Can Law Firms Use AI for Client Memos Responsibly? A Guide to Disclosure and Review

A partner asks an associate for a memo on a new fact pattern by Friday. The associate drafts it with help from a generative AI tool, polishes it, and sends it up. Nobody’s quite sure whether that’s fine, whether the client needs to know, or who’s on the hook if a cited case turns out not to exist. That uncertainty is the real problem most firms are wrestling with right now, and it’s solvable.

This guide walks through what law firm AI writing disclosure actually requires, where the real obligations sit, and a review workflow that keeps you on the right side of your duties without grinding the work to a halt.

Key takeaways

• Most bar guidance doesn’t ban AI for drafting client memos, but it holds the lawyer fully responsible for the output, citations, and confidentiality.
• Disclosure isn’t one rule. It depends on your engagement letter, your billing, and whether AI use is material to the representation or the fee.
• The non-negotiable step is human review: an attorney must verify every citation, fact, and legal conclusion before a memo leaves the building.
• Running a draft through an AI detector tells you a memo reads machine-like, but that’s a readability signal, not proof of anything.
• A written internal AI policy beats case-by-case guessing. Spell out which tools are allowed, what data can go in them, and who signs off.

Is it allowed at all?

Short answer: yes, with conditions. The American Bar Association’s Formal Opinion 512, issued in July 2024, doesn’t prohibit generative AI. It maps existing duties onto the technology instead. The same rules you already follow, competence, confidentiality, communication, candor, and reasonable fees, still govern. AI doesn’t get its own carve-out, and it doesn’t excuse a single one of them.

The practical takeaway is that AI is a tool, not a delegate. You can use it to draft a client memo the way you’d use a form bank, a research platform, or a contract assembly system. What you can’t do is treat its output as authoritative or unreviewed. The duty of competence (Model Rule 1.1) now includes understanding the upsides and the risks of the technology well enough to use it safely.

The competence problem in one sentence

If you can’t explain how a tool might get something wrong, you’re not yet competent to rely on it for client work.

When do I actually have to disclose AI use?

Here’s where firms over-correct in both directions. Some assume every AI-assisted document needs a disclaimer. Others assume disclosure never applies. Neither is right.

There’s generally no blanket duty to announce that you used a general-purpose AI tool, just as you wouldn’t tell a client you used a spell-checker or searched Westlaw. Disclosure becomes a real obligation in a few specific situations:

• It’s in your engagement agreement. If your retainer says you’ll inform the client about AI use, that’s now a contractual promise, and the duty of communication (Rule 1.4) backs it.
• AI use is material to the representation. If a client has expressly said they don’t want AI involved, or the matter is sensitive enough that your method affects their decisions, they need to know.
• It affects the bill. This is the one firms miss. If AI dramatically cut the time a task took, you can’t bill the old number of hours. Charging for work you didn’t do crosses into the reasonable-fee rule (Rule 1.5).

A clean way to handle the disclosure question is up front. Add a short paragraph to your standard engagement letter describing your responsible use of technology, including AI, and the client has notice from day one. That turns a recurring judgment call into a one-time drafting decision. Just don’t expect that paragraph to do double duty as confidentiality consent. Opinion 512 says boilerplate engagement-letter language isn’t enough when you’re putting a client’s confidential facts into a tool, which is a separate conversation we’ll get to below.

Mini-scenario: the fee question

Imagine a memo that historically took your associate six billable hours. With AI assistance and a careful review, it now takes two. You bill two hours, not six. And the tool’s subscription? Opinion 512 treats the cost of accessing a general AI product as overhead, the same bucket as your Westlaw seat or your word processor, which usually shouldn’t be billed back to the client at all. What you definitely can’t do is keep billing six and pocket the efficiency. That’s the move that turns a productivity win into an ethics complaint.

What confidentiality rules apply when I paste client facts in?

Confidentiality (Rule 1.6) is the duty most easily broken by accident here. Many consumer AI tools reserve the right to use your inputs to train future models. Drop a client’s facts into a tool like that and you may have handed protected information to a third party.

Before any client data goes into a tool, do three things:

1. Read the data-handling terms. Look for whether inputs are used for training and how long they’re retained.
2. Prefer enterprise or zero-retention setups. Business-tier and API configurations often contractually exclude your data from training. That’s the version a firm should be on.
3. Decide what’s too sensitive to risk. Some matters simply shouldn’t touch a third-party system, terms or no terms. Strip identifiers or keep the work in-house.

When the matter warrants it, get the client’s informed consent, and get it in writing. Opinion 512 is blunt that a generic clause tucked into the engagement letter doesn’t clear this bar, so have the actual conversation. It’s good practice and good defense if anyone ever asks.

How should an attorney review an AI-drafted memo?

This is the heart of responsible use, and it’s where the loudest failures have happened. Lawyers have been sanctioned for filing briefs full of case citations the AI simply invented, complete with realistic names and docket numbers. The 2023 *Mata v. Avianca* matter in the Southern District of New York is the cautionary tale every firm now references: two attorneys submitted a brief built on six fabricated cases and drew a $5,000 sanction. A memo isn’t a court filing, but the same fabrication risk applies, and a wrong memo can drive a wrong decision.

A reliable review pass covers four things, in this order:

• Citations. Pull up every case, statute, and rule in a real research database. If it isn’t there, it doesn’t go in the memo. No exceptions.
• Facts. Confirm the AI didn’t paper over a gap in the record by inventing a detail or restating something incorrectly.
• Legal reasoning. Check that the analysis matches the jurisdiction and the actual holdings, not a plausible-sounding average of what the model has seen.
• Confidential leakage. Make sure the draft didn’t drag in boilerplate or facts from some other matter.

The mindset that keeps you safe: treat the tool like a junior associate, not a finished product. You’d never sign your name to a green associate’s first draft without reading it. Same rule.

Does running a memo through an AI detector help?

Sometimes, but not the way people expect. A managing partner worried that an associate leaned too hard on AI might run a memo through a detector hoping for a verdict. Detectors don’t give verdicts.

These tools estimate the *probability* that text statistically resembles machine writing, based on patterns like predictability and sentence rhythm. As we cover in our breakdown of why AI detectors output probabilities, not proof, a score is a signal, not evidence. They throw false positives on exactly the kind of formal, formulaic prose legal writing tends to produce, and a 2023 study by Liang and colleagues found detectors are biased against non-native English writers. OpenAI even pulled its own classifier in July 2023 because it wasn’t accurate enough to trust.

So what’s a detector actually good for in a firm? Two honest uses. First, as a *readability* check: a memo that scores as heavily machine-like usually reads stiff, generic, and over-hedged, which is a quality problem worth fixing no matter who or what wrote it. Second, as a prompt to read more carefully, never as the basis for an accusation. If you want to make an AI-assisted draft read like a person actually wrote it, that’s a writing-quality task, and there are more articles on AI writing and detection that go deeper on the craft.

Building a firm AI policy

Stop deciding this memo by memo. Write it down once. A workable policy answers four questions:

1. Which tools are approved? Name the specific enterprise or zero-retention products lawyers may use.
2. What data is allowed in them? Draw the line between general legal questions and client-confidential facts.
3. Who reviews and signs off? Make a named human responsible for verification before anything goes out.
4. How is it disclosed and billed? Standardize the engagement-letter language and the billing approach.

A one-page policy plus a trained team will quietly resolve most of the situations that otherwise turn into anxious hallway conversations.

The bottom line

Law firms can absolutely use AI for client memos, and plenty already do it well. The obligations aren’t new or exotic. They’re your existing duties of competence, confidentiality, communication, and reasonable fees, applied to a new tool. Disclose when it’s material, in your agreement, or relevant to the bill. Verify every citation and fact, every time. And don’t mistake a detector score for proof of anything.

If part of your goal is making AI-assisted drafts read like a careful human wrote them, that’s a writing-quality job worth doing right. You can try PaperBleach to see how a draft scores and read more naturally, or see plans and pricing if your team needs volume.