AI for H-1B petition preparation means using software to read documents, fill forms and draft letters for a cap-subject petition. An attorney still checks every output before filing. The software is most useful where a reviewer can compare its result with a document, such as a passport number or a date of birth. Where the answer depends on the job itself, the lawyer decides, and recent rules make that group larger. Weighted selection and the modernization rule both put more job facts into the filings. This guide is mostly about where the line between the two groups falls.
The guide starts with how a cap petition is built and what changed in 2025 and 2026. It then sorts H-1B tasks by how checkable the AI output is. The hard cases follow: consistency across filings, the support letter and RFE responses. It closes with the professional rules, a review routine and a pilot test. Legal citations link to the Federal Register, ABA and EOIR texts, and the firm-level advice is our own.
Key Takeaways
- Software for H-1B work reads documents, drafts text or checks fields, and each type needs a different kind of review.
- The registration, the labor condition application, Form I-129 and the support letter must describe the same job.
- Under the July 2026 signature rule, the petitioner signs Form I-129 itself, and the lawyer answers for every AI output under ABA Formal Opinion 512.
- A pilot on closed files shows how accurate a tool is before it touches a live petition.
How an H-1B Cap Petition Is Built
Congress caps initial H-1B grants each fiscal year. INA section 214(g) limits the total to 65,000, plus up to 20,000 under the advanced-degree exemption, as DHS recites in its weighted selection final rule. A cap case moves through three filings:
- Registration. The employer registers each requested worker electronically, and only selected registrants may file a petition. For FY 2027, AILA's cap-season page lists a registration period of March 4 to March 19, 2026. Selected petitioners could file beginning April 1, 2026.
- Labor condition application (LCA). The employer obtains a certification from the Department of Labor, as the Wage and Hour Division describes. On it, the employer promises to pay the higher of two figures. One is the actual wage paid to similarly employed workers, and the other is the prevailing wage for the occupation in the area.
- Form I-129. With a certified LCA in hand, the employer files the petition with USCIS. The filing carries evidence that the job is a specialty occupation and that the worker qualifies.
That evidence typically includes an employer letter describing the position, a description of the duties, proof of the worker's degree and any credential evaluation. Work at a client site adds documents from the end client. The mix depends on the case, and this is the part of the file that AI tools are most often asked to help draft.
What Changed in 2025 and 2026
Three recent rules changed what goes into those filings and how they are signed. Weighted selection and the modernization rule affect the content of the petition. The third, USCIS's signature rule at 91 FR 25479, took effect July 10, 2026 and comes up again in the section on professional rules.
Weighted Selection
DHS published the weighted selection final rule at 90 FR 60864 on December 29, 2025. It took effect February 27, 2026, in time for the FY 2027 registration season. Each registration enters the selection pool according to its wage level:
- Level I enters once.
- Level II enters twice.
- Level III enters three times.
- Level IV enters four times.
DHS sets the level at the highest Occupational Employment and Wage Statistics (OEWS) level that the offered wage meets or exceeds. The level is read for the SOC code in the area of intended employment. DHS says the aim is to incentivize employers to offer higher wages or to petition for higher-skilled positions.
That design puts job facts, such as the occupation code and the offered wage, into the registration itself. According to Faegre Drinker's January 2026 summary, registrations must include the SOC code, the area or areas of intended employment and the OEWS wage level. That data must be consistent between the registration and any petition filed. The final rule has two sections on this point, titled "Consistency Between the Registration and the Petition" and "Certifying the Contents of the Registration." Read both in full before you set a firm standard.
The H-1B Modernization Rule
The H-1B modernization rule, published at 89 FR 103054 on December 18, 2024, took effect January 17, 2025. Seven of its provisions bear directly on how a petition describes the job:
- A position may allow a range of degree fields if each is directly related to the duties. "Directly related" means a logical connection between the degree or its equivalent and the duties of the position.
- The criteria in 8 CFR 214.2(h)(4)(iii)(A) use the word "normally," which the rule clarifies does not mean "always."
- The petitioner must show a bona fide position in a specialty occupation as of the start date. USCIS may request contracts or similar evidence.
- For work at a third party's site, the third party's requirements are the most relevant to whether the position is a specialty occupation.
- The labor condition application must properly correspond to the petition.
- USCIS generally defers to a prior eligibility determination involving the same parties and facts, unless a material error in the prior approval is discovered.
- USCIS may conduct site visits, and refusing one may result in denial or revocation.
Most of these provisions concern what the petition says about the job, which is the text a drafting tool writes. The next section therefore sorts H-1B tasks by how checkable the tool's output is.
Where AI Fits in an H-1B Workflow
Sorting tasks starts with the tools, because software sold as AI for this work does three different jobs. Each fails in its own way, and the way it fails decides how much review its output needs.
Three Kinds of Tools
- Extraction tools read a document and return values, such as a passport number or a date of birth. Their typical error is a wrong value, which a reviewer finds by comparing it with the source.
- Drafting tools write new text, such as a support letter or a cover letter. Their typical error is a fluent statement that no document supports, which is harder to see because it reads well.
- Checking tools compare fields across documents and flag empty fields, missing signatures or conflicting dates. Their typical error is silence, because a tool flags only the conflicts it was built to look for.
Most products combine two or three of these, so a firm should ask which part of a workflow each one handles. The review effort follows the error type, and the review routine later in this guide uses the same split.
Task-by-Task Guide
With those error types in mind, the table below sorts common H-1B tasks by how checkable the output is. It is our assessment, not a measured result. Start with the Review Level column, which says who must sign off before the work moves on.
The three review levels mean different things:
- Under standard review, a trained paralegal compares the output with its source.
- Under attorney review, an attorney reads the output against the file and the rules.
- Under attorney decision, the attorney makes the call, and any tool output is background.
Keeping One Job Consistent Across Four Documents
The rows marked attorney review or attorney decision share a theme. The registration, the labor condition application, the Form I-129 and the support letter must all describe the same job. The weighted lottery depends on that match at registration, and the modernization rule requires the LCA to properly correspond to the petition. Before a petition is signed, reconcile these facts across all four documents:
- Employer legal name and address
- Beneficiary name and date of birth
- Occupation and SOC code
- Worksite addresses, including any third-party site
- Offered wage and wage level
- Start and end dates
- Job title and the duties listed under it
- Required degree field
Each document carries a different part of the load. The registration fixes the wage level that weights the lottery, and the LCA commits the employer to a wage and a worksite. The support letter explains the duties, so a mismatch with the other three shows up there first.
An Illustration
Consider a hypothetical software developer role, offered as an example and not a real case. The registration and the LCA use an occupation and wage level that match the duties in the employer's position description. A drafting tool then writes the support letter from the employer's website, which says its engineers lead architecture across the company. The letter credits the developer with duties above the wage level the firm chose, and the employer never confirmed those duties.
The attorney catches the problem by reading the letter against the LCA and the position description. The fix is to confirm the real duties with the employer and rewrite the letter from them. No tool in the workflow could have settled the question, because the answer sits with the employer.
Writing the Support Letter and Answering an RFE
The support letter is where the consistency chain meets the modernization rule's "directly related" test, so it earns the closest review of any AI-drafted document. The same discipline carries over to an RFE response. This section takes the letter first and the RFE second.
Tying Duties to Degree Fields
The test asks for a logical connection between each duty and the degree field or fields. A drafting tool can propose those connections, and the attorney confirms them duty by duty against the position description, as in this illustrative developer role.
The rule allows more than one field only when each is directly related. List a field in the letter only if the duties connect to it. A letter that only restates the legal definition of a specialty occupation without tying it to these duties leaves the connection unproven.
Client-site work adds a second task. The rule makes the end client's requirements the most relevant. The tool should therefore list the documents to request from the client and flag drafts that rely on the petitioner's own requirements. Any statement about available work also needs contracts or similar evidence behind it, since USCIS may request them.
RFE Responses
An RFE is a written request from USCIS for more evidence before it decides a petition. AI helps most at the start of the response. It can:
- List each point the notice raises.
- Match evidence already in the file to each point.
- Draft an outline for the attorney.
The rest of the response calls for judgment about strategy and new evidence. Every statement it adds has to fit the same consistency chain as the original petition. Any case or policy citation in a draft needs to be opened and read by a person, for reasons the next section explains.
Rules That Govern AI Use on a Petition
The drafting and review tasks above sit inside professional-conduct rules, and those rules say who answers for the output. The sections below cover the ABA's generative AI opinion, the EOIR memo on hallucinated content, USCIS's signature rule and California's SB 574.
ABA Opinion 512 and State Opinions
ABA Formal Opinion 512, issued July 29, 2024, applies the existing Model Rules to generative AI:
- Under Rule 1.1, lawyers need to understand the benefits and risks of the tools they use.
- Under Rule 1.6, lawyers must keep client information confidential unless the client gives informed consent.
- Under Rule 1.5, a lawyer may bill for time spent entering information into a tool and reviewing the draft for accuracy and completeness. Learning the tool is generally not billable.
State bars add their own opinions, and Florida Ethics Opinion 24-1 (2024) says lawyers must protect the confidentiality of client information when using generative AI. Check your state bar's current guidance before you write a firm policy. Our guide to attorney oversight of AI in immigration covers the oversight side in more detail.
The EOIR Memo on Hallucinated Content
EOIR Policy Memorandum 25-40, "Use of Generative Artificial Intelligence in EOIR Proceedings," was issued August 8, 2025. The memo does not ban generative AI. It calls hallucinated legal citations or arguments the greatest potential risk. It says practitioners who submit hallucinated or erroneous AI content likely violate their conduct obligations. The memo is written for EOIR proceedings, so by its terms it does not govern USCIS petitions. Lawyers who file in both places can still adopt its verification expectation as a firm-wide habit.
Signatures on the Petition
USCIS published its interim final rule on signatures on immigration benefit requests on May 11, 2026. It requires a valid signature on every benefit request, and its preamble treats a handwritten signature from the requestor as the general rule. Electronic signatures are accepted only in limited settings such as guided e-filing. Its invalid examples include:
- Signatures created by signature software.
- Signatures made by someone other than the requestor, such as an attorney.
If USCIS accepts a request and later finds the signature invalid, it may reject or deny it, and a denial lets USCIS keep the fee. In an H-1B workflow, the petitioner signs the Form I-129 itself, by hand or through a filing method the rule accepts. No packaging tool should place that signature for it. Form instructions can allow narrow exceptions, so check the current Form I-129 instructions.
Client Data and California SB 574
Passport details, I-94 records and Social Security numbers fill an H-1B file, so where a tool stores and processes them matters. California SB 574 (Chapter 858, Statutes of 2026) was signed September 30, 2026 and takes effect January 1, 2027, according to Ropes & Gray.
A Holland & Knight summary says new section 6068.1 of the Business and Professions Code bars lawyers from entering confidential, nonpublic or personally identifying information into a generative AI tool. The one exception is a system whose access is restricted to the attorney and authorized persons who are themselves under confidentiality obligations. Our guide to immigration software data security lists what to ask a vendor about storage and access.
A Review Routine for AI-Assisted Petitions
Those rules point to a review routine, and the routine works best when each check targets a known error type. The table lists the six error types we see as the main risks in AI-assisted H-1B work, with the check and the owner for each.
Run the transcription, invention and reuse checks first, then the drift check. The attorney of record finishes with a review of the legal decisions before the petitioner signs. Keep a simple count of corrections by error type on each file. Pause a tool on a petition type if its drafts need substantial rewriting on more than one petition in five. Resume when the template or intake questions are fixed.
Test a Tool Before You Trust It
The same error types give a firm a way to test a tool before it touches a live petition. A vendor demo shows a tool at its best, so a pilot on closed files whose outcome the firm already knows is more informative. Set the go or no-go thresholds before you run it.
- Select five to ten closed petitions that include different situations, such as a third-party placement, a lower wage level and a prior RFE. Remove client identifiers under your firm policy or get consent to use the files.
- Run the tool from the same intake data the firm used.
- Compare each output with the version the firm filed, line by line, and tag every difference with an error type from the table above.
- Add one trap, such as a position description whose duties conflict with the wage level. See whether the tool flags the conflict or repeats it.
- Record the reviewer time per file for the tool-assisted draft and for the original work.
- Decide go or no-go against the thresholds, such as zero uncorrected transcription errors in the final forms and an agreed ceiling on invented statements.
Alongside the pilot, ask each vendor four questions:
- Where are drafts stored, and who can open them?
- Does the vendor use client data to train models?
- Does the tool show the source document behind each filled field?
- Does an attorney approve every page before it leaves the system?
US Immigration AI's AI Legal Document Drafting works from attorney-approved templates and pulls client details from intake into the forms. It flags empty fields, missing signatures and conflicting dates for attorney review. For the form-filling step on its own, see how the platform can auto-fill USCIS forms from intake data.
What the Evidence Says About Time Saved
The pilot also produces the numbers that published sources lack. The research for this guide covered vendor pages, bar-press articles and survey reports through October 2026. It found no independent study that measures hours saved or RFE-rate changes from AI on H-1B petitions. The time-saving figures on vendor pages give no method, so they cannot be compared across products. For the wider picture of AI in a practice, see our overview of how immigration lawyers use AI.
A firm can build its own evidence by tracking three things across at least one cap season:
- Reviewer hours per petition, before and after adoption.
- Corrections per file, by error type.
- RFE rates, by petition type.
Grow Your Caseload With US Immigration AI
US Immigration AI drafts H-1B petitions and supporting letters from your intake data and attorney-approved templates, then holds every file for lawyer review. If you are planning for the next cap season, you can see the workflow on a short call. Schedule a demo to walk through it with the team.
Frequently Asked Questions
