How to Use Claude for Legal Document Analysis: What Works

Claude won’t replace lawyers. It’ll expose the ones who skip verification.

The pitch sounds perfect: upload a 200-page merger agreement, ask Claude to flag risks, get a summary in seconds. But here’s what the tutorials don’t tell you: Stanford research found legal AI tools hallucinate in roughly 1 out of 6 queries – including citing overturned precedent as current law. A federal court ruled in February 2026 that Claude chats aren’t protected by attorney-client privilege. And the 200K context window everyone celebrates? It comes with an undocumented output cap that truncates summaries mid-sentence.

Why Claude Actually Works for Legal Document Analysis (and Where It Doesn’t)

Claude handles contract-heavy workflows better than most alternatives because of one architectural decision: its context window. Claude Sonnet 4.6 processes up to 1 million tokens (as of March 2026, now generally available), which translates to roughly 750,000 words. That’s an entire commercial lease, a deposition transcript, and the associated exhibits – in one conversation.

Compare that to tools that force you to chunk documents into segments. When you split a contract, the AI loses cross-references. It can’t see that Section 4.2’s indemnity clause conflicts with the limitation of liability in Exhibit C. Claude keeps the full context.

But the context window creates a false sense of security. According to multiple legal AI analyses, Claude’s output is capped at 4K-8K tokens depending on the model – even when processing 200K tokens of input. Request a detailed summary of a 150-page agreement and you’ll get 3,000 words that stop mid-analysis with no warning. The model doesn’t tell you it ran out of space. It just stops.

Configuration: The Part Most Guides Skip

Sign up, upload a PDF, type a prompt. That’s the standard tutorial. It’s also how you accidentally violate client confidentiality rules.

Start here instead: decide whether you’re using Claude.ai (the web interface) or the API. They have different data retention policies. Per Anthropic’s documentation and confirmed by legal analysis sources, the standard API allows your inputs to be used for model improvement unless you explicitly opt out. The web interface has similar terms. For privileged legal work, you need an enterprise agreement with a Data Processing Addendum that specifies deletion timelines and breach notification obligations.

If you’re using the free or Pro plan ($20/month), you’re subject to usage limits that reset every 5-8 hours. The Nevada State Bar’s 2026 review noted Claude is suitable for “routine, non-confidential tasks” – but not ready for privileged legal work yet.

Next: redaction protocols. Before uploading a client contract, strip identifying information. A Boston IP firm that uses Claude for patent license reviews built a Python script that replaces entity names with placeholders like [LICENSEE] and [PATENT_OWNER]. This takes 15 minutes upfront. It prevents accidental disclosure if your data is cached or logged.

Pro tip: Create a dedicated folder for Claude-ready documents. Name it “Claude_Redacted” and train your team to never upload from your main client files. One wrong drag-and-drop can send privileged information to a third-party server.

Writing Prompts That Actually Produce Usable Legal Analysis

“Summarize this contract” produces garbage. Claude needs constraints.

Effective legal prompts follow this structure: role + document type + specific extraction points + output format. Example: “You are a commercial litigation attorney with 10 years of experience. Analyze the attached MSA for: (1) indemnity scope, (2) limitation of liability caps, (3) dispute resolution mechanisms, (4) any unusual termination clauses. Provide output as a bullet list with specific section references.”

The role-setting isn’t cosmetic. Testing across legal workflows shows that specifying expertise level dramatically changes output sophistication. “You are a senior attorney” produces more nuanced clause interpretation than no role specification.

For contract comparison, upload both versions and prompt: “Compare Version A and Version B. Identify all substantive changes in: payment terms, warranties, IP ownership, and confidentiality obligations. Ignore formatting and typo corrections. Flag any changes that increase our client’s risk.”

Output format matters because legal teams need portable results. Request tables for multi-document comparisons. Request JSON for clause extraction that feeds into contract management systems. Request numbered lists with section citations for internal memos.

Here’s the constraint most lawyers miss: specify what NOT to do. “Do not provide legal conclusions or recommendations. Extract factual terms only.” This reduces hallucination risk and keeps Claude in its lane – data extraction, not legal advice.

Handling Multi-Document Analysis Without Hitting Context Limits

Upload five NDAs and ask Claude to identify common risk clauses. Simple, right?

Each document adds to your token count. Five 20-page NDAs might total 100K tokens. That fits comfortably in Claude’s window. But if you’re using the API and your next request pushes you over 200K tokens, Anthropic’s pricing documentation reveals a trap: premium pricing of $10 input / $37.50 output per million tokens kicks in automatically. The entire request gets charged at the premium rate, not just the overage.

Workaround: use meta-summarization. Process each NDA individually, extract key points into a structured summary (200-300 tokens each), then combine the summaries and analyze those. You’ve condensed 100K tokens down to 1,500 tokens while retaining the information you actually need.

For due diligence document sets – think 50+ contracts – use Anthropic’s Batch API. It offers a 50% discount on both input and output tokens by processing requests asynchronously. You submit the batch at 5pm, results are ready by morning, and you’ve cut your API costs in half.

The Hallucination Problem: What It Looks Like in Legal Context

Claude hallucinates less than GPT-4 according to Stanford’s Center for Legal Informatics. “Less” doesn’t mean “never.”

Legal hallucinations aren’t random gibberish. They’re plausible-sounding fabrications. Claude might cite a real case name with an invented holding. It might confidently state that a statute uses specific language that doesn’t exist. In one documented example from Stanford’s HAI research, a legal AI tool (using RAG architecture, which many Claude integrations employ) cited the pre-Dobbs “undue burden” standard for abortion restrictions as current law – months after Dobbs overturned it.

This failure mode is particularly dangerous because the output reads correctly. It uses proper legal terminology. It formats citations properly. A junior associate skimming the summary might not catch it.

Verification workflow: every case citation gets checked in Westlaw or Lexis. Every statutory reference gets confirmed in the actual code. Every “according to” statement gets traced back to a primary source. No exceptions. A 2025 study found that ChatGPT-4 achieved 68% viability in contract-related responses when evaluated by legal experts – which means 32% of outputs were problematic.

The community-built Claude legal skill on GitHub reports an F1 score of approximately 0.62 on ContractEval benchmarks for clause extraction. That’s useful for first-pass review and issue flagging, but it’s explicitly “not a replacement for attorney review on material deals.”

When Claude Gets Jurisdiction Wrong

American legal rules differ by state, circuit, and time period. Documents that seem relevant due to semantic similarity may be legally inapplicable.

If you ask Claude about California employment law and it pulls from a Delaware statute, the answer is wrong even if the reasoning sounds right. RAG systems retrieve based on text similarity, not binding authority. Stanford’s study highlighted this: retrieval occurs, but the retrieved document can be inapplicable.

Mitigation: include jurisdiction explicitly in every prompt. “Under New York contract law, analyze this non-compete clause.” Then verify that Claude’s cited sources actually come from New York precedent, not federal or out-of-state cases that don’t bind.

Claude vs. Legal-Specific Tools: The Honest Comparison

ChatGPT, CoCounsel, Spellbook, Harvey, Lexis+ AI. What’s the actual difference?

Claude is a general-purpose LLM. It wasn’t trained specifically for legal work. Legal-specific tools like Spellbook or CoCounsel are built on top of foundation models (sometimes Claude, sometimes GPT) and then fine-tuned on legal datasets. They include clause libraries, playbooks, and built-in compliance checks.

Claude’s advantage: flexibility and cost. At $3 input / $15 output per million tokens via API, it’s cheaper than most legal-specific subscriptions. You can craft custom workflows without being locked into a vendor’s predetermined templates. For routine document summarization, clause extraction, or first-draft generation, Claude performs comparably to tools that cost 10x more.

Claude’s disadvantage: no legal guardrails. Spellbook automatically flags inconsistencies and missing clauses in contracts. CoCounsel integrates with Thomson Reuters’ legal research database. Lexis+ AI provides Shepard’s citation validation. Claude gives you raw LLM output with no verification layer.

Use Claude when: you’re doing high-volume, low-stakes document triage; you need flexible prompting for non-standard document types; or you’re building a custom legal workflow via API.

Use legal-specific tools when: you need compliance automation; you’re handling regulated industries with specific clause requirements; or your firm prioritizes vendor liability over cost savings.

There’s no universal “best.” It’s a cost-risk tradeoff. Firms with in-house technical talent often build custom Claude workflows via API. Firms without that capacity buy legal-specific SaaS.

The Privilege Problem: What the February 2026 Ruling Means

US v. Heppner (SDNY, February 2026) was the first federal decision to address whether Claude chats are protected by attorney-client privilege. The court said no.

The reasoning: Claude isn’t a lawyer, and “discussion of legal issues between two non-attorneys is not protected.” The defendant had no reasonable expectation of confidentiality when using a public AI tool. And Claude explicitly disclaims providing legal advice in its terms of service.

Under English law, the outcome might differ – litigation privilege there doesn’t require attorney involvement. But in U.S. jurisdictions, this ruling creates a roadmap for opposing counsel. If you use Claude to analyze a client matter and don’t have a DPA or enterprise agreement, those chats may be discoverable.

Practical implication: ABA Formal Opinion 499 requires informed consent when using technology that introduces new confidentiality risks. Template language: “We will use AI-assisted analysis subject to strict redaction protocols and retain full responsibility for final output. AI-generated content is not protected by attorney-client privilege unless covered by an enterprise agreement.”

Skip this conversation with clients and you’ve got both an ethics violation and a potential waiver of privilege.

Building a Verification Checklist That Actually Catches Errors

“Review AI output” is useless advice. What does that mean in practice?

Start with a two-tier system. Tier 1 (quick check, under 5 minutes): confirm document type matches request; spot-check 3 random citations; verify all extracted dates and dollar amounts against source document; check that output didn’t truncate mid-sentence.

Tier 2 (full review, 15-30 minutes): verify every case citation in primary sources; cross-reference all statute numbers; check jurisdiction for all cited authorities; compare extracted clauses against original document sections; validate that “according to” statements trace to real sources; flag any definitive legal conclusions (Claude shouldn’t be making those).

Use Tier 1 for low-risk tasks like internal memos or preliminary issue-spotting. Use Tier 2 for anything that goes to a client, gets filed with a court, or informs a negotiation strategy.

Document your process. When you catch an error, log it: date, document type, error type, what the prompt was. After 20-30 logged errors, patterns emerge. Maybe Claude consistently misidentifies indemnity clauses in SaaS agreements. Now you know to manually review that section every time.

When to Stop Using Claude and Call a Human

Novel legal issues. Ambiguous contractual language that requires judgment calls. Multi-jurisdictional questions where the law conflicts. Regulated industries with strict compliance requirements.

Claude accelerates document processing. It doesn’t replace legal reasoning. If your instinct is “this feels complicated,” don’t let Claude’s confident-sounding output override that instinct. Escalate to a senior attorney.

Cost Management: How to Avoid the 200K Token Pricing Trap

That 1-million-token context window sounds great until you see your API bill.

Here’s what Anthropic’s pricing docs bury: when your input exceeds 200K tokens, the entire request – every single token – gets charged at the premium rate of $10 input / $37.50 output per million tokens (for Sonnet 4.5/4.6). Not just the tokens over 200K. All of them.

A 250K-token request costs $2.50 input + $9.38 output (assuming 2,500 output tokens) = $11.88. At standard rates, the same request would cost $0.75 + $3.75 = $4.50. You just paid 2.6x more because you crossed the threshold by 25%.

Strategies to stay under 200K: use prompt caching (stores repeated context, charges 90% less on cache hits); implement meta-summarization (process chunks, combine summaries); use the Batch API for non-urgent work (50% discount); switch to Haiku 4.5 for simple extraction tasks ($1/$5 per million tokens).

For law firms doing high-volume document review, the subscription vs. API calculation flips. Claude Pro at $20/month includes higher rate limits and access to Opus models. If you’re processing 50+ documents per week, the subscription is cheaper than API usage. If you’re doing 5 documents per month, API wins.

What Comes Next: Where Legal AI Is Actually Heading

Claude Code (released January 2026) introduced agent teams – multiple AI instances collaborating on complex tasks. Legal applications are starting to experiment: one agent redlines a contract, another checks citations, a third generates a summary memo.

But the regulatory environment is tightening. The National Center for State Courts released guidance in February 2026 emphasizing that attorneys “should never submit AI-generated content to courts without thorough review and citation checking.” If hallucinations are discovered after filing, you must correct immediately and notify the court and opposing counsel.

Expect more rulings like Heppner. Expect state bars to issue ethics opinions explicitly addressing AI use. Expect malpractice insurers to start asking about your AI verification protocols.

The firms that thrive won’t be the ones using the most AI. They’ll be the ones with the best verification workflows.

Frequently Asked Questions

Can Claude draft a contract from scratch, or does it only analyze existing documents?

Claude can generate contracts from scratch if you provide detailed parameters (parties, terms, jurisdiction, key clauses). But the output is a first draft that requires substantial attorney review. Legal-specific tools like Spellbook or CoCounsel are better for this because they include clause libraries and compliance checks. For analysis and summarization of existing documents, Claude performs well. For drafting, it’s a starting point – not a finished product.

What’s the actual file size limit when uploading documents to Claude?

Via the web interface, you can upload up to 5 documents at a time. Each document is converted to tokens, not measured by file size. A 200-page PDF might be 150K tokens or 80K tokens depending on formatting and density. The context window is 200K tokens (standard) or 1M tokens (Sonnet 4.6, generally available as of March 2026). Practically, you can upload several large contracts simultaneously – but remember that output is capped at 4K-8K tokens, so even if you upload 500 pages, the response length is limited.

How do I know if Claude is citing real cases or making them up?

You don’t – unless you verify. Claude sometimes generates plausible-sounding citations that are completely fabricated. The National Center for State Courts guidance is explicit: “Always check citations directly in primary sources, and verify case names, holdings, and references independently.” Use Westlaw, Lexis, or Google Scholar to confirm every case citation Claude provides. If the case doesn’t exist or the holding is wrong, don’t use the output. Stanford research found that legal AI tools (including RAG-based systems) hallucinate in roughly 1 out of 6 queries, so this isn’t a rare edge case – it’s a documented failure mode.

Start with one low-stakes document you’ve already reviewed. Run it through Claude. Compare the outputs. Log the errors. Then build your verification checklist around the mistakes you actually encounter.