How to Use Claude for Legal Work Without Violating Client Confidentiality

Using Claude for legal drafting and research can cut hours off your week. The challenge is doing it in a way that is consistent with your professional responsibility obligations, particularly the duty of confidentiality under Model Rule 1.6 and its state equivalents.

This guide walks through the practical options, from fully manual redaction to automated PII interception, so you can choose an approach that fits your workflow and risk tolerance.

Start With the Right Question

Before looking at tools, it helps to be clear on what the actual risk is. The concern is not that Claude will publish your client's information or share it with opposing counsel. The risk is that text you send to Claude.ai or any consumer AI product travels to a third-party server, where it may be retained, reviewed, or used in ways your client has not consented to.

Bar associations in New York, California, Florida, and elsewhere have issued guidance making clear that attorneys must evaluate AI tools against confidentiality obligations before using them for client matters. "Reasonable precautions" is the standard, and what counts as reasonable is still being defined, but the trend in guidance is toward requiring some form of data protection before using consumer AI services.

Option 1: Manual Redaction (Slow, Error-Prone)

The most obvious approach is to remove or replace identifying information yourself before pasting anything into Claude.

For simple tasks like analyzing a short contract clause, this is feasible. You read through the text, replace the client's name with a placeholder like [CLIENT], swap the opposing party for [OPPOSING PARTY], and remove any specific identifying numbers.

The problems with manual redaction are real. It is time-consuming for long documents. It is easy to miss things, especially in dense documents with multiple parties, addresses, dates of birth, and case numbers scattered throughout. And it produces no audit trail. You have no record of which real entities corresponded to which placeholders in a given conversation, which matters if a compliance question comes up later.

Manual redaction also breaks down under the consistency requirement. If you replace "John Doe" with "Client A" in one prompt and forget in the next, Claude's responses become incoherent. You end up spending more time managing the substitutions than you saved by using AI in the first place.

Option 2: Write Generic Prompts (Limited Use Cases)

For some tasks, you can avoid the problem entirely by keeping prompts generic. Instead of pasting a specific contract, you describe the clause type and ask Claude to draft language for it. Instead of summarizing a specific deposition, you ask Claude to produce a summary template.

This works for template-building and legal research on general questions. It does not work for matter-specific analysis, which is where AI tools provide the most value. The whole point of asking Claude to review a specific contract or summarize a specific deposition is the specificity.

Option 3: Use an Enterprise AI Agreement

Anthropic and OpenAI both offer enterprise tiers that include stronger data protection commitments, including agreements not to train on your data and data processing agreements that may satisfy certain compliance requirements.

For large firms with dedicated IT and procurement, an enterprise agreement is worth pursuing. For small firms of 5 to 50 attorneys, enterprise pricing and minimum seat counts typically make this impractical. Enterprise AI contracts are designed for organizations with compliance teams, not for a four-attorney litigation boutique.

Option 4: Automated PII Interception (The Practical Solution)

The most practical approach for small and mid-size firms is a tool that sits between your keyboard and the Claude API, automatically detects and replaces PII before any text is transmitted, and then restores real names in the output on the client side.

This is what PrivateClaude does. Here is how to set it up and use it in your daily workflow.

Step 1: Set Up Your Account and API Key

PrivateClaude requires a Claude API key from Anthropic. Go to console.anthropic.com, create an account, and generate an API key. This takes about five minutes. You pay Anthropic directly for API usage, which for moderate legal work typically runs a few dollars a month.

Add the API key in PrivateClaude's settings. The key is stored locally and used to authenticate requests to Claude's API.

Step 2: Create a Matter

Before starting a conversation, create a Matter for the case you are working on. Give it a name and an optional reference number that matches your case management system. The Matter is the container that keeps entity mappings consistent across all conversations related to that case.

This matters for multi-session work. If you come back to a matter next week and paste in a document that mentions the same client, the system will use the same pseudonym it assigned in the first session.

Step 3: Type or Paste Your Prompt Normally

You do not need to change how you work. Paste the contract, deposition excerpt, or case memo exactly as it appears. PrivateClaude scans the text using named entity recognition before it sends anything to Claude, detecting names, addresses, phone numbers, emails, SSNs, dates of birth, company names, and case numbers.

Each detected entity is replaced with a realistic pseudonym. "John Doe" becomes "Paul Roberts." The client's address becomes a different real-format address. The SSN becomes a different valid-format number. Claude receives the scrubbed version and never sees the original.

A badge on each message shows how many entities were masked, so you know the interception ran.

Step 4: Review the Entity Mapping Table

The sidebar shows a live table of every real value and its pseudonym replacement, with the entity type and timestamp. Before you rely on Claude's output for anything, it is worth scanning this table to confirm the entities you expected to be masked actually appear in it.

If you see a name that was not caught, you can note it. The Pro tier includes cross-variation matching, so "J. Doe" and "Mr. Doe" resolve to the same pseudonym as "John Doe," which handles most of the edge cases in real documents.

Step 5: Restore Real Names in the Output

When Claude responds, the output uses the pseudonyms. Each message has a "Show Real Data" toggle that swaps all pseudonyms back to the original values. This substitution happens in your browser and is never transmitted anywhere.

You can copy the de-anonymized text directly to your document. The output reads naturally with real names, and the PII never traveled to any server.

Step 6: Export the Mapping Table for Your File

Before closing a conversation, download the entity mapping table as a CSV. Store it with the matter file. This is your documentation that you took reasonable precautions: a timestamped record of which real entities appeared in which AI conversations and what pseudonyms were used.

This is the audit trail that manual redaction cannot produce.

What This Workflow Looks Like Day to Day

In practice, using PrivateClaude feels almost identical to using Claude.ai. You open the interface, find the matter, start a new conversation, and paste your text. The PII masking happens automatically and takes less than two seconds even for large documents. The main difference is the mapping table in the sidebar and the toggle button on each response.

For most attorneys, the adjustment period is less than one session. After that, it becomes the default way to use Claude.

The Starter tier at $0/month covers three matters and 50 messages, which is enough for a realistic test. The Pro tier at $29/month removes all limits and adds full entity coverage including SSNs and company names. You can start at deidentification-llm.vercel.app.

Frequently Asked Questions

Do I need to inform my clients that I am using AI tools? Some state bar guidance suggests yes, or at least that engagement letters should address AI use. Check your jurisdiction's specific guidance. Using a PII-safe interface strengthens the position that you have taken reasonable precautions regardless.

What if Claude's response introduces a pseudonym in a form I did not expect, like a possessive? PrivateClaude handles possessives and plurals in the de-anonymization step. "Roberts's" or "the Roberts matter" maps back correctly to the original name.

Can I use PrivateClaude for document drafting, not just analysis? Yes. The most common use cases are drafting demand letters and motions, summarizing depositions and discovery responses, analyzing contracts for specific provisions, and legal research on matter-specific questions. All of these work the same way through the interface.

What if I paste a document and the entity table shows zero detections? The interface shows a "No PII detected" indicator. This happens for documents that are genuinely free of identifying information, like a pure legal research memo on a general question. The prompt is sent to Claude without modification in that case.