AI-Assisted Petition Drafting for Immigration Attorneys

Immigration attorneys who use AI for petition drafting in 2025–2026 are doing it one of two ways. Some have a system: confidentiality setup done, tool configured, prompts tested, workflow documented. They produce better first drafts and file faster. The others are pasting things into ChatGPT and hoping for the best. That approach has two failure modes: professional responsibility problems from client data exposure, and petition quality problems from unreviewed AI output.

This guide covers the systematic approach. Eight steps, in the order they need to happen.

Step 1: Confidentiality Setup

Nothing else in this guide works without this step. Configure it before pasting any client information into any AI tool.

The obligation is from ABA Formal Opinion 512 (July 2024), which applies Model Rules 1.1 (competence), 1.6 (confidentiality), and 5.3 (supervision of non-lawyers) to AI use. Rule 1.6 requires reasonable measures to prevent unauthorized disclosure of client information. Pasting a client name and A-number into consumer AI is not a reasonable measure.

What "configured correctly" means:

• Use ChatGPT Enterprise or Claude Enterprise (not free tier, not Team tier for Claude, not Plus for ChatGPT)
• Obtain and retain your Business Associate Agreement (BAA) documentation
• Disable conversation data training in your account settings
• Brief everyone in your office who will use these tools on the five identifiers that must never be pasted: A-numbers, passport numbers, Social Security numbers, full name combined with date of birth, and I-94 numbers

The anonymization rule: Before any client document or fact pattern goes into AI, substitute [BENEFICIARY] for the client name and strip all government-issued ID numbers. The AI does not need the real name or numbers to draft a petition section. For a full setup walkthrough, see AI Confidentiality for Immigration Attorneys: The Practical Setup.

Step 2: Tool Selection

The right tool depends on your firm's size, budget, and what kind of help you need.

Three categories exist: general-purpose AI configured for legal work, legal-specific AI tools, and immigration-specific tools.

General-purpose AI (Claude, ChatGPT): The starting point for most attorneys. Claude Sonnet or Opus via Claude Enterprise produces better long-form legal argument sections and handles Kazarian Step 2 totality arguments with less attorney editing. ChatGPT Enterprise (GPT-4o or o3) is comparable for shorter structured tasks and follows explicit word count instructions more precisely. Side-by-side output comparisons on four EB-1A tasks show where each model has an edge: see Claude vs. ChatGPT for Immigration Attorneys: EB-1A Output Compared.

Immigration-specific tools: DraftyAI ($79/month), Visalaw.AI ($220/month), and Visas.ai ($130/seat) provide pre-built EB-1A workflows but at higher cost. Immigration Copilot processes your client's full exhibit file before drafting, eliminating the bracket-replacement problem where you have to manually substitute real exhibit references for [EXHIBIT X]. For a full comparison including BAA availability by tier, see Best AI Tools for Immigration Attorneys 2026.

The key question before selecting: Does this tool have a signed BAA for my firm? Many tools advertise "enterprise" tiers that do not include a BAA. Verify before using it for client work.

Step 3: Intake and Eligibility Triage

The first AI task in any new case is eligibility triage: given what this client has, which criteria are plausible, which are strong, and where are the gaps?

This is not a task AI can complete on its own. It requires the attorney's judgment on which criteria to pursue. But AI can organize the question faster than you can without it.

What to feed the AI:

• The client's CV or resume (anonymized: replace name with [BENEFICIARY])
• A brief summary of their field, employer, and primary accomplishments
• Any initial evidence they have already provided

What to ask: Use a criteria-triage prompt that asks the model to evaluate each of the 10 EB-1A criteria (or 8 O-1A criteria) against the evidence summary, rate each as strong/possible/unlikely, and flag the top three to pursue.

What to do with the output: This is a first-pass filter, not a case strategy. The AI will sometimes overrate criteria where the evidence is thin and underrate criteria where unusual evidence types exist. Attorney review of the triage output is not optional.

For a complete set of intake triage prompts configured for EB-1A and O-1A work, including the system prompt that maintains legal context across the intake conversation, see 30 AI Prompts for Immigration Attorneys, Category 1.

Step 4: Evidence Mapping and Gap Analysis

After triage, the attorney has a target set of criteria. The next task is mapping which existing evidence supports each criterion and identifying what is missing.

AI does this well if you give it structured input. The evidence mapping prompt takes a list of documents the client has provided and asks the model to map each to the criteria it could support, with a brief rationale. The output is a structured table: document type, relevant criterion, strength of fit, and what additional evidence would strengthen the argument.

The gap analysis: After the mapping, ask separately: "For each of the three criteria we are pursuing, what additional evidence types are commonly used and are absent from the current evidence set?" This produces the gap list the attorney needs to advise the client on what additional materials to obtain before filing.

The full evidence mapping prompt and gap analysis prompt are in the AI prompts guide, Category 2. For how to structure evidence architecture to reduce RFE risk, see EB-1A Evidence Architecture for RFE Prevention.

Step 5: Drafting Criteria Sections

This is where AI saves the most time per petition.

A criteria section draft starts with the attorney providing: the legal standard for that criterion, the specific evidence the client has (with exhibit references), and any expert letter excerpts relevant to this criterion. The model produces a 300–600 word draft argument. That draft is a starting point, not a filing.

What AI does well here:

• Produces correct CFR citation format on the first try when the system prompt includes the regulation
• Includes the legal argument structure (standard → evidence → analysis) without prompting
• For Criterion 5, Claude produces the before/after field analysis and citation-versus-adoption distinction more reliably than ChatGPT

What AI consistently gets wrong:

• Employer-specific framing on Criterion 5 (the contribution must be field-significant, not employer-significant)
• Vague superlatives without comparators ("among the leading researchers" without a percentile or population)
• Generic characterizations of contributions that describe what the work is, not why it mattered to the field

The edit workflow: For each criteria section, the attorney's review task is specific: verify every exhibit citation, replace every vague superlative with a specific comparative data point, and confirm that the field-significance argument is present (not just a description of the work). This is a 15-minute edit task, not a rewrite.

Criterion-specific failure patterns: Each criterion has its own AI default error.

Criterion 5 (original contributions): the employer-specificity trap. AI will frame a company-changing product as a field-changing contribution. A product that changed how one company's engineering team works is not a field-significant contribution. The fix is explicit: add "the contribution must be significant to the field at large, not to the beneficiary's employer" to every C5 prompt.

Criterion 9 (high salary): AI handles this well when given Bureau of Labor Statistics OES wage data, but will sometimes pull from an employer's internal compensation study rather than the national field comparator. The national comparator is what the criterion requires. Always specify it explicitly.

Criterion 8 (critical role in a distinguished organization): AI frequently conflates the two-prong argument into one. The organization's distinction must be established first, independently, with external evidence. The beneficiary's critical role within it is the second prong. Merged into one paragraph, the distinction disappears and the argument weakens. These should be two separate sections in the draft.

For O-1A work: the eight O-1A criteria under 8 CFR 214.2(o) follow different definitions than EB-1A. The O-1A salary criterion has a different comparator standard. Include the correct O-1A regulatory language in the system prompt separately from your EB-1A system prompt.

For the complete prompt library across all 10 EB-1A criteria and 8 O-1A criteria, see 30 AI Prompts for Immigration Attorneys, Category 4. For a detailed look at where AI output diverges between Claude and ChatGPT on criteria sections, see Claude vs. ChatGPT for Immigration Attorneys.

Step 6: Expert Letter Workflow

Expert letters are the highest-stakes AI task in any EB-1A petition. They are also where AI errors are hardest to catch, because the errors look like legitimate legal writing.

The workflow has two parts: briefing the expert and drafting the letter.

Briefing the expert means producing the memo you send to the expert before they write. A good briefing memo explains the legal standard in plain language (not attorney jargon), names the specific contribution to address, specifies what evidence the letter should cite, and tells the expert what the letter should not include (credential-based praise, vague characterizations of the field). AI turns this from a 60-minute writing task into a 15-minute editing task. The briefing memo is lower-risk: if the briefing is imprecise, the expert will write a different letter anyway.

Drafting the letter means producing a complete letter for the expert to review and sign. This is higher-risk. Four AI errors appear consistently in expert letter drafts, and they persist if the expert rubber-stamps the draft without careful review:

1. Credential bleed: the expert's credentials appear inside the factual argument about the beneficiary's contributions, rather than in the bio section where they belong
2. Generic superlatives: "among the leading researchers" without a denominator or comparator
3. Hedge closers: "I believe this work may have been influential" instead of a declarative expert opinion
4. Vague field definitions: "the field of computer science" instead of "the subfield of LLM training efficiency research"

Each of these is detectable with a quick grep-style review. Before any AI-drafted letter goes to an expert, run the four-item checklist. A letter that passes the checklist is ready for expert review. A letter that fails goes back to the attorney edit stage before the expert sees it.

For the full briefing prompt, drafting prompt, and the complete review checklist, see AI-Drafted EB-1A Expert Letters: What Works and What USCIS Flags.

Step 7: Kazarian Step 2 and RFE Response Drafting

Step 2 of the Kazarian framework (the final merits determination) requires the attorney to argue that the beneficiary, considered in totality, has achieved sustained national or international acclaim at the very top of the field. This is a synthesis task. It requires pulling together evidence from multiple criteria into a single cumulative argument.

AI does this better than almost any other petition task. The reason: Step 2 does not require looking up specific regulatory language or exhibit citations the model cannot verify. It requires organizing and framing evidence that the attorney has already assembled. Given a complete list of Step 1 criteria met, the key data points for each, and the field-definition statement, the model produces a coherent Step 2 argument that the attorney can then refine.

What a complete Step 2 argument needs: Four components are required. First, the field definition: narrow enough to support a comparative argument ("transformer-based interpretability methods for large language models"), not the parent discipline ("machine learning"). Second, the evidence base: each Step 1 criterion with its key data point pulled together in one paragraph (citation percentile, salary percentile, award selectivity, expert testimony). Third, the comparative argument: where the beneficiary sits relative to the defined population, not just a list of their accomplishments. Fourth, the declarative "very top" claim: a direct statement that the totality places the beneficiary at the very top of this field, not merely above the median.

AI (Claude specifically) produces all four components if you provide the evidence base. The most common attorney edit on Step 2 drafts is tightening the field definition. AI broadens it to sound more impressive. Broader fields mean a larger comparator population. A larger population makes the "very top" claim harder to support. The attorney's job is to push the field definition back to the specific subfield where the evidence is strongest.

If an RFE arrives: AI-drafted RFE responses save the most time on the research phase. The attorney inputs the RFE language, the existing petition evidence, and any new evidence the attorney intends to add. The model drafts the response to each specific officer concern. The attorney's review task on an RFE response is more intensive than on initial criteria sections, because RFE responses are arguing against a specific officer's stated deficiencies.

For RFE response prompts and a worked example of a Step 2 challenge response, see 30 AI Prompts for Immigration Attorneys, Category 5. For patterns from recent USCIS RFEs and what officers are flagging, see EB-1A RFE Response Guide.

Step 8: QA Review Before Filing

The last step is also the one most often skipped. Every AI-drafted petition needs a systematic QA pass before filing.

This is not a proofreading pass. It is a fact verification pass.

What to check in QA:

Exhibit citations: Every (Exhibit X) reference must point to an exhibit that exists in the packet, contains what the petition claims it contains, and uses the correct exhibit number. AI will sometimes cite exhibits that were mentioned in the input but numbered differently, or invent exhibit numbers entirely.

Factual claims: Every specific claim (percentile figures, citation counts, dollar amounts, publication venues) must be verified against the actual exhibit. The model will occasionally paraphrase numbers (writing "approximately 200 citations" when the exhibit shows 214) or pull numbers from context rather than the specific document.

Legal standard compliance: Every criteria section must pass the employer-specificity check for C5 (is the contribution framed as field-significant, not just employer-significant?) and the independence check for expert letters (does the letter establish the expert's independent basis for evaluation?).

Step 2 field definition: The field named in the Step 2 argument must match the field named in the criteria sections and the expert letters. Inconsistent field definitions across the petition give USCIS grounds to apply a broader comparator population than the petition intends.

The AAO pattern check: Before filing, cross-reference the petition's three weakest exhibits against recent USCIS Administrative Appeals Office decisions. If you find a non-precedent AAO decision that discounted the same type of evidence on the same criterion, add a sentence distinguishing your evidence from the pattern the AAO flagged. Officers do not always cite AAO decisions in their reasoning, but they read them. A petition that pre-addresses the pattern a recent denial identified is harder to deny on that basis.

The consistency check: Read the petition in sequence: criteria sections, expert letters, Step 2 argument, cover letter. Every data point that appears in more than one section must be stated identically. A citation count of "214 independent citations" in the criteria section that becomes "over 200 citations" in the cover letter creates an inconsistency USCIS will note.

For a complete QA approach covering the petition's evidence architecture, see EB-1A Evidence Architecture for RFE Prevention. For what USCIS adjudicators are looking for in the QA-sensitive sections, see USCIS AI Adjudication of EB-1A Petitions: What Attorneys Need to Know.

The Complete Workflow in Practice

An attorney who has completed the setup steps (confidentiality configured, tool selected, system prompt written) runs through these eight steps per petition:

Phase	AI Task	Time with AI	Time without AI
Step 3: Intake triage	Criteria scoring prompt	30 min	90 min
Step 4: Evidence mapping	Document-to-criteria mapping	20 min	60 min
Step 5: Criteria drafting	Section-by-section drafts (10 criteria avg)	3–4 hours	15–20 hours
Step 6: Expert letter briefings	Briefing memo per expert (6–8 avg)	2 hours	8–12 hours
Step 6: Expert letter drafts	First draft per letter	1 hour	6–9 hours
Step 7: Step 2 / cover letter	Totality argument draft	45 min	3–4 hours
Step 7: RFE response (if needed)	Response draft per RFE concern	1 hour	4–6 hours
Step 8: QA review	Exhibit verification (unchanged)	4–6 hours	4–6 hours

The time savings from a well-configured AI workflow are roughly 60–80% on drafting tasks. QA review does not compress. The risk of skipping QA or reducing QA time because the AI draft looks clean is the primary petition risk in an AI-assisted workflow.

For a real case example of this workflow applied end-to-end on a EB-1A petition, see EB-1A Petition Drafting: From 200 Hours to 40.

Where to Start

The single highest-return first step is the system prompt. Before running any petition task through AI, write a system prompt that includes: the EB-1A legal standard from 8 CFR 204.5(h), the Kazarian two-step adjudication framework from the USCIS Policy Manual, Vol. 6, Part F, Chapter 2, the criteria list you most frequently encounter, and your preferred exhibit citation format. Store it in a Claude Project or ChatGPT Custom GPT. That system prompt does more to improve AI output quality than any other single step.

After the system prompt, the highest-value task to test first is criteria section drafting on a case you have recently completed. Run the criteria section prompt on a section you already know the answer to. Compare the output to what you filed. The gaps you find will tell you exactly what your review checklist needs to catch.

For the tested prompts to start with, see 30 AI Prompts for Immigration Attorneys. For the model comparison on those first drafts, see Claude vs. ChatGPT for Immigration Attorneys. For the safety and ethics background, see Is AI Safe for Your Immigration Practice?.

Immigration Copilot handles the full workflow in this guide on a single platform: document classification, evidence-to-criteria mapping, AI-drafted petition sections with exhibit citations sourced from your client's actual uploaded documents, and expert letter briefing generation. The bracket-replacement problem (where every AI draft requires substituting [EXHIBIT X] placeholders with real exhibit references) is solved at the document ingestion stage, before any drafting begins.