failures of the raw chat window. Built properly, a system handles all three: citations grounded against real case law, isolated infrastructure the public model never touches, a log of who reviewed what and when.

The bright line between "ChatGPT helped me draft this" and "ChatGPT got me sanctioned" runs along citation accuracy. In 2023, a New York lawyer in Mata v. Avianca cited fabricated cases ChatGPT had hallucinated. The sanction was $5,000, plus an order to notify his client and each real judge whose name had been attached to a fabricated opinion. Every legal-tech vendor has told that story a hundred times, usually to sell something. The honest read is that the lawyer skipped a verification a paralegal would have done in 1995, and nothing in the workflow caught it.

ABA Formal Opinion 512 (2024) codified the obvious. Under Model Rules 1.1, 1.6, and 5.1/5.3: know what your AI tool can and can't do, keep client data out of tools with unclear retention, verify outputs the way you'd verify a paralegal's. Every legal-tech article quotes it. Almost none say what compliance actually looks like at the workflow level, probably because the answer is boring. You need a review step on every AI workflow before anything leaves the building, client data isolated from any public model, and a documented supervision chain. Same shape as supervising a junior associate, with a tool where the person used to be.

A February 2026 ruling made the confidentiality risk concrete, and it lands harder than most lawyers assume. In United States v. Heppner, the first federal case to address it head-on, the Southern District of New York held that materials a defendant created by feeding his case details into a public consumer AI tool were not protected by attorney-client privilege or the work-product doctrine. Roughly 31 of those AI-generated documents went to prosecutors. The tool there was Anthropic's Claude, but the reasoning covers any public model, ChatGPT included: you have no attorney-client relationship with an AI platform, and the platform's own terms let it use and disclose your inputs, so there is no reasonable expectation of confidentiality. The court left one door open. Had counsel directed the AI use, privilege might have held. Which is the whole argument for a controlled system: a public chat window is a third party sitting in the room, and the same work done inside a counsel-directed setup stays protected.

Confidentiality is one axis a controlled system fixes. Accuracy is the other, and it's where the overselling happens. No AI system is accurate on day one, whatever the demo showed you. Realistic curve: around 60% on day one, 85% by month three once the team has fed back corrections, past 95% on the workflows people use daily. Judge it on the 95%. The day-one number is sales theater.

A custom build is software somebody actively maintains, so it keeps getting better at exactly those things while the public tool stays where it is. Those limitations are the work a build exists to do.

The three options, and what they really cost

The headline prices are easy to find. What's harder is the work each option puts on the firm, and which one fits how you actually run.

The three options most firms compare:

option	direct cost	hidden cost	best fit
ChatGPT Business (formerly Team)	about $25/seat/mo (~$6K/yr for a 20-person firm)	You enforce the guardrails by hand: a review step on every output, citation checks, keeping client data out of the public model	You want the raw tool for drafting, summarization, and admin, and you can run your own review step. The entry point at any size
Legal-specific SaaS (CoCounsel, Spellbook, Lexis+ AI, Paxton, MyCase IQ)	roughly $50–$850/seat/mo (~$12K–$200K/yr for a 20-person firm)	Vendor lock-in, and you live with their default workflow even where it's two clicks from what you actually do	A mainstream workflow (legal research, contract review) that a legal-tuned tool already fits well
Custom intelligence layer (fractional CAIO retainer, builds included)	~$15K–$25K/mo, scaled to scope (a solo build sits at the low end)	Onboarding time (a few weeks to the first live build), and senior-partner attention to shape what gets built	You want an AI hire that builds systems around your own data, maintains them, and grows them over time, instead of operating a tool yourself. Fits any firm, solo or large, even one whose work an off-the-shelf tool could cover

Most firms read those price tiers as a budget question. It's a fit question. Every firm has repetitive, high-value work, so that tells you nothing about which tier to pick. What you want done with the work does.

Pay $25K/mo for a custom build and then refuse to give it partner attention, and you've wasted the money. The reverse costs more: a 30-attorney firm hand-building workflows on top of raw ChatGPT, when what it actually wants is those systems built and maintained for it, is spending senior-partner hours on work a hire should own. Headcount doesn't settle any of this. The question is whether you want an AI hire building and maintaining custom systems around your data, or a static tool you operate yourself.

When ChatGPT is enough, when a legal tool fits, when you build

Reach for ChatGPT Business (formerly Team) when you want the raw tool for drafting and summarizing and the admin nobody bills for, and you're fine running your own review step. Get fluent here first, whatever your size.

Legal-specific SaaS (CoCounsel for research, Spellbook for contract review, Paxton for litigation prep, Lexis+ AI for citations) makes sense when your workflow is genuinely standard and a tuned tool already fits it. The models are trained on legal corpora, which lowers the hallucination rate. But the confidentiality and audit controls these tools lead with aren't unique to them, a custom build delivers the same, and SaaS adds two costs of its own: lock-in, and a workflow shaped to the vendor instead of to you. It's the right call when your work really is standard. The moment it isn't, you're bending the firm around someone else's product.

Build a custom intelligence layer when you want an AI hire who builds systems around your own data and keeps rebuilding them as the models and the firm move, rather than operating a fixed tool yourself. Every firm has work worth automating, so that's no signal at all. Neither is whether an off-the-shelf tool could technically cover the work. What matters is what you want done with it.

The case is sharpest when the work is specific to you: the same prompt typed by four different people this week, outputs that need the same manual fix every time, a SaaS default flow two clicks off from what you actually do, client-data rules quietly stopping the team from leaning in. But it holds even where a legal tool would fit fine, because a build gives you that workflow without the lock-in, keeps improving in the shape of your firm, and folds it into one system instead of one more subscription. Size has nothing to do with qualifying.

The custom layer is connective tissue around ChatGPT (or Claude, or whichever foundation model is ahead this quarter): firm-specific intake routing, drafting agents trained on your templates, research agents that respect your jurisdiction, document QC, conflict-check automation, all behind one UI with role-based permissions. The model underneath is a commodity. What you own is everything wrapped around it.

What a custom legal intelligence layer actually looks like

Usually it's four or five firm-specific agents wired into one system. For a litigation-heavy boutique, that looks like:

• An intake agent that classifies inbound leads by case type, runs an initial conflict check, generates the first-draft engagement letter, and routes the file to the right paralegal.
• A drafting agent trained on the firm's templates and prior winning briefs, producing first-pass drafts a partner can edit in 20 minutes instead of starting from a blank page.
• A research agent that searches Westlaw or Lexis (depending on the firm's existing subscription) and produces citation-verified memos with the case quotes inline.
• A discovery summarization agent for long document sets (depositions, contract collections, regulatory filings) that produces a paragraph summary per document with timestamps and key-passage links.
• A secure document-handling agent for any workflow that touches privileged client data, running on a hardened instance with no public-model exposure (NVIDIA's NemoClaw or similar).

Five components, one authentication layer, one data layer, one UI. A paralegal texts the intake agent like a colleague. A partner asks the research agent for a memo and gets one back with the citations already checked. Swap the underlying model when a better one ships; the system around it stays.

That system is what an AI consultant who actually ships builds hands over. The same pattern is already running in wealth management, another compliance-bound vertical: start with the one workflow that costs the most time, build it around the practice, expand from there. Tool recommendations are free and everywhere. The custom-layer framing is the operator's answer to "should we use ChatGPT?", and nobody selling a SaaS subscription has any reason to give it to you.

How to test this before you commit

Pick one workflow and run a 90-day pilot. It's a paid quarter, not a free trial, and that's the point: a quarter instead of a year, a working build at the end, and an easy exit if the value doesn't show.

The workflow should be specific. "AI for our firm" is too vague to act on. Examples that work: "draft demand letters from intake forms in our personal injury practice," "summarize incoming discovery PDFs into one-page case-fact briefs," "auto-generate engagement letters from the intake-conflict-check output," "produce client-update emails from the case-management activity log." Each is one workflow, one output, one team that uses it. The first working version comes together fast. Rough, but enough that a partner can see the shape and judge whether the value is real.

Judge it on trajectory. A demo on its own proves very little. Full team adoption runs one to two quarters, so nobody will be living on the tool by day 90, and that's expected. What you're watching for by then is smaller: output people trust without re-checking every line, a workflow that matches how the firm actually works, and at least a few of the people who touch it reaching for it before they reach for the old way.

Don't sign a monthly retainer off one good demo. Sign it once that signal shows up. When you do, the first month of the retainer goes straight into the next build, because the pilot already did the discovery. If the signal never shows, you walk, having spent three months on it. A fast-moving firm on a focused workflow can see payoff inside that first month. A larger or slower one should plan for up to two quarters, because the pace is set by how fast the team changes the way it works, and teams are slow.

Fire me anytime.

No SaaS vendor offers the contract shape most working partners actually need: monthly, no lock-in, where the person has to keep delivering enough value that you don't want to fire them. That's the fractional CAIO retainer. The first build is the down payment. What you're buying is someone who keeps the workflows you already have current as the models and tools underneath them shift, and who ships new ones as the firm changes: a new practice area, a new bottleneck, a tool worth wiring in. AI systems drift if nobody tends them. So the standing job is keeping the existing builds sharp and finding the next workflow worth automating, month after month. The price doesn't move when the scope does. Either side can end it anytime.

If we're a fit, the first step is the 90-day pilot. Pick the workflow, build it, see if the value is real. If it is, the retainer is the obvious next move and there's nothing left to sell you. If it isn't, we shake hands and you've spent a quarter instead of a year.

For any firm ready to put its work into maintained, custom-built systems, the math is straightforward. A year of fractional CAIO retainer costs less than a single underperforming associate. One ChatGPT-hallucinated citation that reaches a filed brief costs more than two months of it. And the cost of not having an AI system in 2027, when half your peers do, is harder to put a number on. It isn't zero.

Start with one workflow. Run a 90-day pilot. Fire me anytime. Book a 30-minute call and we'll walk through one of the live builds.

Frequently asked questions

Is there a ChatGPT for lawyers? No single "ChatGPT for lawyers" product exists. You have three real choices. ChatGPT itself, general-purpose, about $25/seat/mo on the Business plan (formerly called Team). A tier of legal-specific AI tools that wrap foundation models with legal-tuned training and confidentiality controls, at roughly $50–$850/seat/mo (CoCounsel, Spellbook, Lexis+ AI, Paxton, MyCase IQ). Or a custom intelligence layer built for the firm, using whichever foundation model suits each workflow. Which one is right depends on whether you want the raw tool, a vendor's pre-built tool for a standard workflow, or an AI hire building and maintaining custom systems around your own data. Marketing budget is the worst possible tiebreaker.

How much does ChatGPT for lawyers cost? ChatGPT Business (formerly Team) is about $25/seat/mo. ChatGPT Enterprise (with stricter data controls, longer context, admin controls) is roughly $60/seat/mo at typical firm sizes. The cost that's easy to miss isn't the subscription, it's the review step every output needs before it leaves the building. Build that into the workflow and the tool pays for itself. Legal-specific SaaS runs roughly $50–$850/seat/mo, and a custom intelligence layer (a fractional CAIO retainer with builds included) runs roughly $15K–$25K/mo, scaled to scope.

What is the 80/20 rule for lawyers? The 80/20 rule for AI in a law firm: 80% of the value comes from automating the right 20% of repetitive workflows. Drafting, summarization, and intake almost always sit in the 20%. Strategy, judgment, and client relationships live in the other 80%, which should stay human. The mistake most firms make is automating the wrong 20%: generic prompts for legal research instead of firm-specific automation of a specific repeated task.