A few months ago, a senior partner forwarded me an email from a “legal AI” company. The pitch was familiar: $500 per month per attorney, AI-powered contract review, “trained on legal data,” demo available next Tuesday.
I was reading it on my second monitor while Claude Code was, on my first monitor, helping me build a script to extract every change-of-control clause from forty-seven NDAs in a folder. The script took me about twelve minutes to draft, plus another six refining the prompt. The output was a clean spreadsheet with clause text, deal name, and a short note flagging the unusual provisions.
The legal AI company’s demo would have done roughly the same thing. For $500 per month. Per attorney.
This is the article I wish I had read two years ago.
What Claude Code actually is, in lawyer terms
If you’ve heard of Claude—Anthropic’s chatbot, similar to ChatGPT—Claude Code is the same underlying model with one important difference. Instead of just chatting with you, it can read and write files on your computer. You give it instructions in plain English, and it does the work.
That sounds like a small thing. It’s not. The shift from “Claude can answer my questions” to “Claude can do things on my computer” is the difference between asking a junior associate for advice and asking them to draft a memo, redline a contract, or build you a tool.
For the curious: Claude Code runs in your terminal, costs $20 per month for the Pro tier (same price as the regular Claude subscription), and uses the same model as the chat interface. Anthropic doesn’t charge a “professional use” or “legal industry” surcharge. It just works.
What I use it for, in actual practice
Contract review at scale
The clearest win, by a margin. Last quarter I had to review fifty-three commercial agreements for a client running carve-out due diligence. Standard task: identify assignment, change-of-control, exclusivity, non-compete, and termination provisions, then flag the unusual ones.
The old workflow was four days of partner-associate ping-pong. The new workflow:
- Drop the PDFs in a folder.
- Tell Claude Code: “Read each contract in this folder. For each one, extract the assignment, change-of-control, exclusivity, non-compete, and termination clauses. Output a markdown table with deal name, clause type, clause text, and your analysis of whether the clause is standard or unusual. Flag anything you can’t find in a clear list at the bottom.”
- Wait twenty minutes.
- Spend two hours reviewing the output, correcting the three clauses Claude misread, and chasing the four it flagged as missing.
Total time: about three hours instead of four days. The accuracy isn’t 100%—I’d estimate Claude caught 92% of clauses correctly on the first pass. But that 92% was the boring 92%. The remaining 8% was where my judgment actually mattered.
I want to be specific about one thing: I did not upload these contracts to Claude’s web interface. I ran Claude Code locally, against files that never left my hard drive. I’ll write more about confidentiality in a separate post, but the short version is: if you’re a lawyer, you cannot ethically paste client documents into a chatbot. Claude Code’s local file workflow is one of the reasons it works for legal practice.
Memo drafting from research
When I have a research question—say, “what’s the trend in Delaware courts on appraisal arbitrage in 2025?”—I don’t ask Claude to research it. I do the research myself, in Westlaw or Lexis, and save the cases and articles to a folder. Then I tell Claude Code: “Here are eight cases and three commentary articles on appraisal arbitrage. Draft a five-page memo summarizing the current trend, highlighting the most recent Delaware Chancery decisions and any divergence from the historical line of authority.”
What I get back is a memo that sounds like me, because Claude has read my prior memos and learned the patterns. What I do not get is an “AI-generated brief” with fabricated citations. The cases are real because I selected them.
This is a small thing I want to underline: AI does not replace research. It accelerates writing about research you’ve already done. Lawyers who lose their jobs to AI will be the ones who tried to make AI do the research itself.
Building personal tools
This is where Claude Code separates from every “legal AI” SaaS I’ve tested.
A few weeks ago I noticed I kept getting asked the same question by junior associates: “what’s the standard structure of a representations and warranties insurance policy?” I had explained it eleven times in six months. So I asked Claude Code to build me a small interactive guide—a single HTML file—that walked through the structure with examples and let an associate click through to see redline samples.
It took forty minutes. The file lives on a shared drive. I haven’t been asked the question since.
Try doing that with Harvey or CoCounsel. You can’t. Those tools are designed to be used. They are not designed to let you build things. That is a fundamental difference, and it is the difference that “wrappers don’t beat the model—workflow does” is pointing at.
What it can’t do (yet)
I want to be honest about the limits, because the legal-tech industry’s worst sin is overpromising.
It cannot replace partner-level judgment. Claude Code can draft a competent memo. It cannot tell you whether to file the motion. The judgment about strategy, client relationship, and risk tolerance is still yours.
It cannot give you confidence in case law. I never let Claude Code generate citations. Hallucinated case names are a known problem with all LLMs, and the consequences for a lawyer are professional discipline, not a polite correction. I source citations myself, then ask Claude to weave them into prose.
It cannot draft truly novel structures. If you’re doing the third NDA of the day, Claude is faster than a junior associate. If you’re structuring a first-of-its-kind cross-border earnout with a contingent equity component, Claude will produce something that looks plausible but isn’t load-bearing. You still draft those.
It is not a substitute for malpractice insurance. Whatever Claude generates, you sign. If it’s wrong, you’re wrong. Treat its output the way you’d treat a first-year associate’s draft: assume it’s wrong until you’ve verified it.
So why are partners still paying $500/month for legal AI SaaS?
Three reasons, in my observation.
The first is risk aversion. A partner who buys Harvey can tell the firm’s IT and risk committee, “this is a vetted legal-tech vendor with SOC 2 compliance and a privacy policy reviewed by our outside counsel.” A partner who tells the firm “I use Claude Code on my laptop” is making a personal-responsibility argument, even though the underlying privacy posture is often better.
The second is time. Learning Claude Code takes maybe four hours to get to competence and another twenty to get to fluency. Twenty-four hours is a lot when you’re billing 2,200 a year. Most partners would rather pay $500 a month than spend a weekend learning a new tool.
The third is the demo effect. Legal AI vendors run excellent demos. They show you the thing you want to see, on the data they pre-loaded. The demo is real, but it’s not a substitute for the daily workflow. By the time you find out the tool doesn’t fit your workflow, you’ve signed a year-long contract.
I’m not saying don’t buy them. I’m saying: try Claude Code first, for $20, for a month. If after a month you can’t make it work, go ahead and pay $500. Most lawyers I’ve taught this to never make the upgrade.
How to actually start
If you’re a lawyer reading this and you want to try, here’s the minimum viable path.
- Subscribe to Claude Pro at claude.ai for $20/month.
- Install Claude Code from the official Anthropic documentation. It runs in your terminal—Mac, Windows, or Linux.
- Pick one task. Just one. The best starter task is “review this contract and tell me anything unusual.” Drop a contract you’ve already reviewed (so you know the right answer) and see what Claude finds.
- Spend a week comparing Claude’s output to what you’d have done yourself. Note what it gets right, what it misses, and where you have to correct it.
- After that week, decide.
That’s it. There is no certification, no implementation, no IT ticket. You don’t need partner approval, because the cost is a personal expense. You don’t need IT to install software because everything runs from your own terminal.
If after a month it isn’t saving you ten hours a week, I’m wrong about all of this. Email me at [email protected] and tell me what didn’t work. I’ll write about it.
FAQ
Is it ethical to use Claude Code for client work? The same rules apply as with any technology. You owe your clients confidentiality and competence. Don’t upload privileged documents to a chatbot. Use Claude Code locally with files on your own machine, review the output, and treat AI-generated work product the way you’d treat any other input you didn’t write yourself: verify before you sign.
What about Anthropic’s data policy? Claude’s API and Claude Code do not train on your data by default. The chat interface (claude.ai) has separate settings—turn off training in the privacy settings before using it for any sensitive work.
Do I need to know how to code? No. Claude Code lets you give instructions in plain English. The “code” part is what Claude writes for you when you ask it to do something. You read English; it writes whatever it needs to write.
Will my firm’s IT department block this? If you install on a personal device, no. If you install on a work device, ask first. Some firms have approved Claude for use; others have a blanket “no AI” policy that hasn’t been updated since GPT-3.
Should I tell my clients I’m using AI? The professional consensus is shifting toward yes, especially for novel uses. I tell clients I use AI tools in my workflow the same way I tell them I use Westlaw and Microsoft Word. The judgment, advice, and signature are mine.
This is the first review on this site. There will be more—including comparisons with Harvey, CoCounsel, Spellbook, and other “legal AI” tools I’ve actually paid for. If you’d like to be notified, the RSS feed is at the top of every page. If you have a tool you want me to test, email [email protected].