Forms 1040, 1120, and 1065 are the bread and butter of public accounting. Today you learn how to use Claude and ChatGPT to draft, classify, and tie out workpapers — without violating IRS section 7216 disclosure rules.
By the end of this lesson you will know which tax-prep tasks AI handles well, how to redact PII before pasting into a chat window, the prompt patterns that produce review-ready Form 1040, 1120, and 1065 workpapers, and the IRC section 7216 disclosure traps that can shut down a practice.
Tax preparation is a sequence of six tasks: intake, classification, data entry, computation, review, and delivery. AI handles three of these well — classification, narrative drafting, and review-checklist generation. It handles three poorly — final computation, e-filing, and any task where a wrong answer is unrecoverable. Tax software is built to follow the Code and regulations to the letter; an LLM is built to produce plausible text. Use each for what it is.
For a Form 1040 with a Schedule C, AI is excellent at reading a year of Stripe payouts and proposing a categorization scheme: advertising, contract labor, supplies, utilities, professional fees. It is excellent at drafting the depreciation narrative once you give it the asset list and the section 179 election context. It is not excellent at computing the section 199A QBI deduction — leave that to your tax software, which is built around the regulations under section 199A.
For a Form 1120 (C-corp) or Form 1120-S (S-corp), AI shines at Schedule M-1 reconciliation narratives, drafting Schedule G ownership disclosure language, and explaining accumulated adjustments account (AAA) movement to the client. It is unreliable for state apportionment factor calculations — every state's rules differ, and a confidently wrong answer here costs real money.
For a Form 1065 partnership return, the highest-leverage AI use is K-1 footnote drafting and section 754 step-up explanations. Partners read footnotes; they ignore numbers they don't understand. A clear footnote drafted by AI and reviewed by you is worth more than three hours of recasting the same explanation by hand.
Before you paste a single client number into Claude or ChatGPT, read IRC section 7216 and the implementing regulations at Treas. Reg. section 301.7216-1 through 301.7216-3. The statute makes it a misdemeanor for a tax-return preparer to "knowingly or recklessly disclose" or "use" return information for any purpose other than preparing the return — without specific written consent from the taxpayer.
The IRS's position, articulated in Rev. Proc. 2013-14 (still the operative consent-form guidance), is that sending return information to a third-party software vendor for processing constitutes a "disclosure" requiring section 7216(a) consent unless the vendor is acting solely as the preparer's agent under Treas. Reg. section 301.7216-2(d). OpenAI, Anthropic, and Google are not your agents by default. Their consumer terms of service are not section 7216 consents.
The fix is one of three approaches. First — best — use an enterprise tier (ChatGPT Enterprise, Claude for Work, or Microsoft 365 Copilot) where the vendor signs a written agreement that they will not train on your data and will act as your agent for return-preparation purposes. Second, redact every piece of identifying information before prompting. Third, obtain written section 7216 consent from each client whose information will be processed by AI, using the format prescribed in Rev. Proc. 2013-14.
Most of the leverage in AI-assisted tax prep comes from a simple discipline: never let identifying information leave your machine. Replace the client's name with "Taxpayer A," replace the SSN with "XXX-XX-1234," replace the EIN with "XX-XXXXXXX," and replace specific addresses with the city only. Then prompt.
# Schedule C categorization prompt
You are a senior tax preparer reviewing a sole proprietor's books.
Below is a list of bank-statement transactions for tax year 2025
for Taxpayer A, a freelance graphic designer in Austin, TX.
Categorize each transaction into the appropriate Schedule C
expense line (8 through 27a). For each transaction, output:
- Line number on Schedule C
- Reasoning (one sentence)
- Confidence (high / medium / low)
Flag any transaction that may be (a) personal, (b) capitalizable
under section 263(a), or (c) subject to the section 274 meals
limitation.
Do not invent transactions. If a transaction's purpose is unclear,
output "needs clarification" and list the question to ask the client.
Transactions:
[paste redacted transaction list here]
This prompt works because it does four things at once: it sets the role, specifies the output schema, forbids hallucination ("do not invent transactions"), and gives the model a graceful fallback ("needs clarification"). Compare that to a vague prompt like "categorize these transactions for me," which produces confident garbage about half the time.
The single most dangerous AI failure mode in tax prep is the silent invented number. The model adds a depreciation figure that wasn't in the trial balance because it "looks reasonable." The fix is to force the model to quote the source.
# Trial-balance tie-out prompt
Below is the adjusted trial balance for Taxpayer B's Form 1120
(C-corp, calendar year 2025). For each line on Page 1 of the
1120 (lines 1a through 30), do the following:
1. Identify the trial-balance accounts that map to that line.
2. Quote the exact account name and balance from the TB.
3. Compute the line total.
4. Flag any TB account that does not map to a 1120 line.
If a 1120 line has no supporting TB account, output "no supporting
account found" — do not estimate, do not infer.
Trial balance:
[paste redacted TB here]
The "quote the exact account name and balance" instruction is the safety latch. When the model is forced to cite the source row, it cannot fabricate a number without it being immediately visible to you. This pattern — quote the source before answering — is the single most useful defensive prompt in any AI tax workflow.
AICPA Statement on Standards for Tax Services (SSTS) No. 1 (revised 2024, effective for engagements beginning on or after January 1, 2024) requires the preparer to have a "reasonable basis" for every position taken on a return. If you used AI to draft a position memo, the workpaper file must show what you prompted, what the model returned, and what you did to verify it. A screenshot of the chat is acceptable; a vague note like "used ChatGPT" is not.
For peer review under AICPA Quality Management Standards (QMS, effective for systems of quality management designed by December 15, 2025), a reviewer needs to be able to reconstruct your reasoning. Save the prompt, the response, your verification step (cite to Code, Reg, or PLR), and your conclusion. Treat the AI like a junior preparer whose work you are reviewing — because that is exactly what it is.
Before Day 2, do the following on a single anonymized client file (a friend's Schedule C, a sample 1040 from a CPE course, or a fictional taxpayer you construct):
Before moving to Day 2, you should be able to answer: