Form 8867 Due Diligence With AI: Interview Notes and Checklists

A tax preparer can use AI to turn a rambling client interview into structured notes, run the Form 8867 checklist questions against those notes and flag missing answers, as long as the preparer asks the questions, judges the answers and signs. AI organises the file. It cannot hold the knowledge or the diligence the form certifies.

The practice owner who reviews other people's files has a different problem from the preparer. You are less interested in how one preparer takes notes than in whether twelve preparers produce files that look the same, show the questions asked and survive a second look. That is where a consistent note structure, and AI to apply it, helps most.

Follow me on Instagram@sagnikteaches

What a due diligence file has to show

Form 8867 is the paid preparer's due diligence checklist. It applies to returns or refund claims that include certain credits and a filing status: the earned income credit, the child tax credit (including the additional child tax credit and credit for other dependants), the AOTC education credit and head of household filing status. The current instructions (revised November 2025) set out four requirements.

Connect on LinkedInSagnik Bhattacharya

1. Knowledge: interview the client, ask adequate questions, document the questions and the answers at the time (on the return or in your notes), and review enough information to decide eligibility and work out the amounts. Information that looks incorrect, inconsistent or incomplete cannot be ignored.
2. Complete Form 8867 truthfully and accurately, and complete the actions it describes for each credit or status claimed, including the worksheets.
3. Submit Form 8867 in the manner required, with the return.
4. Keep five records for three years: a copy of Form 8867; the worksheets (the official ones or your own); copies of documents the client provided that you relied on; a record of how, when and from whom the information was obtained; and a record of any additional information you relied on, including the questions you asked and the client's answers.

The three years run from the latest of the return's due date (not counting extensions) and, depending on your role, the date a signing preparer e-filed it, the date it was presented to the client for signature if it was not e-filed, or, for a non-signing preparer, the date their part went to the signing preparer. Paper or electronic storage is allowed.

Subscribe on YouTube@sagnikteaches

The penalty is charged per failure, for each credit or status. The instructions give $650 for each failure on a return filed in 2026, so a return claiming all four items that fails on every one could cost $2,600. The amount is adjusted for inflation each year, so check the current figure before you put it in a training note. Multiply that by a season's files and the case for a consistent note structure makes itself.

What AI is for in this process, and what it is not

The preparer's duty is personal. A model cannot ask a follow-up question, notice that a client hesitated or decide that two answers are inconsistent in a way that matters. What it can do is take the messy output of a conversation and put it into the same shape every time, then compare that shape against a checklist. That is clerical, and it is a good use.

If the AI's output ever replaces the preparer's question, the diligence has been outsourced and the form's certification is no longer true. The timing matters too. Because the instructions ask for questions and answers to be documented at the time, the raw notes typed during the call are the contemporaneous record. The AI's tidy version is a working copy made afterwards, and the file should keep both.

Structure the interview first, then the AI

Notes are only as good as the interview. Give preparers a fixed set of question blocks, one per credit or status, and train them to ask each block every time, not only when the client volunteers something. For a typical individual client the blocks are:

• Household: who lived at the home, for how many months, and where each person lived the rest of the year.
• Relationship and support: how each dependant is related to the client, who paid for what, and whether anyone else could claim them.
• Income: sources, amounts, and how it is documented, including any self-employment records.
• Filing status: marital status at year end, who paid the cost of keeping up the home, and any separation details.
• Education credit: student status, enrolment, qualified expenses and any prior-year claims.
• Documents seen: which documents the preparer actually looked at, and which the client only described.

The point of the last block is honesty. "Client states" and "document reviewed" are not the same level of evidence. A reviewer reading the file should see at once which is which. The client tax organiser template can collect part of this before the appointment, which shortens the interview and leaves time for the questions that need a person.

A note template the AI fills and the preparer signs off

Decide the layout once for the whole practice. Here is a compact one:

CLIENT: [initials / file ID]  YEAR: [tax year]  PREPARER: [name]  INTERVIEW DATE: [date]  METHOD: [in person / phone / video]
INFORMATION FROM: [who gave each block of information, how and when]
CREDITS AND STATUS CLAIMED: [EITC / CTC-ACTC-ODC / AOTC / HOH]
HOUSEHOLD: [name or initials, relationship, age, months in home, evidence seen]
INCOME: [source, amount, evidence seen]
SUPPORT AND CLAIMS BY OTHERS: [facts, who else could claim]
FILING STATUS FACTS: [marital status, who maintained the home, evidence]
QUESTIONS ASKED AND ANSWERS: [question, client's answer, follow-up asked, outcome]
INCONSISTENCIES NOTED AND HOW RESOLVED: [list]
DOCUMENTS REVIEWED: [list with dates]
STILL TO CONFIRM: [list]

Then the prompt that applies it:

You organise a tax preparer's raw interview notes into the template below. Use only what the notes say.
Do not add facts, assume relationships or fill blanks. If something is missing, write "NOT RECORDED".
Mark each statement as "client states" or "document reviewed" depending on the notes.
After the template, list: (1) fields marked NOT RECORDED, (2) any statements in the notes that conflict with each other, (3) any answer that seems incomplete for the credit being claimed.
Do not give tax advice or say whether the client qualifies.
Raw notes: [paste]
Template: [paste]

Run this in a business plan where your content is not used for training by default, and strip identifiers first. Use initials and file numbers, no full names, addresses or tax identification numbers. Your firm's written information security plan should say that, and it is worth writing down if it does not.

Worked example: a head of household and child credit file

Say a six-preparer practice sees a client who wants to file as head of household and claim a child tax credit for two children and the earned income credit. The preparer's raw notes, typed during the call, read in part: "Kids with her all year, ages 7 and 10. Works part time retail, plus some cash cleaning. Ex pays nothing. Rent in her name. Showed lease and school letters, no bank statements." The AI returns the template filled in, with the household block showing both children, months in home "all year (client states)", school letters reviewed and no evidence of the second claim. Its gap list says: cash cleaning income amount not recorded; marital status at year end not recorded; no record of whether the ex-partner could claim the children; no document for the cash income.

That list is the value. The preparer returns to the client with four targeted questions, records the answers and notes "cash income: client provided a handwritten log, three entries per week, reviewed". The preparer then answers the Form 8867 questions and signs. The owner reviewing the file sees the same sections in the same order as every other file, and the "still to confirm" box is empty because it was resolved, not ignored.

Using AI to find inconsistencies, with a cautionary example

Consistency checks are the best second job. Ask the AI to read the structured notes and list statements that do not fit together: income too low for the stated housing, a child claimed as living at home all year who is also listed as attending a boarding school, or a dependant whose age does not match the stated relationship. These are prompts for the preparer to ask a question, not conclusions.

A realistic mistake: an AI summary of a client's notes says "both children lived with client all year" because the notes said "kids with her all year". But earlier in the same notes, the preparer had written "older child at dad's most weekends and summers". The summary dropped the second line. The fix is the instruction "list statements that conflict" and the owner's rule that the preparer reads the raw notes beside the AI output before relying on either. Summaries lose caveats; the preparer should check that the nuance survived.

Generate the follow-up questions, then ask them yourself

A useful second prompt takes the structured notes and proposes questions the preparer has not yet asked. It does not decide anything. It is a prompt for the conversation.

Read the structured interview notes below. For each credit or status claimed, list up to five follow-up questions a careful preparer would still want answered, based only on gaps or tensions in the notes.
Phrase each as a neutral question to the client, in plain language. Do not suggest answers. Do not say whether the client qualifies.
Group by credit or status. Notes: [paste]

Illustrative output for the head of household file: "Who paid the main costs of the home during the year, and can you show how? Were you and the children's other parent ever married, and if you were still married at year end, did your spouse live somewhere else for the whole of the last six months of the year? Has anyone else said they will claim either child? What records show the cash cleaning income?" The preparer picks the questions that matter, asks them, records the answers and strikes the rest. The record shows the questions asked, which is exactly the evidence of diligence the form expects.

Second example: a student dependant and an education credit

A different client, a parent, wants to claim the AOTC for a child at college. The raw notes say: "Daughter second year, full time, tuition about 9k, paid with loan and some from us. Has 1098-T. Didnt get scholarship I think." The structured version flags: scholarship status uncertain ("I think"); expenses paid by loan versus family not split; form received but amounts not recorded; years of prior claim not recorded; enrolment half-time or more not confirmed in documents. Each item becomes a question. The preparer asks for the college's account statement, which settles the scholarship point, and records the years already claimed. Without the structured list, the vague "I think" would have gone into the return as "no scholarship".

Third example: self-employed income with thin records

A sole trader with an earned income credit claim brings a shoebox of receipts and says income was "about 18,000, mostly cash". The notes show no invoices and no bank deposit pattern. The AI's gap list notes: no document supports gross receipts; expenses recorded as a lump; no mileage log mentioned. The preparer asks for any records the client has, such as app payments, text confirmations or a handwritten ledger, and records what was and was not available. The file does not need perfect records, but it must show what was asked and what the client provided, and the preparer's reasoning for accepting or questioning the figures. The AI helps by making that gap impossible to overlook.

Where AI-structured notes go wrong in tax files

• Tidy but wrong relationship labels. A model may turn "partner's son" into "child" or "stepchild". Keep relationship words exactly as stated, then have the preparer confirm.
• Rounded numbers. "About 18k" becomes "$18,000". Instruct the tool to quote amounts as given, with the client's own qualifier.
• Silent merging of two conversations. If a client calls back, new notes can overwrite old ones. Date every note block and keep both.
• False reassurance. A neat summary looks complete. Make the NOT RECORDED list part of the standard output, so incompleteness is visible.
• Identifiers pasted by habit. Test with a dummy file to confirm nobody pastes full tax identification numbers into the tool.

A one-page practice standard to give every preparer

1. Ask every interview block for every claimed credit or status, every time.
2. Write raw notes during the call; do not rely on memory.
3. Remove identifiers before using any AI tool; use file IDs.
4. Run the structuring prompt; read the raw notes next to the output.
5. Resolve every NOT RECORDED item by asking the client, and record the answer.
6. Mark each fact as client stated or document reviewed.
7. Answer Form 8867 yourself, using the notes and worksheets, and sign.
8. File notes, documents list, worksheets and the form together under the tax year.
9. Delete AI chat history once the saved file is complete.

Put that on one sheet, laminated if you like. A standard that fits on a page gets followed in February, when nobody reads a manual.

Busy-season capacity: a quick sum

Illustrative figures for a six-preparer practice with 400 files that carry Form 8867 in a season. If structuring the notes by hand takes about twelve minutes a file and the AI-assisted route takes about five including the preparer's read-through, the saving is roughly seven minutes a file, about 47 hours across the season. Spend part of it on the thing that matters, the callbacks: even two extra minutes per file asking follow-up questions before filing uses less than a third of the time saved. Measure your own timings in the first week before promising anyone a figure.

Owner's review: what you sample and what you look for

As the supervisor, you cannot read every file in full in busy season. Sample, and use the same structure to speed it up. For each file in the sample:

1. Are all template sections filled, and are NOT RECORDED boxes resolved or explained?
2. Does the "documents reviewed" list match what is in the file folder?
3. Are there questions asked and answers recorded for any statement that looked unusual?
4. Do the Form 8867 answers match the notes, line by line?
5. Are the credit worksheets and the form saved with the return and indexed?
6. Is there evidence the preparer, not the tool, made the decisions (initials, date, a short reasoning line)?

Ask AI to prepare a one-page review sheet for each sampled file, with the file's gaps and any conflicts on top, and tick the boxes yourself. A reviewer who has to find the problems in raw notes will skip files; a reviewer given a ranked list reads more.

Retention and where the records live

Whatever the AI produces becomes part of your records only if you save it. Decide what is the record: the preparer's raw notes, the structured version or both. A defensible rule is to keep the raw notes as the original and the structured version as the working file, and to record which tool and date produced the structure. Keep the completed Form 8867, worksheets, notes, documents reviewed and the record of follow-up questions together, indexed by tax year and file ID, and apply the retention period from the current instructions plus any longer period your firm's policy sets.

Set a calendar rule for deletion. Keeping everything for ever means carrying the risk of a data breach for ever. Delete AI chat history for client work regularly and keep the saved file, not the chat, as the record. If you are building the wider process, the tutorial on small tax practices using AI through busy season shows where this fits among the other uses.

Training preparers to use the routine without leaning on it

Run one calibration session each season. Pick three anonymised files: one clean, one with a hidden inconsistency and one thin on evidence. Have each preparer write notes from a scripted client call, run them through the template and compare. The aim is to show that the AI structures what the preparer asked, and gives back nothing that was not asked. A preparer who sees the NOT RECORDED list for their own thin interview learns to ask more.

Also teach the rule against leaning on the tool: if the AI says there are no gaps, the preparer still reads the notes. If you want the practical side of getting clients to provide cleaner information up front, the tutorial on turning handwritten forms into spreadsheet data covers one way.

How to check the routine is working

After the first month of busy season, pull ten files and test them against the six review questions. Count how many had unresolved NOT RECORDED boxes, how many needed a callback to the client and how long review took per file. If review time fell and callbacks happened before filing, not after, the structure is doing its job. If callbacks are still arriving after returns are filed, the interview blocks need tightening.

Review the process again at year end, using any notices or questions the practice received. Treat them as free feedback on which blocks to strengthen. If you would like help setting up the template, the prompt and the review routine across your preparers, that is what my AI implementation consultation covers, and the six-preparer example above is the sort of workflow we would map together.

Questions practice owners ask about AI and due diligence files

Can AI complete Form 8867 for the preparer?

It can organise the notes that the form's questions rely on, but the preparer must answer the form, because the due diligence duty belongs to the person signing the return. Use AI to structure and cross-check, then have the preparer confirm each answer from the interview and the documents.

Should I put client documents into an AI tool?

Only after deciding, in writing, that the tool's data terms are acceptable for taxpayer information and that your firm's security plan covers it. Prefer notes with identifiers removed, such as initials and the last four digits only, and never paste full tax identification numbers.

How long should due diligence records be kept?

Three years under the current Form 8867 instructions, counted from the latest of the return's due date (ignoring extensions) and, depending on your role, the date it was e-filed, presented to the client for signature or passed to the signing preparer. Records can be paper or electronic. Check each season's instructions and your own retention policy before you set a schedule.

What does 'knowledge' mean for a preparer?

The preparer must not ignore information that seems incorrect, inconsistent or incomplete, and must ask reasonable follow-up questions and record the answers. An AI summary cannot hold that knowledge for the preparer. It can point to inconsistencies, but the preparer still has to ask.

Further reads

• What an AI Implementation Looks Like in a Small Accounting Firm — What a sensible AI rollout looks like in a small firm.
• Can AI Fill In Forms and Supplier Portals for You? — What AI can and cannot do on forms and portals.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check readiness before adding AI to regulated work.
• How Much Does AI Cost a Small Accounting Firm? — Budget for AI across a small accounting firm.
• Compliance Calendar: Track Licences, Insurance and Certificates — Track retention and review dates on one calendar.
• ChatGPT Prompts for Tax Preparers: Client Emails and Checklists — A redaction-first prompt pack for tax preparers: reminders, extension notices, document checklists and engagement follow-ups, using placeholders only.
• Circular 230 and AI: Due Diligence When a Tool Drafts the Work — A per-return checklist and sign-off sheet that turns the IRS's June 2026 AI alert and Circular 230 diligence rules into a routine.
• ChatGPT for Tax Preparers: How to Redact a Return Before You Paste — A redaction checklist for tax preparers: what counts as taxpayer information, how to strip a PDF properly and which placeholders keep the answer useful.
• WISP for Tax Preparers: Add Your AI Tools Without Gaps — One added section for your existing WISP that covers AI tools: register, vendor due diligence, data rules, MFA and incident steps, with sample wording.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the Instructions for Form 8867 (revised November 2025) and the About Form 8867 page on the tax agency's website, checked November 2026; always read the current-year instructions for exact wording and amounts.