AI Legal Research vs. AI Drafting: Where Claude Belongs

The most common and costly mistake firms make with AI is treating it as a monolith. It is not. "AI for legal work" covers two distinct tasks: finding authority and creating work product. They carry different risks, require different tools, and rely on different workflows. Conflating them is how lawyers end up filing briefs citing cases that do not exist and getting sanctioned for it.

This post draws that line. Mastering this distinction makes every AI tool decision in your practice easier to evaluate and defend.

The Two Categories of Legal AI Work

Start with the underlying task, not the tool. Every AI-assisted legal task is either a finding task or a synthesizing task.

Finding tasks are about retrieving specific, verifiable facts from an authoritative source. You need a case that stands for a proposition. You need the current text of a regulation. You need to know whether a controlling circuit has addressed an issue. These tasks demand certainty. An answer that sounds plausible but is fabricated is worse than no answer at all.

Synthesizing tasks are about taking information you already possess and shaping it into persuasive, organized work product. You have the research memo. You need the motion. You have the deposition transcripts. You need the argument. These tasks require craft, structure, and expression. A poorly structured argument can be edited. A brief citing cases that do not exist is a sanctionable offense that destroys your credibility with the court.

The question to ask before any AI-assisted task: is this a finding task or a synthesizing task? The answer tells you which category of tool belongs in the workflow.

Retrieval AI: Built for Finding

Westlaw AI-Assisted Research, Lexis+ AI, CoCounsel (formerly Casetext), and similar platforms are built on curated legal databases. When these tools return a case, it is generally because the case exists in that database. The hallucination risk is fundamentally different because the output is tied to a specific retrieved document, not generated freely from training data.

These tools are not perfect. They can miss cases. A retrieved case can be distinguished, overruled, or misread by the attorney. But the failure modes are familiar: they are the same ones that existed before AI. An attorney who misreads a Westlaw result is making a familiar type of research error. An attorney who files a brief with fabricated citations from an LLM, as happened in Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023), is making a new and more dangerous category of mistake.

Purpose-built legal research platforms belong at the front of your workflow. They produce the verified research memo. That memo is the input to the next stage.

Generative AI: Built for Synthesizing

Claude is a large language model. Its training included an enormous range of text, including legal text, but it does not have a live connection to Westlaw. It cannot retrieve the current text of a statute. When it produces a case citation, it is generating a plausible-sounding citation based on patterns in training data. Some citations will be real. Some will be invented. Claude cannot tell the difference, and neither can you without checking.

This is not a defect. It is a classification. It tells you where Claude belongs in your workflow and where it does not. For a full breakdown of the professional responsibility exposure this causes, see the companion post on AI citation hallucinations in legal filings.

Where Claude earns its keep for attorneys:

• Drafting a motion from a research memo you provide
• Restructuring an argument for clarity or persuasive flow
• Playing opposing counsel to stress-test your theory of the case
• Drafting client communications, contract provisions, and demand letters from facts you supply (see also the PI demand letter template)
• Organizing and analyzing medical records and deposition exhibits into structured summaries (see the guide on AI medical chronologies)
• Turning deposition notes into organized summaries
• Structuring a settlement negotiation protocol from facts and authority you supply

Every item on that list is a synthesizing task. Claude is working on material you control and verify. It is not being asked to retrieve authority from its training data.

The Clean Room Workflow

The practical application of this distinction is what the Leveraged Attorney course calls the Clean Room method. The workflow runs in two distinct stages:

Stage 1: Research (retrieval tools): Use Westlaw, Lexis+, or another purpose-built platform to find and verify authority. Draft a research memo that contains verified citations, quoted passages, and your analysis of how each authority applies. Every citation in this memo has been confirmed to exist and to say what you say it says.

Stage 2: Drafting (Claude): Open a Claude session. Paste the research memo as context. Instruct Claude to draft the motion, brief, or document using only the authority in your memo. Claude is synthesizing from verified inputs, not generating citations from training data.

The separation is the protection. Module 3 of the Leveraged Attorney course covers Stage 1 in depth, including how to organize a research memo for maximum Claude usability. Module 4 covers the drafting workflow itself, including how to structure prompts so Claude stays within the authority you have provided rather than supplementing with generated citations.

For attorneys who want to see the drafting stage applied to motions and briefs specifically, the post on drafting motions and briefs with Claude walks through the full technique.

For the complete picture, see the full attorney workflow guide for Claude.

The Hybrid Tools: Where It Gets Complicated

The market has responded to attorney demand with hybrid products that combine retrieval and generation. Westlaw's AI features return retrieved passages alongside generated summaries. Harvey drafts documents with a research layer underneath. Lexis+ AI mixes retrieved caselaw with generated analysis.

These tools can be good. The relevant question for any hybrid platform is: which parts of this output are retrieved and which parts are generated? Sophisticated legal AI buyers ask this question on every tool evaluation call. If a vendor cannot give you a clear answer on what is retrieved versus generated, that is a reason not to buy the product.

The practical rule for any hybrid tool output: treat retrieved text as verified only after you have read the source document. Treat generated summaries and generated analysis as drafts requiring your review. For risk management purposes, assume the generative layer in a hybrid tool will hallucinate on the same basis as any other large language model. The retrieval layer grounds it, but the boundary between them is not always visible in the output.

What This Means for AI Tool Decisions

Attorneys who understand the research-versus-drafting distinction evaluate AI tools with precision. They do not buy a drafting tool expecting it to replace Westlaw. They do not reject Claude because it hallucinated when asked to retrieve cases, having never understood that citation retrieval is outside its design envelope.

A useful framework for evaluating any legal AI product:

1. Is this tool primarily retrieval-based or generation-based?
2. For hybrid tools: what is the source of each output element, and how can I verify the retrieved portion?
3. What workflow does this tool fit into, and what stage of that workflow does it own?
4. What does the vendor say about citation accuracy, and is that claim audited or anecdotal?

The attorneys who get the most from AI are not the ones with the largest tool stack. They are the ones who understand precisely what each tool can and cannot do, assign tasks accordingly, and maintain human review at the stage where output enters the public record.

The Practical Takeaway

Before any AI-assisted task, ask the question: am I finding something or expressing something?

If you are finding, the tool needs a verified legal database behind it. If you are expressing, synthesizing, or structuring, Claude is a capable partner once you have supplied verified inputs.

This mental model eliminates most of the confusion in the legal AI market. It also eliminates most of the professional responsibility risk. The attorneys who face sanctions for AI-generated citations are, almost without exception, using a generative tool at the finding stage. The workflow separation described here prevents that error at the architectural level.

The Leveraged Attorney curriculum is built around exactly this discipline. The Clean Room method is not a theoretical safeguard. It is the operational standard that makes high-volume AI-assisted practice possible without accumulating the kind of exposure that ends careers.