Claude for Immigration Lawyers: Complete 2026 Guide

Three numbers define the current state of Claude adoption in immigration practice.

Attorneys are using Claude for immigration work. Most of them are using it wrong.

The wrong looks like this: a solo practitioner on a $20/month Pro plan, pasting client documents into a conversation, getting competent drafting output, and not realizing they have just routed client-identifying information through infrastructure that Anthropic can use for model training. No BAA. No contractual protection. Bar complaint waiting to happen.

The right looks different. An attorney on Enterprise, with a Project configured per client matter, using a system prompt that embeds the legal standard and output format, running Opus 4.8 for high-stakes criterion arguments and Sonnet 4.6 for the ten tasks that do not require that level of reasoning. Saving three to five hours per petition. Producing first drafts that require thirty minutes of editing, not three hours.

The gap between these two uses is not about AI sophistication. It is about knowing which plan has a BAA, what a Project actually does, and which tasks justify the per-token cost difference between the two models.

Which Claude Plan You Actually Need

This is the most commonly misunderstood piece of Claude for legal work, so it gets first position.

There are three relevant consumer plans. Two of them cannot be used with client data.

Claude Pro ($20/month): No BAA. Do not use with client-identifying information. Suitable for tasks that contain zero client-specific data: drafting generic criterion frameworks, analyzing public USCIS policy text you copied from uscis.gov, or drafting template language for a type of expert letter. The moment a specific client name, A-number, date of entry, or case-specific fact enters the conversation, you are in confidentiality violation territory.

Claude Team ($25-30/seat): No BAA. Same restriction as Pro. The Team plan adds collaboration features and slightly higher usage limits, but it does not change Anthropic's data handling terms. Attorney confidentiality obligations do not scale to fit your billing tier.

Claude Enterprise (~$60/seat): BAA available. This is the tier you can use for client work. The BAA requires activation by your organization's administrator in the account settings (see Anthropic's privacy policy for what Enterprise data handling covers). After activation, Anthropic contractually commits to not training models on your inputs and to meeting HIPAA-level security controls. For immigration attorneys, the direct HIPAA connection is limited since immigration matters do not typically involve protected health information. The BAA functions as a proxy: it establishes the contractual data protection baseline that satisfies ABA Rule 1.6 confidentiality obligations.

One thing the BAA does not do: it does not create attorney-client privilege over communications with Claude. The Heppner ruling addresses this directly (more on that in the confidentiality section). The BAA protects against unauthorized disclosure; it does not transform AI assistance into privileged communication.

If your firm cannot justify $60/seat for Enterprise, the practical alternative is to use Pro or Team only for non-client work, and use the API directly for client work with appropriate access controls and data handling agreements. API customers with qualifying usage have BAA access under separate terms.

Setting Up Claude Projects for Immigration Work

Claude Projects are persistent workspaces inside Claude.ai. Think of them as a combination of a system prompt, a document library, and a conversation history that persists across sessions. When you open a Project, every conversation you start inherits the system prompt you wrote and has access to the documents you uploaded.

The right structure for immigration attorneys: one Project per active client matter.

What to put in the Project knowledge base (the "Instructions" section stores your system prompt; the files section stores documents):

System prompt (Instructions field): The EB-1A or O-1 legal standard, the criteria list, your output format preferences, and a constraint on what Claude should not do. The next section covers the exact template.

Permanent documents for EB-1A matters: The USCIS Policy Manual, Vol. 6, Part F excerpt you care about. The client's current resume. A criteria mapping skeleton. The Kazarian v. USCIS, 596 F.3d 1115 (9th Cir. 2010) two-step framework language you use in petitions.

Per-conversation uploads (not in the permanent library): The specific exhibits you are working with in a session. An expert letter draft you want reviewed. A USCIS RFE notice.

File limits: 30MB per file, up to 20 files per conversation. The Project knowledge base itself has no stated file count limit on paid plans. For EB-1A cases with large PDF evidence packages, compress scanned documents before upload. Most PDF scanners produce files larger than 30MB if the resolution is high; 150 DPI produces clean text extraction at under half that size.

One practical note: Projects do not replace your case management system. Treat them as a drafting workspace, not a document repository. The canonical evidence file stays in your DMS or S3-backed document system. What goes in the Project is the reference material Claude needs to produce accurate output for that case.

For the initial setup walkthrough and a confidentiality-focused configuration guide, see the AI practice confidentiality setup guide.

Writing Your System Prompt

A system prompt is the standing instruction Claude reads at the start of every conversation in a Project. Without one, you are re-explaining the context each time, which wastes tokens and produces inconsistent output. With a well-structured one, every conversation starts with Claude already oriented to the legal standard, your output format, and the constraints that matter.

Four components go into a working EB-1A system prompt:

1. Role definition. Tell Claude what it is and is not. Be specific.

You are a legal drafting assistant helping an immigration attorney prepare 
EB-1A extraordinary ability petition materials. You have no independent 
legal judgment. You draft based on evidence provided by the attorney.

2. Legal standard. Give Claude the actual regulatory text it will need, not a paraphrase.

The EB-1A standard requires the petitioner to demonstrate extraordinary 
ability through either a one-time achievement (major internationally 
recognized award) or evidence meeting at least 3 of 10 regulatory criteria 
at 8 CFR 204.5(h)(3). Under Kazarian v. USCIS, adjudication occurs in two 
steps: Step 1 (criterion counting) and Step 2 (final merits determination 
asking whether the evidence establishes sustained national or international 
acclaim and that the beneficiary is among the small percentage at the very 
top of their field). At Step 2, cite the standard: "very top of the field 
of endeavor."

3. Output format. Specify what you want back.

When drafting criterion arguments:
- Open with the regulatory language for the criterion
- State the evidence in the next paragraph, citing exhibits by number
- Analyze why the evidence meets the standard in paragraph three
- Do not use bullet points within criterion argument sections
- Length: 300-500 words per criterion unless instructed otherwise

When drafting Step 2 totality arguments:
- Include field definition, competitive population argument, 
  and "very top" standard language
- Cross-reference at least three criteria
- Target 600-800 words

4. Constraint. State what Claude should not do.

Do not state CFR section numbers from memory. If you need to cite a specific 
regulation, write [CFR CITE NEEDED] and I will verify. Do not invent exhibit 
numbers or case citations. If the evidence provided does not support a claim, 
say so rather than drafting around it.

That last component is the most important. Claude will comply with explicit constraints more reliably than it will self-police. The CFR citation constraint alone prevents a category of errors that take attorneys time to catch in review.

For a library of 40+ task-specific prompts that work with this system prompt as a base, see the immigration attorney AI prompt library.

Opus 4.8 or Sonnet 4.6: The $48/Month Decision

Both models have a 1M token context window on paid plans. Both are available in Claude.ai Enterprise via the model selector. The decision is about reasoning depth and cost.

Task	Model	Reason
Criterion 5 original contributions	Opus 4.8	Multi-step legal reasoning required
Kazarian Step 2 totality argument	Opus 4.8	Cross-criteria synthesis
RFE response narrative	Opus 4.8	Filed with minimal editing
Evidence gap analysis	Sonnet 4.6	Structured output, lower stakes
Document summarization	Sonnet 4.6	Speed more important than depth
Expert letter briefing	Sonnet 4.6	Template-like output
Cover letter outline	Sonnet 4.6	Repetitive across cases

Claude Opus 4.8: $5 input / $25 output per million tokens (API pricing). Anthropic's own legal summarization guide recommends Opus 4.8 for "high accuracy" use cases. Opus produces more complete legal arguments, handles multi-step reasoning tasks better, and is more likely to catch the places where your evidence does not fully support the claim you are trying to make. For immigration petition work, the tasks that justify Opus: Criterion 5 original contributions arguments, Kazarian Step 2 totality analysis, RFE response narratives where the specific facts matter, and any section where you are going to file what Claude produces with minimal editing.

Claude Sonnet 4.6: $3 input / $15 output per million tokens. Forty percent cheaper per token. For most iterative and supporting tasks, the quality difference from Opus is not worth the cost difference. Sonnet handles well: evidence gap analysis (what criteria do we still need to support?), document summarization, generating questions to ask an expert witness, drafting cover letter outlines, and producing structured intake checklists. Faster response times as a secondary benefit.

The practical decision for most solo and small-firm attorneys who access Claude through claude.ai Enterprise rather than the API: use Opus for drafting sessions where you plan to produce something close to final. Use Sonnet for exploratory sessions, document review, and preparation tasks. The model selector is per-conversation, so you switch as needed.

If you are working via the API directly (more likely if you are building a custom workflow or integrating with practice management software), the cost difference compounds quickly. A 10,000-word petition involves roughly 15,000 tokens of output. At Opus pricing, that is $0.375 per petition in output costs. At Sonnet, $0.225. Not a meaningful difference at one petition per week. At twenty petitions per month, you are looking at $90/month vs. $54/month in output tokens alone, before input costs. At that volume, run Sonnet for all supporting tasks and reserve Opus for the criterion argument sections you expect to use.

Eight Workflows Where Claude Saves 30+ Minutes

These are tasks where attorney time before AI took between 30 minutes and two hours, and where a well-prompted Claude session produces a usable first draft in under ten minutes.

1. Evidence gap analysis. Paste the client's document inventory. Ask Claude to map each document to the EB-1A criteria it supports and identify criteria where the evidence is thin or absent. A 45-minute review session becomes a 5-minute conversation. Output: a structured gap list your client intake interview can address directly.

2. Expert letter evaluation. Paste a draft or submitted expert letter. Ask Claude to evaluate it against three dimensions: (a) does it establish the expert's independence from the petitioner, (b) does it address the petitioner's specific contribution rather than restating their CV, and (c) does it support the Criterion 5 "original contribution of major significance" standard with specific technical language. Most expert letter evaluation used to take 20-30 minutes per letter. Claude flags the same issues in under two minutes.

3. RFE first draft: Step 2 challenge. Upload the RFE notice. Ask Claude to identify the specific Step 2 deficiency the officer cited, then draft a response narrative that addresses it using the evidence already in the record. The first draft will require fact-checking and attorney review, but starting from a structured argument saves significant time compared to drafting from scratch.

4. Criterion 5 argument, first pass. Provide the citation analysis, paper abstract, and expert letter on the contribution. Ask Claude to draft the Criterion 5 argument section. For most well-supported contributions, the Opus output requires 20-30 minutes of attorney editing rather than the 90-120 minutes it takes to draft from scratch. See the EB-1A petition guide for the evidence architecture that makes this work best.

5. Expert letter briefing document. Before a call with a prospective expert witness, ask Claude to draft a briefing document: what criterion the expert's testimony supports, what specific facts they should address, what the "major significance" standard means in concrete terms for this field, and what language to avoid. Experts who receive a clear brief write better letters. This used to take 30-45 minutes to draft per expert. With Claude, 5-10 minutes.

6. Petition cover letter. Upload the criteria summary and evidence list. Ask Claude to draft the petition cover letter in the format your supervising attorney uses, citing the specific criteria satisfied and the primary evidence for each. Cover letter drafting is repetitive across cases. Building a Claude template for it saves meaningful time at scale.

7. High-salary percentile analysis. Paste salary data from the BLS OES survey or industry compensation report. Provide the petitioner's salary. Ask Claude to calculate the percentile, identify the appropriate comparison group, and draft the Criterion 9 high remuneration argument section. This is one of the more mechanical criteria sections and Claude handles it well when the input data is clean.

8. O-1 vs. EB-1A eligibility memo. For clients who might qualify for either path, ask Claude to produce a comparative eligibility analysis based on the evidence in the record. It will not give you a legal opinion, but it will give you a structured framework covering how the evidence maps to each standard, which criteria are stronger under each petition type, and what additional evidence would strengthen the weaker path. Use this as a starting point for your own analysis, not a substitute for it.

What Claude Cannot Do Reliably

Three specific failure modes. These are not edge cases.

CFR citations from memory. Claude will produce specific CFR section numbers with confidence. Some percentage of those citations will be wrong: wrong section number, outdated version, or correct section but not for the proposition cited. The correct handling is the constraint in your system prompt: require Claude to write [CFR CITE NEEDED] rather than generate citations from training data. You verify against the current electronic CFR before filing. The EB-1A extraordinary ability criteria live at 8 CFR 204.5(h)(3) if you want to paste the actual text into your system prompt rather than relying on Claude's memory of it.

Recent USCIS policy without grounding. Claude's training data has a knowledge cutoff. USCIS issues policy memoranda, AAO decisions, and updates to the Policy Manual continuously. If you ask Claude about current EB-1A adjudication trends, what the latest AAO decisions say about Criterion 5, or how a recent policy memorandum affects Step 2 analysis, Claude's answer will reflect its training data, not the current adjudication environment. Ground it: paste the actual policy text or AAO decision into the conversation, then ask your question against that document.

Processing times and visa bulletin predictions. Claude cannot predict monthly DOS visa bulletin outcomes or current USCIS processing times. Do not ask it to. This is a category error. The data changes monthly and Claude cannot access live data without tools that are not available in standard Claude.ai sessions.

Confidentiality: Enterprise Tier Is Not Optional

The confidentiality analysis for Claude use in immigration practice runs through two rules.

ABA Model Rule 1.6 requires attorneys to make reasonable efforts to prevent the unauthorized disclosure of client information. "Reasonable efforts" in 2026 has been interpreted by multiple bar association ethics opinions to include ensuring that AI tools used with client data have appropriate contractual data handling protections. The specific requirement varies by jurisdiction. The consensus position is that consumer and small-team AI tiers, without contractual commitments on data retention and training, do not meet the "reasonable efforts" standard for client confidential information.

The BAA in Claude Enterprise is the practical mechanism that satisfies this standard for Anthropic products. It commits Anthropic to specific data handling terms: no training on your inputs, defined retention limits, and HIPAA-level security controls as the minimum baseline. The BAA must be activated by an administrator. Default Enterprise accounts do not have it active; you have to turn it on.

What the BAA does not do: it does not make Claude's outputs privileged. It does not make Claude a member of your firm's legal team. Privilege analysis still depends on the circumstances of each communication.

The Heppner ruling reinforces a point that should be obvious but often is not: AI tools are software. They do not have professional obligations, they do not hold confidences in a legal sense, and using them is the attorney's decision with the attorney's professional responsibility attached. Getting the tier right is the minimum. The attorney still needs to exercise independent judgment on every output before it goes anywhere.

For a step-by-step confidentiality compliance setup, including how to configure Enterprise BAA, what to include in your firm's AI use policy, and how to handle state-specific ethics requirements, see the AI immigration practice confidentiality setup guide.

Getting Started: Your First EB-1A Case in Claude

If you have a Claude Enterprise account and have never used it for EB-1A work, here is a concrete starting sequence.

Step 1: Create a Project for the matter. In Claude.ai, go to Projects and create a new one named with the client matter identifier you use in your DMS (not the client's full name; use your internal case number). Set the project to Private.

Step 2: Write the system prompt. Use the four-component template from the system prompt section above. Customize the output format section to match how your petitions are actually structured. If you always draft criterion arguments as three paragraphs of specific lengths, specify that. If you use a particular citation style for exhibits, specify that.

Step 3: Upload the base documents. Start with the USCIS Policy Manual, Vol. 6, Part F text (download the relevant chapters as PDF), the criteria mapping skeleton for this case, and the client's current CV. These stay in the Project and provide grounding for every conversation.

Step 4: Start with evidence gap analysis. Your first conversation should be the gap analysis workflow above. Paste the document inventory. Ask Claude to map evidence to criteria and identify gaps. This gives you two things: a quality check on Claude's output for this case, and a structured work plan for the intake phase.

Step 5: Draft one criterion argument. Pick the criterion with the strongest supporting evidence. Upload the relevant exhibits. Run the Criterion 5 or equivalent argument section. Review the output against the actual exhibits: does every factual claim have support in the document you provided? If Claude invents a fact, you have found a hallucination. Correct it, note what prompted it, and add a constraint to your system prompt to prevent recurrence.

Step 6: Iterate. Once you have drafted and reviewed two or three criterion sections, you will know where Claude is reliably useful for your writing style and where it requires more editing. Adjust the system prompt accordingly. The first case takes longer than subsequent ones; you are also building your template.

Three platforms build petition workflows on Claude's API, adding document classification, knowledge base generation, and evidence grounding on top of what vanilla Claude provides: Visalaw.AI, DraftyAI, and Immigration Copilot. If you prefer a configured product over building your own workflow from scratch, any of those three will run Claude under the hood with immigration-specific structure already built in. For the trade-off analysis between using Claude directly versus a vertical platform, see the AI tools comparison for immigration attorneys.

For attorneys who want to see Claude's output compared to ChatGPT on specific EB-1A tasks before committing to the platform, the Claude vs. ChatGPT side-by-side comparison runs four identical drafting tasks through both models with output annotations.

---

The gap between attorneys who have built a working Claude workflow and those who have not is measurable in hours per case. At ten cases per year, saving three hours per petition is thirty hours annually. At twenty cases, sixty hours. Those are numbers that matter for a solo or small-firm practice.

The setup time is real: writing a good system prompt, configuring Projects correctly, activating the BAA, and learning which tasks Claude handles well for your specific writing style takes several hours upfront. The return on that time starts compounding quickly.

For a broader overview of AI tools across the full immigration practice, the AI in legal practice hub covers ethics, tool selection, and workflow across all visa categories.

If you are ready to build the workflow, start with one case, one Project, and the system prompt template above. Sign up for Immigration Copilot if you want the document classification, evidence grounding, and petition generation built in from the start.