ChatGPT for accountants: where it helps, where it invents, and what stays out of it

For an accountant, ChatGPT is quickest at the words around the numbers: the client email, the engagement-letter paragraph, the plain-English version of a notice. This guide covers what each plan costs, the work it does well, the mistakes OpenAI itself warns about, what to keep out of it, OpenAI's own settings step by step, and prompts for client work.

This page is the ChatGPT chapter of our AI for accountants map, which covers ledger, tax research and bookkeeping tools and the professional rules in full. The question here is narrower: you have a ChatGPT account, or you are about to open one, and you want to use it on real engagements without a client's return ending up somewhere it shouldn't.

Equerry, which publishes this guide, makes an assistant that keeps track of client correspondence and calls. It does not prepare returns, keep books or research tax law, and it comes up only near the end. Nothing here is tax, accounting or legal advice. What this page says OpenAI documents is taken from its own help center and pricing page, read in a browser on 1 October 2026.

What ChatGPT costs a solo accountant, plan by plan

ChatGPT is OpenAI's general-purpose assistant, on the web and in apps for iPhone and Android; it is not tax software or a ledger. For a small practice the plan decides how much you can use it, how much it remembers, and whether your conversations can train OpenAI's models by default. Dollar figures come from OpenAI's help articles; the pricing page shows local currency.

OpenAI's Go article adds that it does "not support annual billing or the option to pay for multiple months in advance for ChatGPT Go, Plus, or Pro subscriptions". For one practitioner the realistic options are a personal plan with the training switch turned off, or, if there are at least two of you, a Business workspace that is excluded from training by default. The plan you pick does not answer whether a client's information may go in.

Where ChatGPT saves an accountant time

It earns its keep on language, not numbers.

• Explaining a notice or a standard in plain English. Paste the IRS's public explanation of a notice and ask for a version a business owner can follow; check it against the source.
• Engagement letters and client emails, with placeholders. A scope paragraph, a fee-change email, a document reminder, with [Client], [Entity], [Year] and [Amount] in place of the real details.
• Excel help on a dummy file. Describe a workbook's layout, or upload a copy with invented figures, and ask for the formula.
• Working through public guidance. An IRS publication chapter or a revenue procedure pasted in full: ask for the main points, each tied to the sentence it comes from.

Where it goes wrong: figures, citations and code sections

OpenAI's help article "Does ChatGPT tell the truth?" (read 1 October 2026) is direct about it: "Sometimes, it might sound confident—even when it's wrong." The errors it lists include "Fabricated quotes, studies, citations or references to non-existent sources" and "Incorrect definitions, dates, or facts." In accounting work they land in particular places.

• Authority that doesn't check out. A Code section, regulation paragraph or revenue ruling it names may not say what it claims, or may not exist. Treat every citation as a lead to pull and read.
• Figures in running text. OpenAI says its data analysis tool "Enables accurate calculations", and still advises: "Always verify quotes, data, technical information or references to external documents." Recalculate any figure that will reach a client or a return.
• The wrong year. OpenAI: models "are trained on data up to a certain point and responses do not incorporate information about events beyond that, unless tools are used," though search tools "are enabled for all models by default." Ask which year an answer assumes and confirm it at the source.
• It sees only what you give it. The sources OpenAI's Memory article lists are your chats, files and connected apps, not your ledger or tax software unless you upload or connect them. Asked anyway, it can give one of the "Overconfident answers to ambiguous or complex questions" OpenAI warns about.

OpenAI's own advice is the working rule: "Use ChatGPT as a first draft, not a final source."

What to paste into ChatGPT, and what to keep out

The test under every row: could it identify a client, or did it come from their return? Your firm's policy overrides the table.

Two sets of rules bear on that line. The IRS describes the Treasury regulations under IRC §7216 as "rules and consent requirements relating to the disclosure or use of tax return information by tax return preparers" (Section 7216 information center, read 1 October 2026). And for AICPA members in public practice, the AICPA Code of Professional Conduct has a Confidential Client Information Rule (1.700.001); beside it, the Code notes that "A nonauthoritative article, Ethics Staff Insights: AI through an ethics lens, discusses threats to compliance with the code when members use AI to provide services" (Code updated through September 2026, read 1 October 2026). Check how both apply with your firm, and check whether consent is required before client information goes into any tool.

Connecting an account is a different act from pasting. OpenAI's Memory article counts "content from connected apps, such as Gmail" among the sources Memory can use, and says disconnecting an app "does not delete past conversations that already used app content." For a practitioner whose inbox carries every client, that grant is wider than any paste; more in can ChatGPT read my emails.

What OpenAI says happens to what you type

No vendor setting decides your confidentiality obligations. What OpenAI documents is how training and retention work (help center, read 1 October 2026):

• The training switch covers new conversations. "When Improve the model for everyone is off, your new conversations won't be used to train OpenAI models." It removes nothing: turning it off "does not delete or hide saved chats."
• A rating reopens one conversation. After you opt out, if you give a thumbs up or thumbs down, "the entire conversation associated with that feedback may be used to train OpenAI models."
• Temporary chats are kept for a while. "OpenAI may keep a copy of a temporary chat for up to 30 days for safety purposes."
• Firm workspaces start from a different default. "By default, OpenAI does not use content from ChatGPT Business, Enterprise, Edu, or ChatGPT for Healthcare workspaces to train its models," and "Your organization's privacy, retention, and access policies still apply."

Nothing there calls a conversation confidential, and this guide doesn't either.

OpenAI's own steps for three settings

Quoted from OpenAI's help articles, read 1 October 2026. If your screen differs, your screen wins.

Training, from "Data controls in ChatGPT". Web (signed in):

1. "Open your account menu."
2. "Select Settings."
3. "Select Data controls."
4. "Select Improve the model for everyone, turn it off, and select Done."

iOS and Android:

1. "Open the sidebar."
2. "Select your profile icon to open Settings."
3. "Select Data controls."
4. "Turn off Improve the model for everyone."

A chat outside history and memory, from "Temporary chat in ChatGPT":

1. "Open a new chat."
2. "Select Temporary."
3. "Choose Personalized or Unpersonalized before you send your first message."

Unpersonalized "Does not use memory, custom instructions, or plugins," and "You cannot change personalization after the conversation starts." Background in our guide to ChatGPT Temporary Chat.

Memory, from "Memory in ChatGPT":

1. "Open Settings."
2. "Select Personalization."
3. "Select Memory."

Where client names have already reached regular chats, OpenAI cautions: "Deleting a chat alone does not necessarily delete a separate saved memory created from that chat." Our walkthrough: how to delete ChatGPT memory.

ChatGPT prompts for accountants' client work

Each gives ChatGPT a narrow job, uses placeholders for client details, and asks for output you can check.

• Explain a notice. "Here is the IRS's public explanation of a [notice type]: [paste]. Rewrite it for a business owner with no tax background, under 250 words. Add nothing that isn't in the text, and mark every step that depends on the taxpayer's facts."
• Chase the documents. "Write a short, friendly email to [Client] listing the documents we still need for the [Year] return: [list]. Ask for them by [Date] and explain in one sentence why the date matters. Include no figures."
• Get the formula. "My workbook has transaction dates in column A, vendor names in column B and amounts in column C. Write a formula that totals column C by vendor and by month, and explain each part so I can check it."
• Summarise public guidance you can verify. "Summarise the [publication chapter / revenue procedure] pasted below in eight points, quoting the sentence each point comes from. Where a point needs a judgment the text doesn't make, say so."

None of them asks for a figure to rely on, a tax position, or an email to send unread.

Claude for accounting

Claude, Anthropic's general assistant, has its own plans, prices and data settings, documented in Anthropic's help center. Our guide to Claude for accountants covers it in the same order as this page.

What stays outside ChatGPT: the client record, thread by thread and call by call

ChatGPT can remember, so the gap is not memory but what reaches it: what you type, paste or upload, or an account you connect. Typing falls behind when a client replies mid-review; connecting grants the account, not one engagement. And what you can audit afterwards is largely a summary: in OpenAI's words the memory summary is "a high-level view" that "does not necessarily include every detail or source that Memory can reference."

The record an engagement actually runs on sits between those two: did the controller send the fixed-asset schedule, what did the owner say on Tuesday's call about the year-end bonus, what did we promise to have ready before the partner review? That is the client side of the practice, covered in our practitioner's map of AI in an accounting firm.

Equerry is a personal assistant who helps you get things done. In an accounting practice its job is the client record: it keeps what you send it and briefs you from that, while the books, the return and your tax software stay where they are. The ways in:

• Copy it on the engagement email. Your Equerry has its own mail.getequerry.app address: CC it on a live thread with the client, or forward the thread once it has happened; replies that keep the address on the thread keep arriving. It never asks for your inbox, and the original stays in your own mailbox. Ask about it later. What Equerry remembered from it is listed in Settings → Your memory, each line marked as from that email, and you can delete any line.
• Say it out loud after the call. When a client call ends, record a short voice note in Equerry: what was agreed, what the client is sending, what you owe them. Equerry is not on the call and records nothing during it; the note is one you make afterwards.
• Send the PDFs. Share a PDF or a screenshot to Equerry from your iPhone's share sheet. Before tax return information goes to any tool, Equerry included, check §7216 consent and your firm's policy.
• Type a note. After a meeting, write down what matters. When an answer draws on a note you saved, it carries a source chip: tap it to see the passage Equerry used and open the original.
• Ask before the next meeting. "What is still outstanding on the Reyes engagement?" is answered from what you sent it. If you allow calendar access and set up Meeting Prep, you get a short brief ahead of each client meeting, drawn from what you have shared.

The boundaries are short. It works only from what you send, and from your calendar if you allow that. Copying it on a thread shows its address to everyone else on that thread. Your material is processed so it can be searched, partly by the outside providers listed on our AI disclosure page. Equerry claims nothing under §7216, the AICPA Code or any other rule; sending it client information is the same decision you make for ChatGPT, taken under your firm's policy and with the client's consent where that is required. It is a narrower decision: chosen threads, notes and files rather than a connected account.

For the same questions in a law practice, see ChatGPT for lawyers. Keep ChatGPT for the plain-English draft and the formula; use Equerry for what each client told you. It is on the App Store for iPhone and iPad.

Sources, read 1 October 2026 in a browser. OpenAI Help Center: Does ChatGPT tell the truth?; Data controls in ChatGPT; Temporary chat in ChatGPT; Memory in ChatGPT; What is ChatGPT Plus? (US Plus price); About ChatGPT Pro tiers (the three Pro plans and their US prices); What is ChatGPT Go? (Go billing; no annual billing for Go, Plus or Pro). OpenAI, ChatGPT pricing (prices shown in local currency). Internal Revenue Service, Section 7216 information center. American Institute of CPAs, AICPA Code of Professional Conduct (updated for all official releases through September 2026), 1.700.001. OpenAI changes its products and pages often; where its current page disagrees with this one, OpenAI's page is right. This guide is general information, not tax, accounting or legal advice.

Last updated 2026-10-02.

All product names, logos, and brands are property of their respective owners. Equerry is not affiliated with, endorsed by, or sponsored by any other product mentioned; comparisons are provided for informational purposes.