How to Use AI to Write Legal Memos Faster and More Effectively

A lawyer can draft a competent legal memo with AI in under an hour, but only by treating the model as a junior drafter rather than a research source. AI writes a credible first draft when the lawyer supplies verified facts and authorities, structures the prompt with care, and reviews every citation before the memo leaves the desk. The work AI handles is structural. The work that requires judgment stays with the lawyer.

This article lays out a five-step drafting workflow, prompt templates for each stage, a verification checklist, and a framework for choosing the right tool. Drafting is already one of the most widespread uses of AI in legal practice today, so this article takes the value of AI in memo drafting as a given and focuses on how to do it well. Done right, this workflow recovers hours of drafting time per memo without compromising accuracy or professional duty.

The Structural Reason AI Drafts Memos Well

Legal memos follow a fixed analytical frame. IRAC and CREAC give the writer a template to fill rather than a blank page to invent. The lawyer knows what each section needs to do, the reader knows what to expect, and the analysis moves in a predictable order. Legal drafting AI handles that kind of structured writing well and stumbles on the unstructured, judgment-heavy parts.

Not every memo is an equally good fit. Internal predictive memos, the ones that answer how likely an outcome is given a set of facts, are the strongest candidates for AI drafting. The audience is internal, the tone is analytical rather than persuasive, and the structure is the most standardized across firms and practice groups. Intra-firm research memos sit close behind. The lawyer's job here is synthesis, and the AI can carry a meaningful portion of that work once the authority is in hand.

Client advisory memos and persuasive briefs need more lawyer involvement at every stage. Client advice carries reputational weight and demands a register the AI rarely lands without rewriting. Persuasive briefs require rhetorical choices, framing decisions, and selective emphasis that no general model handles reliably. AI still has a place in those documents, but the ratio shifts. The lawyer does more of the writing, and the AI does more of the structural support.

Drafting is now one of the most common AI use cases in practicing law. Lawyers at firms and in-house teams are already using Legal AI for first drafts of memos, letters, and routine correspondence. The harder problem is building a workflow that produces a memo worth signing.

How to Write a Prompt That Produces a Usable Legal Memo Draft

The quality of an AI-drafted memo is determined almost entirely by the quality of the prompt. Most lawyers experimenting with AI know this in theory. Far fewer have a working template they can adapt under deadline pressure. A strong legal-memo prompt contains four elements, each one shaping a different part of the model's output. They are role, context, source material, and output specification.

Role

State who the AI is supposed to be in the memo. A typical role line reads, "You are a senior associate drafting an internal predictive memo for a partner." This frame anchors the tone, the analytical depth, and the level of caution the model brings to the draft. A model told to write as a senior associate treats authority more carefully than a model told to write "a legal memo."

Context

Give the jurisdiction, governing law, procedural posture, and any deadline that affects scope. Specificity matters. "Second Circuit, federal employment discrimination, motion to dismiss pending" tells the model what authority is binding, what doctrine governs, and what stage of the matter the memo is meant to inform. A model without this context guesses, and a guessing model produces a memo that misses the point.

Source material

Paste or upload the case opinions, statute text, regulatory guidance, and client facts the memo needs. Tell the model to use only this material as authority. The single most important sentence to add to every legal-AI prompt is this. "Do not cite any authority that is not present in the materials I have provided. If you need authority to support a point, flag the gap rather than filling it." This one sentence is the line between a draft you can verify and a draft that fabricates citations.

Output specification

Tell the model the format you want, including memo type (predictive or advocacy), structure (IRAC or CREAC), section length, citation format, and tone. A clean output line might read, "Draft a predictive memo using CREAC. Each section should run 200 to 400 words. Use Bluebook 21st edition citation format. Tone should be analytical, not persuasive." Precision in this step is what keeps the model from drifting into a generic format that the lawyer has to rebuild from scratch.

A complete prompt you can adapt

Here is the structure applied to a realistic fact pattern. The question is whether an employee's social media post counts as protected concerted activity under Section 7 of the National Labor Relations Act.

• Role. You are a senior associate drafting an internal predictive memo for a partner who supervises labor and employment matters.
• Context. Jurisdiction is the Second Circuit. Governing law is Section 7 of the NLRA. Procedural posture is pre-complaint. The client is the employer. Memo is due in 48 hours.
• Facts. [Insert client facts. The employee's job title, what the post said, who could see it, whether other employees engaged with it, and the disciplinary action the employer is considering.]
• Source material. [Attach the operative NLRA text, the leading Board decisions, and any Second Circuit cases reviewing Board orders on the same question. Use only this material as authority.]
• Instruction. Draft a predictive memo using CREAC. Each section should run 200 to 400 words. Use Bluebook 21st edition citation format. Tone should be analytical, not persuasive. Do not cite any authority that is not present in the materials I have provided. If you need authority to support a point, flag the gap rather than filling it.

Two prompts that don't work

Two patterns produce unusable output. "Find me cases supporting the position that X" invites the model to hallucinate authority. The model has no path to verified case law and treats the request as a writing exercise. "Write me a memo about Y" produces generic content that the lawyer has to throw out and start over. Both patterns share the same flaw. They ask the AI to do legal research instead of legal drafting.

What the Lawyer Does and What AI Does

The five-step workflow looks different on the page than it does in practice. To make it concrete, here is the same Section 7 NLRA fact pattern from the prompt template, walked through end to end. Five stages, one memo, two participants.

1. Inputs assembled

The lawyer compiles the operative facts from the client interview and the personnel file. The governing statute is pulled from a verified database, along with the leading Board decisions and the Second Circuit cases reviewing them. None of this involves AI. The principle is that every authority the AI will be asked to use is already in front of the lawyer before the prompt window opens. The lawyer is doing the research. The AI will do the drafting.

2. Prompt drafted

The lawyer assembles the prompt using the four-element structure. The role is a senior associate writing for a labor and employment partner. The context names the Second Circuit, the NLRA, and the pre-complaint posture. The facts and source material are pasted into the prompt window in full. The output instruction asks for a predictive CREAC memo with Bluebook citation and an explicit rule against citing material not in the source set.

3. AI draft generated

The model returns a structured memo in under a minute. The CREAC outline is intact. The rule statements are accurate but generic, reading more like treatise summaries than working analysis. The application section is thin in two places. It restates the rule rather than applying it to the facts, and it elides the Board's analysis of when individual conduct rises to concerted activity, which is the heart of the matter. The lawyer flags those weaknesses on the first read.

4. Lawyer revision

This is where the memo becomes legal analysis. The lawyer tightens the conclusion section, which had drifted into rhetorical territory. The analysis section gets the most work. The lawyer removes an overreach in the application paragraph that claimed the post was "clearly" protected, because the cases don't support that adverb.

A counterargument the AI missed gets added. The Board's recent narrowing of Section 7 in the social media context cuts against the conclusion, and the partner will ask about it. One supporting case is swapped for a stronger one. The original AI selection was on point but not the most recent or the most analogous. These are substantive edits, and most of them require legal judgment the model couldn't have made.

5. Final verification

The lawyer pulls every citation. Each pinpoint is checked against the source. Quoted language is confirmed letter by letter. Two minor citation format errors are corrected. One case the model cited is verified to exist, in the right reporter, on the right page.

This stage sets the legal floor. Verification confirms that the foundation under the analysis is real, while the analysis itself was already validated in stage four.

The walkthrough shows a division of labor. The lawyer did the parts that demand judgment, and the AI did the parts that demand structure. The workflow itself protects against the failure modes that have produced sanctions in the field, because the lawyer never asks the AI for authority and never signs a memo without verifying every cite.

What Legal Memo Work Requires That General-Purpose AI Can't Provide

Legal memo work is not a place to triage by risk tier. Every memo touches the lawyer's duties of confidentiality, competence, candor, and supervision, and those duties carry architectural requirements. A platform that meets them by design is the right default for the work. A platform that does not meet them is the wrong category of tool.

What follows are six requirements that legal memo work imposes on any AI tool that touches it, regardless of how routine the matter looks. Every one of these requirements holds for routine memos as much as for high-stakes ones.

Citation grounding against primary legal sources

A memo's authority must be real, in force, and accurately quoted. A platform that retrieves from verified legal databases and surfaces the underlying material at the point of citation collapses the verification step into a glance at the source. A platform that predicts the next plausible-looking citation does not. The lawyer verifies every cite either way, but the time and risk profile of those two workflows are not the same.

Matter-level isolation

The duty of confidentiality requires that information from one matter does not migrate to another. Enterprise legal platforms enforce that boundary in their architecture, so prompts and source material from one client engagement stay within that engagement. Consumer AI products do not. Anything pasted into a consumer prompt window may be used to train future model versions or appear in other users' sessions, depending on the provider's policies at any given moment.

Audit logging at the firm level

Supervisory duties require visibility into the AI's role in the work product. A partner supervising an associate who uses AI needs to know which prompts ran, what materials were used, and what the model returned. Without an audit trail tied to the firm rather than to a consumer account, supervision becomes guesswork. Firms that adopt AI without firm-level logging are accepting a supervisory exposure they cannot remediate after the fact.

Data residency and retention controls

Cross-border practice and sector-specific privacy regimes (GDPR, HIPAA, US sectoral rules) require contractual and architectural controls over where prompts go and how long they persist. Legal-specific platforms treat these controls as the operating assumption, built in from the start. A consumer AI tool that processes prompts in an undisclosed jurisdiction and retains them on undisclosed terms is incompatible with the duties owed to a client whose data sits inside those prompts.

Document management system integration

Legal memos live in the same document management system as the underlying case file. A tool that requires copy-paste between systems creates a confidentiality exposure on every memo, because every paste is a fresh opportunity for the wrong material to land in the wrong window. Integration with the firm's DMS keeps the work inside the same trust boundary as the rest of the matter file. Without it, the AI workflow is fighting the firm's own information governance.

Domain-trained reasoning

Legal analysis has structural patterns that benefit from training on legal corpora. A model that treats a case opinion as ordinary prose handles legal reasoning the way a general translator handles a contract, fluently and unreliably. Domain-trained models recognize holdings, distinguish dicta, weight subsequent treatment, and structure analysis the way lawyers do. General models can imitate the surface of legal writing, but the analytical work underneath is not the same.

Every one of these requirements holds for routine memos as much as for high-stakes ones. A tool that lacks them is not "good enough for outlining" or "fine for non-client work." It belongs in a different category of work, not a different price tier of the same work.

Domain-specific AI platforms built for legal work address this category of problem at the system level rather than asking the lawyer to catch fabrication and confidentiality issues after the fact. Harvey is one widely adopted example, used across more than 60% of the AmLaw 100 and by over 142,000 legal professionals worldwide, grounding outputs in verified legal sources, enforcing matter-level isolation, and surfacing the underlying authority for every claim.

Why the "non-legal use only" workaround doesn't hold

Some lawyers say they use general-purpose AI only for "non-legal" tasks. Restructuring an outline. Copy-editing a paragraph. Brainstorming a section heading. The line between legal and non-legal use does not survive contact with a real memo. The outline contains client facts. The paragraph names a party. The section heading reveals the theory of the case. Every prompt that touches the memo touches privileged work product, and there is no version of "I'll only use it for the safe parts" that survives a thoughtful conflicts or privilege analysis.

The Five-Step Workflow for AI-Drafted Legal Memos

The workflow runs in five steps, each one anchored to a human decision the AI cannot make.

1. Frame the assignment

Write the question presented, governing law, jurisdiction, and key facts in a structured prompt before opening the AI tool. The lawyer's framing decisions belong in the prompt itself, settled before the model drafts a word.

2. Pull verified authority first

Identify the cases, statutes, and regulations the memo needs from a trusted database. Do not ask the AI to find them. Hallucinated citations enter the work product when authority retrieval is delegated to a general model.

3. Have the AI draft the structure

Generate an IRAC or CREAC outline using the facts and authorities you supplied. The structure should reflect the analytical frame the memo will follow, shaped by the facts and authorities you gave it.

4. Draft section by section with grounded prompts

Feed the AI the relevant case text and ask it to draft each section, one at a time. Section-level prompts produce stronger output than asking for an entire memo in a single pass.

5. Verify, revise, and own the final product

Check every citation against the source. Rewrite the analysis in your own voice. Sign nothing you have not read.

Every step trades speed for control in the right direction. The lawyer keeps the decisions that demand judgment. The AI handles the structural work it does well.

What Changes for the Lawyer When AI Handles the First Draft

The time saved on drafting moves to the parts of memo writing that demand the most from the lawyer, which are the parts that always mattered most.

Upstream, more time goes to issue spotting, fact development, and choosing which authority matters. A bad question produces a bad memo no matter who drafts it, and AI cannot tell the lawyer which question to ask. The minutes recovered on drafting are minutes the lawyer can put into framing the inquiry. That framing decides whether the memo is useful or beside the point.

Downstream, more time goes to the parts of analysis that demand judgment. Counterarguments. The limits of a case's reasoning. The practical implications for the client. AI will draft an application section that summarizes how the rule fits the facts, but it will not tell the partner that the rule's recent treatment in a sister circuit cuts against the conclusion. Those judgments are what the lawyer is paid for, and they become a larger share of the work as drafting takes less of it.

Junior associates face a training tradeoff with this workflow. Drafting the analysis section is how lawyers build legal judgment. An associate who lets AI carry that section produces a memo on time and learns nothing from the exercise. The right approach is to use AI as a structural assistant, then write the analysis themselves. The structural work is the part safe to delegate. The analytical work is the part that builds the lawyer.

This is the workflow Harvey is built to run, with verified authority pulled before drafting, grounded prompts, traceable citations, and matter-level isolation built in. The more than 142,000 legal professionals using Harvey across over 60% of the AmLaw 100 and in 60-plus countries are not asking it to draft memos for them. They are using it to push the structural work to a system that handles it well, so their own time goes to the parts of memo writing that demand a lawyer.

The firms recovering the most time are not the heaviest AI users. They are the ones that moved early, built their prompt libraries, integrated Harvey with their document management systems, and made the new workflow their team default. The gains compound across memos. Each memo runs faster than the last because the prompt library is sharper, the integrations are smoother, and the firm's review patterns are baked into the workflow.

The fastest way to see whether the same shift fits your team is to request a Harvey demo and run it against a memo your team is drafting this week.