First Drafts, Not Final Answers: Where AI Fits in Legal Work

It’s 9:15 on a Tuesday night. A third-year associate is on page 190 of a 280-page deposition transcript, writing a summary the partner wants by morning. Down the hall, a paralegal is turning a document production into a chronology, one email at a time. The partner is drafting the fourth demand letter this month that looks a lot like the other three.

None of that work is the reason anybody went to law school. All of it has to be right.

That’s where generative AI is useful in a law firm, and also where it can embarrass one. The tools are good at a first pass. They are also capable of sounding certain about things that aren’t true.

A note on where I’m coming from. I’m not a lawyer, and nothing here is legal or ethics advice. I spent almost 18 years running a digital agency and now spend my days using AI to fix real bottlenecks inside businesses. What I understand is how work moves through an office and where it piles up. In firms, it piles up in first drafts.

Where the hours actually go

Look at a month of time entries by task and a pattern shows up. A big share of hours goes to work that’s mostly reading and restating.

• Deposition summaries. Hours per transcript, often written by the most expensive person available that evening.
• First-pass document review. Sorting a production into what matters, what’s privileged and what’s noise.
• Routine drafts. Demand letters, engagement letters, standard motions, discovery requests, simple agreements built from the firm’s own forms.
• Chronologies and comparisons. Building a timeline from emails, or listing what changed between two versions of a contract.
• Research starting points. Figuring out where to look before the real research begins.

Every one of those ends with a lawyer’s judgment. Most of the clock, though, goes to the part before the judgment.

What the ethics opinion asks lawyers to weigh

In July 2024 the ABA issued Formal Opinion 512 on generative AI tools. It’s the most useful single document I’ve read on this, and it’s worth a lawyer’s own read. In brief, it says lawyers using these tools must consider their duties of competence, protecting client information, communicating with clients, supervising the people and services they rely on, advancing only meritorious claims, candor toward the tribunal and charging reasonable fees.

A few points stood out to me as a non-lawyer:

• Lawyers need an appropriate degree of independent review of what the tool produces, and how much depends on the tool and the task.
• Before putting information about a representation into many self-learning tools, the client’s informed consent is required, and general boilerplate in an engagement letter isn’t enough.
• Lawyers billing by the hour bill their actual time, and a lawyer can’t charge a client for learning a tool the lawyer will use regularly.
• The opinion expects state and local bar ethics committees to keep issuing their own guidance.

It’s an ABA opinion. Each firm practices under its own state’s rules, and those can differ. How the duties apply to a particular firm is for the firm and its state bar to decide.

What AI-assisted legal work looks like

Deposition summaries with page and line references

The tool drafts a summary organized by topic, and every point carries the page and line it came from. The lawyer reads the summary, spot-checks the references, and fixes what’s off. If a point has no reference, it doesn’t stay in.

First-pass review the lawyer can audit

The tool sorts documents into likely relevant, likely privileged and likely irrelevant, and says why for each. Lawyers review the privileged pile and sample the rest. Sorting is a first pass. Production decisions stay with the lawyers.

First drafts built from the firm’s own work

Routine letters and documents start from the firm’s own templates and past work product, not from whatever the internet thinks a demand letter looks like. The lawyer edits, decides and signs.

Research leads, never citations

The tool can suggest where to look and what questions to ask. It cannot be trusted to supply authority. Lawyers have been sanctioned for filings that cited cases that don’t exist, and Opinion 512 itself mentions citations to nonexistent opinions among the problems that have come up. Every case, statute and quote gets pulled and read by a lawyer before it’s cited.

Run the numbers

Here’s the formula in words. Monthly hours spent on first drafts and summaries, times the share AI can draft usably, minus the time to review and correct, gives the hours that can go somewhere else. Multiply by what an hour of redirected time is worth to the firm.

Say, for the sake of the math, a made-up eight-lawyer firm spends 160 hours a month across lawyers and paralegals on deposition summaries, routine drafts and chronologies. Say AI can produce a usable first draft for half of that work, the 80 hours. Review and correction take, in this made-up case, 30 percent of the original time, or 24 hours. That leaves 56 hours a month.

What those 56 hours are worth depends on what the firm does with them. Taking on more matters with the same people. Faster turnaround on the work clients complain about. Associates who aren’t writing summaries at 9 PM. If the firm values a redirected hour at a made-up $150, that’s $8,400 a month in capacity.

How any of that shows up in fees is the firm’s decision. Hourly billing, flat fees and AI tool costs each raise questions the opinion addresses directly, so read its fee section before setting a policy.

To run it for real, pull three numbers:

1. Hours per month on summaries, first drafts and chronologies, from time records by task.
2. The share of that work that’s routine enough for a first draft, by your own lawyers’ estimate.
3. What an hour of redirected lawyer or paralegal time is worth to the firm, whether that’s new matters, turnaround or fewer late nights.

What has to be in place first

• A written firm policy. Which tools are approved, for which tasks, and who reviews what. Lawyers and staff should know the answer before they open a tool.
• Tools whose terms someone has read. Opinion 512 says lawyers should understand the terms of use and privacy policy of the tools they use, or consult someone who has. Consumer chatbots with unclear data handling don’t belong near client files.
• Templates and past work product, organized. First drafts are only as good as the examples they start from.
• A verification step for each task. Written down. What gets spot-checked, what gets read in full, who signs off.
• A decision about client communication. When the firm tells clients, when it asks for consent, and how. That’s the firm’s decision under its state’s rules.

Guardrails worth keeping

• Every citation is read by a lawyer. No exceptions, no “it looked right.”
• Client information stays in business-grade tools. Set up so data isn’t used to train public models, with access limits and logs.
• Supervision is real. Partners own the work associates and staff produce with these tools, the same as any other work.
• Check court rules. Some courts require lawyers to disclose AI use in filings. Know the standing orders where you file.
• The lawyer signs, the lawyer decides. AI drafts. It doesn’t advise anyone.

How you’ll know it’s working

• Hours per deposition summary, before and after.
• Turnaround time on routine drafts.
• Errors caught at review, by task, so you know where the tool needs more checking.
• Citations in filings that a lawyer hasn’t read. This number should be zero, always.
• Matters handled per lawyer per month.

Where I’d start

Start with deposition summaries on closed matters. Your lawyers already wrote summaries of those transcripts, so you can compare the tool’s draft to work you trust, line by line, before using it on anything live. The opinion describes this same idea: test a tool on a smaller set you’ve reviewed yourself before relying on it for more.

If this sounds like your week, send me a note through the contact page.