Can AI Help Immigration Law Firms Answer an RFE? What to Check

Yes, for the organising and first-draft work, but not for the judgement. AI can turn a request for evidence into an issue list, match your exhibits to each issue and draft response sections. A lawyer must verify every fact, citation and exhibit against the file, and sign the response.

The danger is fluent writing. A confident paragraph that cites a case for something the case does not say, or states a fact the client never gave you, reads exactly like the rest of the draft. A request for evidence response is a filing on which a client's status can turn, so the process has to make wrong statements visible before they leave the firm.

Follow me on Instagram@sagnikteaches

Read the notice first, with the deadline in hand

Before any tool is opened, a person reads the notice itself. The regulations and the adjudicating agency's own policy manual are clear on a few points that shape everything below. The deadline is printed on the notice, and the longest response period an officer can set is 12 weeks. The response must be received by that date (an online response counts as received on the day it is submitted), so posting a package on the last day is not enough. Everything requested goes in together, in one package, with the original notice. A partial response is treated as a request for a decision on the record as it stands, and the officer will not wait for a second instalment. Extra time to respond cannot be granted. The notice and the current rules are the authority for your file, so check them rather than a summary like this one.

Connect on LinkedInSagnik Bhattacharya

That matters for AI use for one reason: the time you save should go into verification, not into a later filing date. Put the printed deadline in the matter calendar, then work backwards: partner review at least three working days before, client sign-off on declarations a day or two before that, exhibit assembly before that. An AI draft that arrives early only helps if the checks still fit.

Subscribe on YouTube@sagnikteaches

The five-step workflow

1. Issue list. Break the notice into numbered issues with the exact language the agency used.
2. Evidence matrix. Map every exhibit already in the file, and every document you can still get, to each issue.
3. Gap list. Mark issues with thin or missing evidence, and decide what to request from the client or third parties.
4. Draft. Write each section from the matrix, using only material in the file.
5. Verification. A lawyer checks every fact, cite and exhibit reference, and signs off.

AI helps with steps one to four. Step five is the lawyer's, and no software substitutes for it.

Step one: the issue list

RFE notices often run to several pages, with legal standards recited in boilerplate and the actual issues buried in a few sentences. An assistant is good at pulling those out. Prompt it narrowly and tell it not to answer the issues yet.

Below is the text of a request for evidence. Do not draft a response.
1. List each distinct issue the agency raised as a numbered item, quoting its own words.
2. For each, list the specific evidence types the notice says would resolve it.
3. Flag any sentence you could not classify.
4. Give the response deadline as printed on the notice, quoted exactly.
[notice text, identifiers removed]

A realistic output (illustrative) for a notice on an employment-based petition might return three items: whether the role qualifies as a specialised position, whether the employer has shown ability to pay, and whether the beneficiary's qualifications match the role. It might add a note that sentence 14 is ambiguous. What you fix: compare each item with the notice paragraph by paragraph. Models sometimes merge two issues or split one into two, and a missed second ground in a paragraph is the single most expensive error at this stage.

Step two: the evidence matrix

An evidence matrix is a table: issues down the side, exhibits across the top, a mark where an exhibit speaks to an issue. Firms do this on paper or in a spreadsheet already. AI makes it quicker when the file is large, because it can read the exhibit list and the documents and propose the mapping.

Give the assistant exhibit text, not just titles. A title such as "Letter from CFO" tells it nothing about whether the letter addresses ability to pay. Ask it to quote the line in each exhibit that supports the mapping, so you can verify the match in seconds rather than rereading the exhibit.

What goes wrong in matrices

The common failure is an exhibit marked "strong" because its title sounds right. A bank statement exhibit is listed under ability to pay, but the statement is for a different entity from the petitioner. The mapping looked fine; the entity names did not match. Always add a column for entity, date range and identifiers, and check those, not only the topic.

Step three: the gap list drives client contact

Turn weak and empty cells into a request to the client. The assistant can draft a plain-language document request, which a lawyer reviews before sending. Keep it short and specific: what document, why it is needed, by when, and who can supply it. The gap list is also where you decide whether a witness statement is needed or a third-party letter, and which of those you can realistically obtain before the deadline.

From the matrix below, write a client email in plain English (reading age 12) asking for the
documents marked as gaps. One bullet per document: what it is, why we need it, where
they might get it. Do not mention legal standards. Use [first name] and [your name].
[matrix rows with gaps]

Check the draft for any promise about outcomes. Assistants like to reassure ("this will strengthen your case"). Delete anything that sounds like a forecast.

Step four: drafting from the matrix

Draft one issue at a time. A single prompt for the whole response produces a smooth document in which you cannot trace any statement to its source. Instead, give the assistant the issue, the relevant exhibits, and the authorities your firm has already pulled, then require exhibit references for every factual sentence.

Draft the response section for Issue 2 (ability to pay).
Use ONLY the exhibits pasted below and ONLY the authorities in the list below.
After every factual sentence, add the exhibit number and page in brackets, for example (Ex. 12, p. 4).
If a needed fact is not in the exhibits, write [GAP: description] instead of stating it.
Do not cite any case or policy that is not in the list.
Tone: formal, concise, no rhetorical questions.
[Issue 2 text, exhibits, authority list]

Illustrative output: three paragraphs, each factual sentence ending in a bracketed exhibit reference, with one [GAP: Q4 payroll total] flag. That is what you want, because every sentence is now checkable and every missing fact is visible. A drafter who gets no flags should be suspicious rather than pleased.

The authority list is yours

The most serious AI failure in legal drafting is the invented or misdescribed citation. The remedy is structural. Build the authority list yourself from the agency's policy manual, the regulations and the decisions you have read, and tell the assistant it may cite nothing else. If the draft cites anything outside the list, delete the sentence and note it as a defect in the process. Then open each listed authority and confirm it says what the draft says it does. Courts and tribunals have sanctioned lawyers for filing fabricated or misdescribed authorities: the public AI Hallucination Cases database kept by the researcher Damien Charlotin listed more than 2,000 such decisions by November 2026, including monetary sanctions and referrals to bar regulators. Treat it as a professional-conduct matter, not a question of style.

A worked RFE: a four-person firm, 40 exhibits, 9 working days

Take an illustrative firm of four fee earners. A notice arrives on an employment-based petition with three issues, a file of 40 exhibits, and a printed deadline 9 working days away. Here is how the days might split.

Day one (about 3 hours). A senior lawyer reads the notice twice and writes the issue list herself in 20 minutes. She also runs it through the assistant as a cross-check; the tool splits issue 1 into two items, which she accepts after rereading the paragraph. The deadline goes into the calendar with three reviewing milestones.

Day two (about 4 hours). A paralegal builds the exhibit index with entity, dates and page counts, and the assistant proposes the matrix. The paralegal checks every "strong" mark against the exhibit text and downgrades two. The gap list has seven items.

Days three and four. The client request goes out on day two evening, with the plain-English email above after a lawyer edits it. Two of seven items come back on day four; the rest need a call. The lawyer decides to commission a short third-party letter for issue 3 and starts that straight away, because it has the longest lead time.

Days five to seven. Drafting, one issue per prompt. Each section is about 400 to 600 words with exhibit references. The lawyer rewrites the opening of each section herself, because the opening is where the argument lives. Roughly a third of the AI text survives unchanged, a third is edited, and a third is replaced.

Day eight. Verification, described below, takes about 4 hours for 40 exhibits.

Day nine. Final read, signatures on declarations, assembly of the package with the original notice included, as the rules require, and dispatch with tracking early enough to be received before the printed date, or upload through the online account where the case allows it.

A rough sum, illustrative: a firm that spends 22 hours on a response like this by hand might spend 14 to 16 with this workflow. The saving sits in the issue list, the matrix and first-draft prose. The verification hours do not shrink, and should not.

Step five: the lawyer's checklist

Use this as a sign-off form, completed and kept in the matter file. It converts "someone looked at it" into a record.

• Every factual sentence has an exhibit reference, and the lawyer has opened the exhibit and confirmed the fact. Names, dates, amounts and entity names match.
• Every authority has been opened and read for the proposition it is cited for. No authority outside the list.
• Every issue in the issue list has a section, and the notice has been reread against the response for any ground the issue list missed.
• Every declaration or statement has been read line by line with the person signing it, and the signer has corrected what is wrong.
• No unfilled [GAP] markers remain, and no bracketed AI instruction text survives into the filing.
• Inconsistency check: dates of employment, salary, job title and address agree across the response, the original filing and the exhibits.
• Exhibit packaging: every exhibit is legible, complete, translated with a translator's certification where needed, and paginated as the index says.
• Deadline: dispatch method and expected arrival are recorded against the printed date.

Four quick examples of where the checks earn their keep

The entity mismatch. A matrix lists a bank statement under ability to pay. The statement belongs to the founder, not the petitioning company. The check catches it, and the gap list gets a new line.

The softened fact. A draft says the beneficiary "managed a team of about ten". The reference letter says "supervised a group of up to six". Nothing in the file says ten. The sentence is corrected to match the letter, since an overstatement in a response is a problem for the client, not just the lawyer.

The wrong date. A draft says the employer was founded in 2014. The incorporation certificate says 2016, and the website says 2014 because of an earlier trading name. The lawyer asks the client and writes both facts correctly.

The helpful paragraph. The assistant, asked to "strengthen" a section, adds a paragraph arguing that the employer's industry is growing. There is no exhibit behind it. The lawyer deletes it; unsupported assertions weaken a response and sit badly beside the verified material.

Two awkward jobs: the cover letter and the exhibit list

Two parts of the package are tedious and well suited to a supervised first draft. The first is the cover letter, which tells the adjudicator how the package is organised. The second is the exhibit index, which has to agree with the pages behind it.

For the cover letter, give the assistant the issue list and the final exhibit order and ask for a one-page roadmap: each issue, the exhibits that answer it, and nothing argued. A sample result (illustrative) reads: "Issue 1 is addressed at Section A and Exhibits 4, 9 and 15. Issue 2 is addressed at Section B and Exhibits 12, 13 and 21." What you fix: exhibit numbers. They shift whenever someone adds a late document, so regenerate the roadmap after the final order is fixed, and compare every number to the index.

For the index, paste the exhibit titles with page counts and ask for a table with a one-line neutral description for each. Neutral matters. "Letter confirming the beneficiary's duties from 2019 to 2022" is fine; "Powerful letter proving the role is specialised" is argument, and does not belong in an index. A before and after shows the difference:

Before (assistant's first pass): Ex. 9 - Compelling organisational chart demonstrating the complexity of the role
After (lawyer's edit):          Ex. 9 - Organisational chart of the petitioner, dated 3 March, 2 pages

Then count the pages. If the index says 212 pages in total and the compiled file is 209, find the three missing pages before anything else happens. A missing page in an evidence package is exactly the sort of fault that can sink a response, and it is a mechanical check that a person with the printed file can do in ten minutes.

Translations deserve a separate note. Where an exhibit is in another language, the rules require a full translation into English, with the translator's certification that it is complete and accurate and that they are competent to translate. Do not use a general chatbot's translation as the certified one. It can help a lawyer skim what a foreign document says so the matrix is accurate, but the translation that goes in the file is a certified human one, and the lawyer should say in the matter notes which version was used for which purpose.

Confidentiality and client consent

An RFE file holds passports, tax records, medical or family details and financial statements. Before any of it goes into an AI product, settle three things in writing. First, the tool: a business plan with a data agreement and no training on your content, not a personal account. OpenAI and Anthropic both state that their business tiers do not train on customer content by default; on a consumer plan the training setting is something you would have to find and switch off, which is one more reason to keep client files away from personal accounts. Second, minimisation: redact identifiers the task does not need, such as passport and receipt numbers, and replace names with labels in drafting prompts. Third, consent: update the engagement terms so clients know AI-assisted tools may be used under lawyer supervision, and your professional-conduct rules on confidentiality and supervision still apply in full.

For more on redaction and letters, see using ChatGPT for evidence letters and redaction, and the stage-by-stage guide to tools for where case-management software already includes drafting help. If your firm is deciding on software rather than a general chatbot, Docketwise or ChatGPT for forms compares the two approaches. The buyer's checklist for case software helps if automation is the goal.

How to check the process is working

After each response, record the hours spent per stage and the number of defects found in the verification stage that originated in the AI draft (wrong fact, unsupported claim, bad citation). If the count is not falling over five or six responses, tighten the prompts: shorter sections, stricter authority lists, exhibit text instead of titles. If the lawyer's verification time is falling because the drafts are traceable, that is the real saving.

Keep a one-page note on what the notice outcomes were. Whether a response succeeded depends on facts and law, not on tools, so do not draw conclusions from a handful of files. What you can measure is the quality of your own process: time to issue list, time to gap list, and defects caught before filing.

Treat the whole workflow as a supervised drafting aid. The lawyer remains responsible for what is filed, and nothing in this tutorial is legal advice for any particular case. Read the AI hub for the rest of the series on professional practices.

Questions immigration teams ask about AI and RFEs

Can I paste an RFE notice into ChatGPT or Claude?

Only if your firm's tool, plan and client engagement terms allow client data in that product. Use a business plan that excludes content from training, redact identifiers you do not need, and check your professional-conduct rules on confidentiality. If the answer is unclear, use a legal-specific tool with a data agreement.

Does AI know what the agency will accept as evidence?

No. It can repeat common patterns, but it cannot know the adjudicator's reasoning in your file. Read the notice, the statute, the regulation and current policy guidance yourself, and treat any AI statement about what is required as a lead to verify.

How do I stop invented case citations in a draft?

Do not let the assistant supply authorities. Give it the list of cases and policy sections you have already pulled and instruct it to cite only from that list. Then open every authority and check the proposition it is cited for before the draft leaves the building.

Should the client see the AI-drafted version?

Clients should see a lawyer-reviewed draft, not raw output. If you use AI in the process, say so in your engagement terms in plain words. Clients sign declarations and statements that must be accurate, so reviewing every fact with them matters more than the tool you used.

Is a small RFE worth using AI on?

Usually yes for the organising work: a three-issue notice with a 40-exhibit file takes real time to map. A one-issue notice with a clear fix may be quicker to write directly. Use AI where the file is large, not where the legal question is hard.

Further reads

• Docketwise Pricing and the AI Writing Assistant: What You Get — What a case-management AI writing assistant costs and does.
• Immigration Agent AI: Client Document Checklists and Case Updates — Client checklists and case updates.
• Law Firms AI UK: What Solicitors Should Check Before Using AI — Pre-use checks a law firm should run on any AI tool.
• What to Check in an AI Tool's Privacy Policy and Terms — What to read in an AI tool's privacy terms.
• AI Acceptable Use Policy for a Small Professional Firm — A firm policy template for AI use.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the regulations on requests for evidence (response period, no extensions, one complete response, partial response as a decision on the record, certified translations) and the adjudicating agency's policy manual chapter on RFEs (receipt by the due date, online responses); the AI Hallucination Cases database; OpenAI and Anthropic data-use pages. All read 4 November 2026.