Busy season is mostly communication and lookup wrapped around a few hours of actual preparation. AI compresses the communication: organizer reminders, missing-document requests, explanations of an IRS notice, a memo on why the client's situation changed. It also gives you a fast starting frame for research, as long as you treat everything it says about numbers as a placeholder.
Standard deductions, phaseouts, mileage rates, and retirement limits change every year, and a chat model answers from training data that ended months or years ago. It also produces code sections, revenue procedures, and court cases that do not exist, in the same tone it uses for the real ones. The preparer penalty under Section 6694 is yours, not the model's.
Taxpayer data has its own statute. IRC Section 7216 limits how you use and disclose tax return information, and the FTC Safeguards Rule requires a written information security plan. Client data goes into an AI tool only if your firm has approved that tool and its data terms; otherwise you work from de-identified facts, which is enough for most of what follows.
Quick wins this week
- Paste your organizer cover letter and ask for a version a first-time client actually reads: shorter, a checklist, and a plain reason for each document you need.
- Upload the current IRS instructions for a form you rarely see to NotebookLM and ask your question with a page reference. It answers only from the document.
- Draft the three reminder emails you send every year (extension filed, estimates due, still missing documents) as templates with placeholders, then load them into your practice management tool.
- Give the model a notice type (CP2000, CP14, Letter 12C) and ask for a client-facing explanation of what it means and what usually happens next. Fill in the specific figures yourself from the notice.
What AI can do for tax preparers, task by task
Research framing and issue spotting
Give the fact pattern without identifiers and ask for the issues, the likely code sections and forms, the general rule, common exceptions, and what facts would change the answer. Then move to Checkpoint, CCH AnswerConnect, Bloomberg Tax, or the IRS site to confirm every citation and every number. The AI answer is the outline of the research, never the research.
Client organizers and missing-document requests
Paste last year's document list for a client profile (Schedule C with a home office, rental owner, W-2 with equity compensation) and ask for a personalized organizer with a one-line reason for each item and a friendly deadline. For mid-season chasing, list what is still missing and ask for a short email that groups the items and points to your secure upload portal.
Explaining notices and return changes to clients
Describe the notice type and the general situation (income the client forgot, a math adjustment, a penalty) and ask for a plain-English explanation of what the notice is, what the options usually are, and what you need from the client. You insert the actual amounts and deadlines from the notice. Do not paste the notice; it carries the SSN and full name.
Reading new legislation and guidance
Upload the bill text, the IRS release, or the revenue procedure to NotebookLM or Claude and ask targeted questions: effective dates, which taxpayers are affected, what changed from prior law, with page references. Then open each page and read the actual language. Summaries of a long bill are where wrong effective dates and missed exceptions come from.
Engagement letters and 7216 consents
Turn your engagement letter into a fill-in prompt (return types, fee basis, exclusions, records retention) and ask for a draft. For 7216 consents, the wording, format, and font requirements are prescribed by Treasury regulations and IRS revenue procedures, so use the AI only to plain-English the cover explanation and keep the required consent language exactly as published.
Busy-season operations
Describe your intake pipeline (return count, stages, where the bottleneck is) and ask for a triage order and a capacity plan. Use Zapier or Make to move signed engagement letters, uploaded documents, and payment confirmations into your practice management system without hand entry. Test each flow on a dummy client before pointing it at real ones.
Prompts for tax preparers
Replace the bracketed placeholders, paste into any chat assistant, and iterate on the result.
Frame a tax research question
You are a tax research assistant helping a preparer. I will describe a fact pattern with no client identifiers. Return: 1) the tax issues in order of significance, 2) the code sections, regulations, and forms most likely involved, 3) the general rule and common exceptions in plain language, 4) facts I still need to gather, 5) the specific items I must verify in primary sources. Do not state any dollar threshold, rate, or date; write [VERIFY] in its place. Do not cite cases or revenue rulings unless you are certain they exist, and mark every citation [VERIFY]. Fact pattern: [DESCRIBE THE SITUATION, ENTITY TYPE, TAX YEAR, AND STATES INVOLVED]
Tip: Every [VERIFY] is a line on your research log. If the model refuses to produce numbers, that is the point.
Write a personalized organizer letter
You are a tax preparer writing to a returning client for tax year [TAX YEAR]. Client profile: [E.G. MARRIED, TWO W-2S, ONE RENTAL PROPERTY, BROKERAGE ACCOUNT, DEPENDENT IN COLLEGE]. Last year's document list: [PASTE LIST]. Write a friendly email under 250 words with a checklist of documents grouped by category, a one-line reason for each group, a request date of [DATE], and a note about the secure upload portal ([PORTAL DESCRIPTION, NO LINK]). Do not mention specific dollar amounts, credits, or deductions they qualify for.
Tip: Make one version per client profile and save them; the rest of the season is search and replace.
Explain an IRS notice to a client
You are a tax preparer explaining an IRS notice to a client who is anxious and not a numbers person. Notice type: [NOTICE TYPE, E.G. CP2000]. General situation: [ONE SENTENCE, E.G. A 1099 THE CLIENT DID NOT SEND US]. Write an email under 200 words: what the notice is, whether it is an audit (say plainly if this notice type is not), what the options usually are, and what I need from them. Leave placeholders for the proposed amount [AMOUNT], the response deadline [DEADLINE], and my recommended next step [NEXT STEP]. Do not speculate about penalties or interest.
Tip: Fill the placeholders from the notice itself, and never paste the notice.
Question a bill or guidance document
You are a tax analyst. Answer only from the document I have uploaded, and give the section number and page for each answer. Questions: 1) Which provisions affect [TAXPAYER TYPE, E.G. SCHEDULE C FILERS]? 2) What is the effective date of each? 3) What changed compared to the prior rule, as described in the document itself? 4) Which provisions expire, and when? 5) What definitions determine who is covered? If the document does not answer a question, say NOT IN DOCUMENT rather than guessing. Do not use outside knowledge.
Tip: Works best in NotebookLM or a Claude Project with the PDF uploaded; ask one question set per document.
Second-reviewer checklist for a return profile
You are a senior reviewer at a tax firm. I will describe a return in summary form with no identifiers: [ENTITY TYPE, INCOME TYPES, MAJOR SCHEDULES, UNUSUAL ITEMS, STATES]. List: 1) the ten items most often wrong or missed on returns with this profile, 2) consistency checks between forms and schedules I should run, 3) questions a reviewer would ask the preparer, 4) documentation I should have in the file. Do not state numeric limits; mark them [VERIFY].
Tip: Run it before you hand the return to your actual reviewer; it shortens their notes.
Want a prompt for something else? Use the Prompt Builder.
Skills to build
Treating every number as a placeholder
Why: Thresholds, rates, and dates change annually and the model does not know what year it is. One stale figure in a client email is a credibility problem; on a return it is a penalty problem.
How: Prompt for [VERIFY] markers instead of figures. Keep this year's key numbers in a one-page reference you consult, not the model.
Separating facts from identifiers
Why: Section 7216 and the Safeguards Rule are about tax return information, and most research questions need none of it.
How: Describe the situation the way you would to a colleague in the hallway: entity type, income types, the question. No names, no SSNs, no notices pasted.
Verifying citations in primary sources
Why: Models produce plausible code sections, revenue procedures, and cases that do not exist, and they misstate the ones that do.
How: Every citation gets opened in your research platform or on irs.gov before it goes in a memo or a client letter. Log what you checked; Circular 230 diligence is easier to show than to describe.
Working from the document, not the summary
Why: Bills and IRS guidance are long, and the model's summary flattens exceptions and effective dates.
How: Upload the source to NotebookLM or Claude, ask for section and page references, and read each cited passage yourself.
Templating the recurring conversations
Why: You send the same twenty emails to hundreds of clients. Consistent, fast, and reviewed once beats rewritten every time.
How: Save your organizer, reminder, extension, and notice templates as prompts in a Claude Project or custom GPT with your tone rules and the no-figures constraint built in.
Tools worth knowing
Claude
A careful writing and analysis assistant that shines on long documents.
ChatGPT
The general-purpose AI assistant most of your coworkers already use.
NotebookLM
A research notebook that only answers from the sources you give it, with citations.
Microsoft Copilot
AI inside Word, Excel, Outlook, and Teams, with your company's data protections.
Perplexity
An answer engine that cites its sources, built for research rather than chat.
Zapier
Connect thousands of apps with no-code workflows, now with AI steps, agents, and chatbots
Cautions for tax preparers
Section 7216 makes knowing or reckless use or disclosure of tax return information a crime, and Section 6713 adds civil penalties. Sending client data to a third-party AI service can be a disclosure. Before any client data touches a tool, your firm should confirm that the vendor agreement and the 7216 regulations' contractor provisions cover it, or obtain client consent in the prescribed form. When in doubt, use de-identified facts only, and never paste tax return information into a consumer AI tool unless your organization has approved it.
Tax preparers are financial institutions under the Safeguards Rule, and the PTIN application asks you to acknowledge your data security obligations. Add AI tools to your written information security plan: which are approved, in which tiers, what data is allowed, who reviews vendor terms. IRS Publication 4557 and Publication 5708 are the starting points.
The model's training ended before this filing season. It will state last year's standard deduction, an expired credit, or a phaseout that has since changed, and cite a revenue ruling that does not exist. Every figure and citation comes from a primary source you opened this year. Section 6694 preparer penalties do not have an AI exception.
Competence and diligence obligations apply to how you use tools. Drafting with AI is fine; relying on it without verification is not. For the earned income credit, child tax credit, American opportunity credit, and head of household status, Form 8867 due diligence requires the questions and documentation to come from you, not from a model that has never met the client.
Never forward AI output to a client as an answer. Positions, elections, and planning belong to you as the preparer. The model can help you explain a decision you already made; it cannot make one.
Your 30-day plan
- Week 1: Pull out your WISP and add an AI section: approved tools and tiers, banned data (names, SSNs, EINs, notices, returns), who approves new tools. If you have no WISP, start from IRS Publication 5708.
- Week 1: Rewrite your organizer letter and three reminder emails with a chat tool. Save them as templates.
- Week 2: Take five research questions you answered last season, run them through the research-framing prompt, and score the output: right issues, wrong issues, and what it invented.
- Week 2: Upload one current IRS publication to NotebookLM and use it for a week of lookups. Note where it beats your usual search and where it was incomplete.
- Week 3: Build a Claude Project or custom GPT with your templates, tone, and the numbers-are-placeholders rule. Use it for every client email for one week.
- Week 4: Map one intake workflow (engagement letter signed to document reminder sent) in Zapier or your practice management tool and test it on a dummy client. Then brief your team, including the errors the tools made.
Frequently asked questions
Can I use ChatGPT for tax research?
Is it a 7216 violation to put client information into an AI tool?
Will AI replace tax preparers?
Does the IRS allow AI-prepared returns?
Which is better for tax work, ChatGPT or Claude?
Terms used on this page
Related roles
- AI for AccountantsAI drafts the flux commentary, writes the XLOOKUP, explains the K-1 to your client in plain English, and reads the 40-page lease for you. You still own the numbers, the judgment, and the signature.
- AI for BookkeepersAI writes the collections email, drafts the categorization rules, builds the month-end checklist, and explains the P&L to an owner who has never read one. You still decide what the transaction actually was and whether the books are ready to close.
- AI for Financial AdvisorsAI drafts the meeting brief from your notes, turns a 45-minute review into a follow-up letter, and rewrites your explanation of sequence-of-returns risk for a client who hates charts. The recommendation, the best-interest judgment, and the compliance file stay with you.
- AI for Small Business OwnersAI answers the one-star review calmly, drafts the month of posts, turns your voice memo into an SOP, and preps the questions for your CPA. You still make the calls, sign the checks, and own what goes out under your name.