ChatGPT for accounting: how a practice uses it across its work, and the client record it never holds

If you run or manage an accounting practice, the question is not whether one accountant can get a formula out of ChatGPT. It is how the firm uses it: on which plan, for which engagements, under what rule for staff, and with what review before a client reads the result. This page answers that from OpenAI's own pages and the profession's publications, read on 27 October 2026, then names the one record no plan holds: what each client sent and still owes.

Short answer: accounting practices use ChatGPT for the writing around an engagement, for internal procedures and training, and for first-pass summaries that someone then checks. The firm's decisions are the plan (a Business or Enterprise workspace is excluded from training by default and puts admins in charge of retention and access), the paste rule, and the review step. This is the firm-level chapter of our map of AI for accountants; the individual accountant's settings and prompts for client work live on ChatGPT for accountants and are not repeated here.

A disclosure first. This page is published by Equerry, which makes a personal assistant for iPhone: not accounting software, no ledger or accounting-system connection, and it appears only in the last section. Equerry is not affiliated with OpenAI, and nothing here is tax, accounting or legal advice. Every OpenAI statement below was read at help.openai.com or openai.com on 27 October 2026.

How accounting practices use ChatGPT across their work

The most-cited firm-level piece on the subject is Cetrom's, titled "How Accounting Firms Should (And Should Not) Use ChatGPT" and dated 28 December 2023 (read 27 October 2026). Its advice to firms is temperamental rather than technical: "the best course of action is cautious experimentation with a healthy dose of skepticism", "Anything ChatGPT-generated must be carefully vetted for accuracy", and "Don't trust it based on prestige, novelty, or popularity alone." On client data it warned that "if accountants input sensitive data in ChatGPT, it could become part of the database it is trained on" — true of a personal account with training on, and the reason the plan question below matters.

Read that piece beside the prompt libraries and practitioner threads that rank with it, and the firm-level uses sort into three:

• Client communication the firm sends in its own name. The scope paragraph in an engagement letter, the fee-change note, the reminder about outstanding documents, the explanation of a notice in words the client's owner will follow. The firm decides the house style once, writes it into a project's instructions, and every draft comes out in it.
• The firm's own machinery. Onboarding checklists, a close calendar written out as procedures, the acceptable-use policy for the tool itself, training material for new staff. No client in any of it, so it is where most firms start.
• First-pass reading. A long revenue procedure or a standards update pasted in whole and summarised with the sentence behind each point, so a reviewer checks the points against the text rather than reading cold.

What the firm-level sources agree on is the review step: a partner or manager reads the output against the source before a client does. Prompt libraries are a way to standardise the ask; the review is what makes the firm's name safe to put under the result.

ChatGPT prompts for accountants, seen from the firm

A prompt library is an individual's tool until the firm adopts it, and then it is a template like any other: written once into a project's instructions, kept with placeholders where client details would go, owned and reviewed like the engagement-letter master. The prompts themselves sit on the accountant's page linked above; this page stays with the firm's decisions around them.

ChatGPT accounting: what the business plans change for a firm

The individual's plan question is cost and usage. The firm's is control: who can see what, how long it is kept, and whether any of it trains a model. The table reads straight from OpenAI's pages on 27 October 2026; prices are US dollars before tax, and OpenAI notes that "pricing may vary depending on your country and currency type". Zero Data Retention and BAAs are sales-led, not part of self-serve Business.

Two lines in that table are the firm's to act on. The admin line means a partner can read a junior's conversations; say so in the acceptable-use policy. And retention is a lever, not a default: OpenAI's enterprise-privacy page adds that "shorter retention periods may compromise product experience", so a short window trades history for exposure, on purpose.

Projects: one per engagement, with its memory kept inside

The feature a practice leans on is Projects, which "keep related chats, files, and instructions together". On Business and Enterprise, OpenAI's Projects article, read 27 October 2026, allows "40 files per project", and its retention article says that "Files added to a project, including a shared project, are retained until the project or file is deleted or an applicable workspace retention policy removes them." Project instructions "override your global custom instructions", so the house style lives there. A project can be shared with the engagement team with chat or edit access, and "Shared projects automatically use project-only memory and cannot access an individual member's memories or context from outside the project."

Project-only memory is the setting that keeps one client's project from colouring another's. OpenAI's steps, from the same article: open the project, select the more options menu (•••), then Project settings, then under Memory select Project-only memory, then Save. "Changes may take a few hours to take effect", and on Business or Enterprise it "is available when Memory is enabled for both your workspace and your account". There is no global switch: "Each project's memory must be set individually, and also respects your personal and workspace memory settings."

One limit from OpenAI's own FAQ belongs in the firm's expectations: "Project memory does not show a list of memories like personal memory." You can delete or move a chat; you cannot read off what the project has retained about the client.

Files, in the sizes a firm actually sends

OpenAI's upload article, read 27 October 2026, sets the limits a practice bumps into: "CSV files and spreadsheets have a size limit of approximately 50 MB, depending on the size of each row", "up to 80 files every 3 hours", and "Shared upload storage is capped at 25 GB per user and 100 GB per organization" — capacity figures, none of them permission, and whether a client's file belongs in the workspace at all is the next section but one.

What the professional bodies publish

The AICPA's magazine, the Journal of Accountancy, ran a Professional Liability Spotlight on 1 April 2025 under the title "Should I disclose my use of gen AI to clients?", written by a risk consulting director at CNA, one of whose companies underwrites the AICPA Professional Liability Insurance Program. Its starting point is worth quoting exactly, because firms often assume the opposite: "no specific federal law or professional standard applicable to CPAs mandates disclosure when generative AI is used or if client data is used within it." What does bear on the question, in its list: the Gramm-Leach-Bliley Safeguards Rule, which "requires 'financial institutions' — which includes CPA firms — to safeguard nonpublic personal information"; state laws, naming California and Utah; the AICPA Code's Confidential Client Information Rule; and, for tax preparers, Section 1.3, Data Protection, of the Statements on Standards for Tax Services.

Its practical advice is a list of questions a firm should be able to answer before a client asks: "how their information is stored; whether their data is used to further train the generative AI model; who has access; whether their data is shared outside of your organization; how data is retained, by both the firm and any external provider(s)". On engagement letters it borrows the American Bar Association's 2024 opinion that "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use generative AI is not sufficient", and suggests a disclosure and consent cover the firm's quality-control procedures, consent to disclosure to third parties, the firm's responsibility for the client's confidential information, and "A request for the client's consent to engagement performance under those conditions." Where tax return information is involved it points to Regs. Sec. 301.7216-2 and -3.

That is what the publication says. Whether your firm must disclose, obtain consent, or change its engagement letter is a question for your counsel under the rules that bind your practice; this page draws no conclusion about any of them.

A three-partner firm's engagement: what ChatGPT drafts, and what it cannot know

An illustration, invented for this page. Marsh, Tennant & Okafor is a three-partner practice: Marsh runs review and compilation work, Tennant tax, Okafor advisory and outsourced controllership. They run a ChatGPT Business workspace on Standard seats with one project per engagement, each set to project-only memory, and a one-page rule that no client identifier enters a chat.

The engagement is the year-end review for Brightwater Cold Storage, a refrigerated-warehouse operator whose bank wants the reviewed statements by 15 March. Over the engagement, the firm's project does three jobs well:

• The scope addendum. Brightwater's owner asks Okafor, on a call, to add a monthly close for the new site. Okafor pastes the firm's standard addendum with [Client], [Service], [Fee] and [Start date] as placeholders and asks for a version that adds the service without touching the review scope; Tennant reads it against the master letter before it goes.
• The document chase. A courteous reminder to a controller who has sent two of five requested schedules, listing the three outstanding by name with a date and no figures. The firm will reuse it, so it goes into the project's sources.
• The first read of a lease-standard memo. A public implementation memo pasted whole and summarised in ten points with the sentence behind each; Marsh checks three against the text and strikes one.

Then the questions the project cannot answer, because nothing in it knows:

The firm's own good rule created the gap: the more carefully a practice keeps client identifiers out of ChatGPT, the less it can know about any one client, and the more the engagement runs on what three partners remember and can find in their mail. Not a flaw in the tool; the shape of the decision.

What never to paste, and where OpenAI's data terms are explained

One sentence, firm-wide: nothing that identifies a client or came from a client's return enters a chat on any plan until the firm has approved the tool and plan for it and checked whether consent is required. What OpenAI publishes about training, retention and who can read a conversation, plan by plan, is set out in our guide to ChatGPT privacy; it describes OpenAI's handling and leaves your firm's obligations where they were.

The client record across the practice

Every prompt library, and every plan in the table above, is about what the firm says to ChatGPT. The record an engagement runs on is what the client said to the firm: the schedule in the fourth email of a thread, the date the owner moved on a call, the fee a departed senior agreed. It lives in partners' mailboxes and heads, and the firm's paste rule, rightly, keeps it out of the workspace.

Equerry is a personal assistant on your iPhone who helps you get things done. For a partner in an accounting practice its job is that record, not the books: it reads only what you send it, does no accounting, and connects to no ledger, tax or practice-management system. Each partner has their own; it keeps the client record for the clients that partner looks after.

The contrast with the workspace is the unit of sharing. Connecting a mailbox to any assistant grants the account; Equerry takes the one thread you put it on, so what it holds about Brightwater is exactly the correspondence you chose. It reads the thread, not your inbox, and the original stays in your own mailbox.

The honest limits. Equerry never asks for your inbox and cannot find a reply you did not CC or forward. Its address is visible to everyone on a thread it is copied on, so forward instead where that would be awkward. It does not join or record calls, send email, or act in a client's accounts. Sending something to Equerry is still sharing it: what you send is processed by the outside providers named on our AI disclosure page, so the firm's paste rule and the §7216 question apply to it exactly as they do to ChatGPT. Equerry makes no claim about meeting the AICPA Code, the Safeguards Rule or any other rule. Equerry Pro is $39.99 a month or $329 a year.

Keep the workspace for the drafting the firm signs; keep Equerry for what each client told you. For the same questions in a law practice, see ChatGPT for lawyers; for how far ChatGPT itself can see into a mailbox, can ChatGPT read my emails?

Sources, all read 27 October 2026. OpenAI Help Center: ChatGPT Business overview (seat prices, two-seat minimum, training, the Team rename, sales-led options); Projects in ChatGPT (file limits, shared projects, project-only memory and its steps); Uploading files and audio to ChatGPT (spreadsheet size, rate and storage caps); Chat and file retention in ChatGPT (retention of project files); Memory in ChatGPT. OpenAI, Enterprise privacy at OpenAI (updated 8 January 2026; training default, admin access, retention, SOC 2 Type 2, DPA). Cetrom, How Accounting Firms Should (And Should Not) Use ChatGPT (28 December 2023). Journal of Accountancy (AICPA & CIMA), Should I disclose my use of gen AI to clients? (1 April 2025). The firm, client and people in the example are invented. 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 2026-10-28.

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.