ChatGPT for Accountants: What It Can and Can't Do (2026)

Every guide to ChatGPT for accountants is a list of things you can do with it. None of them tells you which tasks to hand over and which to keep. ChatGPT is a good writer and a poor accountant: in a firm it earns its subscription on language work, and it fails on anything that needs your ledger, your workpapers or your client's actual files. What decides each task is not a feature list. It is how much client data the task forces you to paste in.

Short answer: ChatGPT is the right tool for four jobs in an accounting firm: drafting client emails, writing Excel formulas, turning a variance into a sentence, and firm marketing. It is the wrong tool for converting bank statements, coding the bank feed, audit fieldwork, meeting notes, and chasing clients for documents. Each of those has something built for it.

The line that matters is your client's data, not the feature. AICPA ET section 1.700.001 treats client information as confidential, and for tax return information 26 U.S.C. 7216 makes reckless disclosure a misdemeanor. Strip the names and the numbers before you paste. If the task only works with the real numbers in it, use a tool under contract with your firm.

Can accountants use ChatGPT?

Yes. No professional rule bans an accountant from using a general AI assistant; the rules govern the client information you put into it. Karbon's State of AI in Accounting 2026, a vendor survey of nearly 600 professionals released January 20, 2026, reports 98% of firms using AI, 60 minutes saved per employee per day, and 21% with a documented AI policy.¹ Self-reported by a software company, so read it as direction. The tool is already open on the desk; the rule for using it is written nowhere.

Of the 12 recurring tasks in an accounting firm below, ChatGPT is the right tool for 4, the wrong tool for 5, and usable on 3 more for wording only or with the numbers replaced. Here is the work in the order it happens. The last column is the one nobody else publishes.

Prices in this article were checked September 20, 2026, on the vendors' own pages.²

What ChatGPT is genuinely good at in a firm

ChatGPT is good at the tasks where the deliverable is language and the client data can be taken out first. That is a narrower list than any prompt roundup suggests, and it is where the $20 a month pays for itself.²

Client emails and difficult explanations. The best single use. You know what you decided about the S-corp election; the hard part is saying it to someone who does not read tax law. Karbon puts the same task first in its own six.³ Describe the client generically, then put the names and figures back yourself. Drafting the same request for the fifth time? A document request email template beats regenerating it.

Excel formulas. Describe the columns and the result you want, get the formula back as text. It cannot see your workbook, which is why it is safe here: the prompt contains a shape, not data. To have AI act on the open file, that is Microsoft 365 Copilot, a business add-on at $18 per user per month paid yearly, a promotion running July 1 to December 31, 2026 and good for the first year only, against a $21 list price. Billed monthly it is $25.20.²

The first sentence of a variance. "Cost of sales up 14% on volume up 9%" becomes a readable line for the pack. Keep it in percentages and nothing about the client leaves your firm. Numeric writes the same sentence from your actual numbers, which is a different product and a different risk; it starts at $30 per user per month, with flux analysis on a higher tier.²

The tone of a client report, not its figures. ChatGPT will not have the numbers, and pasting them is the problem this article is about. Fathom's Commentary Writer, sold as "AI report commentary with full number visibility", starts at A$59 a month for one company, excluding GST. Australian dollars.²

Firm marketing. Newsletters, a tax-season explainer, website copy, proposals. There is no client data in a blog post about the Section 179 limit. This is the part of the budget with no downside.

Explaining a rule, for wording only. It writes a clean paragraph on a standard and invents code sections with the same confidence. Surgent's write-up names the failure mode plainly: ChatGPT produces "hallucinations, confident assertions containing factual inaccuracies".⁴ Research tools built for tax answer with the authority attached. Blue J publishes "inline citations and source lists" at $1,498 per user per year; TaxGPT sells citation-backed answers with no published price.²

What ChatGPT should not be doing in your firm

The wrong-tool tasks share one trait: they need the client's real records, and a chat box is not where those belong.

Turning PDF bank statements into rows. Karbon lists this as a use of ChatGPT, and at volume it is the worst confidentiality exposure on this page: a full statement is a client's entire month.³ DocuClipper does the job properly, from $20 a month, or $14 a month billed annually, for 60 pages, exporting to XLSX, CSV, QBO, OFX and QIF among others, straight into QuickBooks Online, Xero, Sage or NetSuite.²

Coding and reconciling the bank feed. ChatGPT does not post to a ledger. QuickBooks Online Accounting AI does, and its own help page draws the line where I would: Ready to post "groups the transactions that your Accounting AI is very confident you'd be able to post", and you post them. It is included with QuickBooks Online Essentials, Plus and Advanced, not with Simple Start.² Reconciling differences is an O*NET task statement for accountants for a reason: the judgment is the job.⁵

Meeting notes. ChatGPT was not in the meeting. The only way to get notes out of it is to paste the transcript, which is the whole client conversation, unredacted. Otter.ai records it instead, with automated summaries with action items and outlines, free for 300 minutes a month and $16.99 per user monthly on Pro.²

Chasing clients for documents. ChatGPT can word the request. It cannot send it under your name, hold the file, know what is still missing, or follow up on Tuesday. In Financial Cents' 2024 State of Accounting Workflow Automation Report, a survey of 367 firm owners run between November 9, 2023 and January 18, 2024, "getting info and documents from clients carried the crown according to 65.2% of respondents" as the biggest workflow challenge.⁶ The number one problem in the firm, and the chatbot does not touch it.

Client data: what you can and cannot paste into ChatGPT

Three separate rules govern the paste, and only one of them is about AI. Read them in this order.

1. Your plan decides the contract, not the safety of the paste. ChatGPT is Free at $0, Go at $8 a month, Plus at $20, Pro from $100, and Business at $25 per user per month billed monthly or $20 billed annually for two or more users.² Two training statements I could verify myself on September 20, 2026. OpenAI's developer documentation says "We do not train our models on inputs and outputs through our API". Anthropic publishes "None by default" for model training on Claude Team and Enterprise content, next to Team seats at $25 monthly or $20 annually.⁷ OpenAI's business privacy page, which carries the equivalent statement for ChatGPT Business and Enterprise, refused automated retrieval that day: open it yourself and read the current sentence before you write it into a firm policy.⁷ The received wisdom on the ranking pages is out of date, too. Karbon's guide still tells accountants that ChatGPT "learns from, and is trained on, all information inputs", which is not what the vendors publish today.³

2. AICPA: pasting client data into a tool can be a disclosure. The Confidential Client Information Rule, ET section 1.700.001, provides that a member in public practice shall not disclose confidential client information without the client's specific consent. Writing in the Journal of Accountancy on August 5, 2026, Kelly D. Mullins puts the test in one line: "Before pasting client data into any AI system, stop and ask: Is this confidential client information?" And where consent is needed, "make sure your client understands what information may be used, why it's being used, and who may have access to it."⁸

3. Tax return information is a criminal statute, not only an ethics rule. Under 26 U.S.C. 7216(a), a person in the business of preparing returns who knowingly or recklessly discloses information furnished for the preparation of a return, or uses it for any purpose other than preparing that return, "shall be guilty of a misdemeanor, and, upon conviction thereof, shall be fined not more than $1,000 ... or imprisoned not more than 1 year, or both, together with the costs of prosecution."⁹ IRS Publication 4557 adds the civil side. "According to the FTC Safeguards Rule, tax return preparers must create and enact security plans to protect client data." It points to Publication 5708 for the written plan, warns that failure "may result in an FTC investigation", and lists among its basic steps: "Limit access to taxpayer data to individuals who need to know."¹⁰ If your firm has a WISP, the written information security plan the FTC Safeguards Rule requires of tax return preparers, that plan has to say which AI tools are allowed. Karbon's 21% says most firms have not written that sentence yet.¹

The working rule: strip it before you paste it. Names, EINs, SSNs, account numbers, addresses and the actual dollar figures come out; the structure stays. Ask about a client with these facts, never about your client. If the task only works with the real numbers in it, a statement conversion or a flux on live data, that is the signal to use a tool under contract with your firm instead. Karbon says it in fewer words: "Make sure anything you use is de-identified and anonymized."³

5 ChatGPT prompts for accountants, with the client data taken out

These five prompts cover the firm tasks where ChatGPT is the right tool: a client email, a plain-English explanation of a rule, an Excel formula, a variance sentence, and a tax-season newsletter item. Each is written to run on no client data at all, and each carries what to redact first and what you still have to check yourself.

1. The client email

Write a short email to a small-business client explaining that I am moving
their bookkeeping to accrual for the year end, why it changes their
reported profit, and what I need from them. Plain English, six sentences,
no jargon, friendly but not chatty.

Redact first: client and entity name, the industry if it identifies them, every figure. Still on you: put the real names and numbers back, and confirm the reason you gave is the one that applies.

2. The plain-English explanation

Explain the difference between a Section 179 deduction and bonus
depreciation to a client who has never filed a business return. Eight
sentences, no code sections, one concrete example with round numbers.

Redact first: nothing, if you keep it generic. Still on you: check every rule, threshold and limit against primary authority before it leaves the office. This is the prompt that hallucinates.

3. The Excel formula

I have a sheet with dates in column A, vendor names in column B, and
amounts in column C. Write a formula that totals column C for one vendor
between two dates held in cells F1 and F2. Explain what each argument does.

Redact first: vendor names, real amounts. Describe the shape, not the data. Still on you: test it on rows you can check by hand before trusting a total.

4. The variance sentence

Write one sentence for a management pack explaining that cost of sales
rose 14% while volume rose 9%, and one follow-up question I should ask
the operations manager.

Redact first: dollar figures. Percentages and direction are enough. Still on you: verify the explanation is the true one. The model will write a plausible cause for numbers it has never seen.

5. The tax-season explainer

Draft a 250-word newsletter item for individual tax clients about what to
gather before their appointment, in the second person, no fear tactics,
with a short bulleted list of documents.

Redact first: nothing, there is no client data in a newsletter. Still on you: check the document list against what you actually require, or start from a tax document checklist and have it write around that.

ChatGPT for CPAs: what changes in audit and attest work

In attest work the problem is not accuracy, it is reviewability. A workpaper has to show how you got there, and a chat session is not something a reviewer can re-perform. DataSnipper, which sells the alternative, states the gap accurately: ChatGPT "doesn't interpret your audit methodology, client risk, or internal controls", and its outputs typically require manual verification and lack source citation.¹¹ Its own automation, Test of Details, payroll testing and AR cut-off testing, runs inside Excel where the trail stays in the file. It publishes no price.²

Adjusting entries sit just outside that line. Use ChatGPT for the narrative of a workpaper note with the figures replaced by placeholders, never for the arithmetic, and never as the record of why the entry was made. Whatever drafts the sentence, the entry and the conclusion are signed by a person.

Where a purpose-built document tool beats ChatGPT

Chasing clients for documents is a tracking problem, not a writing problem, which is why the task firms name as their biggest headache is the one a chatbot structurally cannot do.⁶ Send 60 tax clients two follow-ups each at five minutes apiece and that is 10 hours a season on chasing alone, before a single wrong or missing file, and every wrong-year W-2 adds a round trip. That is arithmetic on a 60-client season, not a measured result.

This is the one section where I make the case for my own product. DokuTrak is document intake software for accounting and tax firms: it sends one itemized request per client, tracks what is still missing, and follows up on a schedule instead of from your outbox. Clients upload from a single link with no account to create, which matters when the person you need a 1099 from has not signed into anything since last April. An AI first pass flags the wrong document type, the wrong year and files nobody can read, and you accept each one or send it back. It does not make the call for you and it does not replace your tax software. Solo is $79 a month, Team $199, Agency $449, with a 14-day trial; the plans are on DokuTrak pricing. Content Snare, from $35 a month billed annually with unlimited reminders, and SmartVault do versions of the same job.²

How to choose: the plan, then the task

Answer the plan question before the task question: it is the only one of the four with a legal consequence. Then work down the rest.

1. Which plan am I on, and what does its provider publish about my data?
2. Does this task require real client data? If yes, ChatGPT is out unless you can strip it and still get the answer.
3. Does something I already pay for do it? QuickBooks, your practice management software and Microsoft 365 all ship AI you have already bought.
4. Does the tool make the call, or does it draft and flag? If it makes the call, you answer for it.

Run your own week past those four questions and most of the AI shortlist disappears. What is left, for most firms, is a good writing assistant and one fix for the document chase. For a new client, the onboarding checklist is the list to work from; the client-facing tax document checklist is the one they can follow.

Send your first request free. First client request free, no card. Then a 14-day trial, no upfront charge.

Sources

Footnotes

1. Karbon, State of AI in Accounting 2026, released January 20, 2026, a survey of nearly 600 accounting professionals: 98% of firms use AI, an average of 60 minutes saved per employee per day (approximately 21 hours a month), and 21% with a documented AI policy or strategy. Vendor survey, self-reported. https://karbonhq.com/resources/state-of-ai-accounting-2026/ (retrieved September 20, 2026). ↩ ↩²

2. Vendor pricing and product pages, all retrieved September 20, 2026. ChatGPT plans (Free $0, Go $8/mo, Plus $20/mo, Pro from $100/mo, Business $25/user/mo monthly or $20/user/mo annually for 2+ users, Enterprise contact sales): https://learn.chatgpt.com/docs/pricing · Claude (Pro $20/mo monthly, $17/mo annual; Team $25/seat monthly, $20/seat annual; Enterprise $20/seat/mo annual plus usage): https://claude.com/pricing · Microsoft 365 Copilot ($21.00 user/month paid yearly list, $18.00 promotional July 1-December 31, 2026, first year only, $25.20 billed monthly): https://www.microsoft.com/en-us/microsoft-365-copilot/pricing · Karbon (Team $59/user/mo annual, $79 monthly; Business $89 annual, $99 monthly; Kai AI Coworker): https://karbonhq.com/pricing/ · Blue J (Individual $1,498/year per user, "inline citations and source lists"): https://www.bluej.com/pricing · TaxGPT (no published price, demo request): https://www.taxgpt.com/pricing · DocuClipper (Starter $20/mo monthly, $14/mo annual, 60 pages; Business $111/$78, 640 pages; Enterprise $360/$288, 2,000 pages; 14-day trial): https://www.docuclipper.com/pricing/ · QuickBooks Online Accounting AI ("Ready to post groups the transactions that your Accounting AI is very confident you'd be able to post"; included with Essentials, Plus, Advanced and Intuit Enterprise Suite): https://quickbooks.intuit.com/learn-support/en-us/help-article/bank-transactions/accounting-agent-features/L6pl9rv94_US_en_US · Numeric (Essentials from $30/user/mo, with the Technical AI Assistant and close orchestration; flux analysis on the Growth tier; Growth and Enterprise custom): https://www.numeric.io/pricing · Otter.ai (Basic free, 300 min/mo; Pro $16.99/user/mo monthly, $8.33 annual, 1,200 min; Business $30/$19.99, unlimited; "automated summaries with action items and outlines"): https://www.otter.ai/pricing · Content Snare (Basic $35/mo annual, $42 monthly, 20 active requests; "Unlimited clients / recipients" and "Unlimited reminders" on every plan): https://contentsnare.com/pricing/ · Fathom (Commentary Writer, "AI report commentary with full number visibility"; Starter A$59/month for one company, excluding GST): https://www.fathomhq.com/pricing · DataSnipper (no published price): https://www.datasnipper.com/resources/ai-in-accounting-chatgpt-limitations · DokuTrak plans (Solo $79/mo, Team $199, Agency $449, 14-day trial): https://dokutrak.com/pricing ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

3. Karbon, 6 ways you can use ChatGPT as an accountant now (retrieved September 20, 2026): the six uses are simplify communication, generate marketing strategies and copy, sort data, extract transactions from bank statements, research uncategorized transactions, and learn new Excel skills. "Make sure anything you use is de-identified and anonymized." "AI doesn't know what's proprietary information and what's general knowledge." The page also states that ChatGPT "learns from, and is trained on, all information inputs", which is the older received wisdom and not what OpenAI and Anthropic publish today (see note 7). https://karbonhq.com/resources/use-chatgpt-as-an-accountant/ ↩ ↩² ↩³ ↩⁴

4. Surgent CPE blog, ChatGPT in the Accounting Profession: Applications and Implementation (retrieved September 20, 2026): ChatGPT produces "hallucinations, confident assertions containing factual inaccuracies", and professional publications emphasize mandatory human verification of all outputs. https://blog.surgentcpe.com/chatgpt-in-the-accounting-profession-applications-and-implementation ↩

5. O*NET OnLine, occupation 13-2011.00, Accountants and Auditors, task statements (retrieved September 20, 2026), including "Review accounts for discrepancies and reconcile differences". https://www.onetonline.org/link/summary/13-2011.00 ↩

6. Financial Cents, 2024 State of Accounting Workflow Automation Report, 367 completed responses from accounting, bookkeeping and tax firm owners collected between November 9, 2023 and January 18, 2024: "Getting info and documents from clients carried the crown according to 65.2% of respondents." https://financial-cents.com/resources/articles/2024-report-state-of-accounting-workflow-automation/ (retrieved September 20, 2026). ↩ ↩²

7. OpenAI developer documentation, Key concepts and Your data (retrieved September 20, 2026): "We do not train our models on inputs and outputs through our API", and data sent to the API "is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)". https://developers.openai.com/api/docs/concepts · https://developers.openai.com/api/docs/guides/your-data — OpenAI's business privacy page, https://openai.com/enterprise-privacy/, carries the equivalent statement for ChatGPT Business and Enterprise but returned HTTP 403 to automated retrieval on September 20, 2026, so it is not quoted here. Anthropic, Pricing (retrieved September 20, 2026): "None by default" for model training on Claude Team and Enterprise content. https://claude.com/pricing ↩ ↩²

8. Kelly D. Mullins, "Accounting ethics in the age of AI", Journal of Accountancy, August 5, 2026 (retrieved September 20, 2026), on ET sec. 1.700.001, the Confidential Client Information Rule: "Before pasting client data into any AI system, stop and ask: Is this confidential client information?" and "When consent is required, make sure your client understands what information may be used, why it's being used, and who may have access to it." https://www.journalofaccountancy.com/issues/2026/aug/accounting-ethics-in-the-age-of-ai/ ↩

9. 26 U.S. Code § 7216(a), via Cornell Legal Information Institute (retrieved September 20, 2026): a person engaged in the business of preparing returns who "knowingly or recklessly ... discloses any information furnished to him for, or in connection with, the preparation of any such return" or "uses any such information for any purpose other than to prepare, or assist in preparing, any such return" is "guilty of a misdemeanor, and, upon conviction thereof, shall be fined not more than $1,000 ... or imprisoned not more than 1 year, or both, together with the costs of prosecution." https://www.law.cornell.edu/uscode/text/26/7216 ↩

10. Internal Revenue Service, Publication 4557, Safeguarding Taxpayer Data: A Guide for Your Business (Rev. 6-2024; retrieved September 20, 2026): "According to the FTC Safeguards Rule, tax return preparers must create and enact security plans to protect client data. See Publication 5708 for information on creating a written information security plan. Failure to do so may result in an FTC investigation." Among its basic steps: "Limit access to taxpayer data to individuals who need to know." https://www.irs.gov/pub/irs-pdf/p4557.pdf ↩

11. DataSnipper, 5 Pros and Cons of Using ChatGPT for Accounting (retrieved September 20, 2026): "ChatGPT doesn't interpret your audit methodology, client risk, or internal controls"; outputs "typically require manual verification" and lack source citation. Named automation includes Test of Details, payroll testing and AR cut-off testing. https://www.datasnipper.com/resources/ai-in-accounting-chatgpt-limitations ↩