Short answer
Picking a model — Claude, GPT, or the next one — is the easy part and the wrong headline. For a law firm, the risk isn't which system drafts the brief; it's whether a human reviews and approves before anything reaches a client or a court. Govern the workflow first. The model is interchangeable; the audit trail is not.
The legal press is full of "AI-native firm" profiles this year: small practices that route intake, research, and drafting through a single model and report faster turnaround. The exciting version is a story about a model. The useful version is a story about governance — the rules that decide what software is allowed to do on its own, and what a person must sign off on first.
Here's the distinction that matters for an operator.
Why is "which model" the wrong question for a law firm?
Models are becoming commodities; your obligations aren't. Whichever system you pick can draft a motion, summarize a deposition, or answer an intake call. None of them can be held responsible for what goes out under your name. So the real question isn't "which model," it's "what is allowed to happen without a human in the loop."
Two things are true at once:
- Generation is cheap. Any current model produces a plausible first draft in seconds. That's genuinely useful.
- Judgment is the job. A plausible draft that cites a case that doesn't exist is worse than no draft, because it looks finished.
The moment generation is free, the scarce thing is review. A firm that builds its practice around "which model is smartest" has optimized the abundant input and ignored the scarce one. The firms that hold up build around a simpler rule: the software drafts, a lawyer approves, and every step lands on a record.
What actually goes wrong when firms skip the review step?
Fabricated citations reach real courts, and judges have responded with money and public orders. This isn't hypothetical. Hallucination — when a model states a confident, well-formatted fact that is simply false — has already produced sanctions when lawyers filed AI-generated work without checking it.
The reference case is Mata v. Avianca. In June 2023, U.S. District Judge P. Kevin Castel of the Southern District of New York sanctioned two attorneys and their firm $5,000 after they submitted a brief citing cases that ChatGPT invented, then stood by them when challenged. The order is public and specific: the problem wasn't using a tool, it was submitting its output as fact without verifying it.
Since then, tracking has gotten more systematic. Legal researcher Damien Charlotin maintains a public database of court decisions involving AI-hallucinated citations; it has logged hundreds of cases worldwide, and the count keeps climbing as more filings get caught. The lesson is consistent across all of them: the failure point is never generation. It's the missing approval step.
So what does a "governed" AI setup actually look like?
It looks like a draft that waits. The system does the work — pulls the record, writes the letter, stages the filing — and then stops, holding the action for a person to approve, edit, or kill. Nothing is sent, filed, or billed automatically. Vantrow calls this propose, never commit.
In practice, a governed setup has three properties:
- Staged actions, not automatic ones. The system prepares outbound email, intake summaries, or draft motions and holds them. A human releases them.
- A human approver by name. Someone specific signs off, so accountability is a person, not a setting.
- An audit trail. Every proposal, edit, and approval is logged — what the system suggested, who changed it, who sent it, and when.
That last property is what turns "we use AI" into something you can defend to a client, a partner, or a bar committee. Adoption is real: the American Bar Association's 2024 Legal Technology Survey Report found roughly 30% of law firms reported using AI tools, up sharply year over year — which means the review question is now a here-and-now operating problem, not a future one.
Does governance slow the firm down?
Not the way people fear. The slow part of legal work was never writing the first draft — it was the checking, the sign-off, and the responsibility. Governance doesn't add that step; it makes the step that already exists explicit and traceable.
A governed desk changes where the lawyer spends attention:
- Less time on blank-page drafting and manual record-pulling.
- Same time on review — but now against a draft instead of a void.
- Zero time reconstructing "who approved this and when," because it's logged.
The "AI-native firm" that reports speed usually isn't fast because the model is smart. It's fast because the low-value keystrokes are gone and the high-value judgment is concentrated where it belongs.
What should an operator take from the "Claude-native" headlines?
Treat the model as a replaceable part and the governance as the permanent one. If a competitor's setup would break the day their model of choice changed pricing, deprecated, or hallucinated into a filing, it was never a strategy — it was a bet. Build the rule first: draft freely, approve deliberately, log everything.
Concretely, before you route any real client work through a model, get answers to these:
- What can the system do without a human? (Ideally: nothing final.)
- Who approves, by name?
- Where is the record of what was proposed versus what was sent?
If those answers are clear, the model underneath genuinely doesn't matter much. If they're not, no model is smart enough to cover for it.
FAQ
Is it a bad idea to standardize on one AI model like Claude?
Standardizing on a model is fine for consistency, but it shouldn't be your strategy. Models change pricing, get deprecated, and hallucinate. Build your workflow so any model drafts and a human approves. Then swapping the model is a settings change, not a crisis.
What is "propose, never commit"?
It's the rule that software stages actions instead of taking them. The system drafts the email, summary, or filing and holds it for a person to approve, edit, or reject. Nothing is sent or filed automatically, and every step is logged on an audit trail.
Did a court really sanction lawyers for AI-generated citations?
Yes. In Mata v. Avianca (June 2023), U.S. District Judge P. Kevin Castel in the Southern District of New York sanctioned two attorneys and their firm $5,000 for submitting a brief citing cases ChatGPT fabricated. It remains the reference case, and researcher Damien Charlotin's public database now tracks hundreds of similar decisions.
How many law firms actually use AI tools?
The American Bar Association's 2024 Legal Technology Survey Report found roughly 30% of law firms reported using AI tools, a sharp year-over-year increase. That makes review and approval an immediate operating question, not a future one.
What's the fastest way to check if my firm's AI setup is safe?
Ask three questions: What can the system do without a human? Who approves, by name? Where is the record of what was proposed versus sent? If those answers are clear, the model underneath matters far less than the governance around it.