01Set staff rules
What Client Information Can Staff Put Into ChatGPT and Other AI Tools?
Which ChatGPT, Copilot and Gemini accounts are safe enough for client work, what each vendor says about training, and how a yes gets decided.
- Topic
- Set staff rules
- Written by
- AI Ground Rules
- Last reviewed
- Sources
- 7 checked sources
Client information can go into ChatGPT, Microsoft Copilot, Gemini or any other AI tool only when the exact tool, plan and work account is approved for that kind of data and that task, and the client has not restricted it. The data decision owner gives that yes, and records it. Until then, the answer is “not yet.”
There is one exception. Public material your firm may use, such as your published pages or public guidance, is fine in an approved tool without a separate yes. Client information never counts as public just because a client sent it to you.
This page covers the account question: which kinds of account are safe enough to consider for client work, and how a yes gets decided. For the full staff rules on what may go in, see the AI acceptable use policy.
Three words to know
- Work login: you are signed in with your firm’s account (for example Microsoft 365, Google Workspace or a ChatGPT Business workspace), not a personal email address.
- Tenant: your firm’s own private space inside a cloud service such as Microsoft 365, holding your users, files and settings, kept apart from other customers.
- Training on your data: the vendor uses what you type or upload to improve its future AI models, so your content is kept and reused beyond your own chat.
Which accounts are safe enough to consider?
This table shows what each vendor says on its own pages about training. Checked 14 November 2026. Vendors change plans and names, so check the linked page again before you rely on it.
| Account | Used to train models by default? | Caveats |
|---|---|---|
| ChatGPT Free, Go, Plus or Pro (personal account) | Yes. OpenAI’s pricing page says “Content is used to train our models”, with an opt-out. | A personal account is not an approved tool, even with training switched off. Do not use it for client information. |
| ChatGPT Business, signed in to your firm’s workspace | No. OpenAI’s Business page says “Your data is never used for training our models.” | Not training is not the same as approved. Your firm still decides which data types and tasks are allowed. |
| ChatGPT Enterprise, signed in to your firm’s workspace | No. OpenAI’s Enterprise page says business data “is not used to train our models by default.” | As above. Check the settings your admin has chosen. |
| Microsoft Copilot Chat, signed in with your work account | No. Microsoft says prompts and responses aren’t used to train foundation models and get enterprise data protection. | Look for the green shield next to New Chat. Web search queries go to Bing, which sits outside the Microsoft 365 service boundary and has its own data terms. Microsoft says those queries have user and tenant identifiers removed. |
| Microsoft Copilot (the paid licence), signed in with your work account | No. Prompts, responses and the files it reads through Microsoft 365 aren’t used to train foundation models. | It can read the emails and files you already have access to, so overshared SharePoint or OneDrive folders become an AI problem. Same Bing caveat for web search. |
| Gemini in a qualifying Google Workspace edition, work account | No. Google says Workspace does not use customer data for training “without customer’s prior permission or instruction”, and content is not human reviewed for training outside your domain (Privacy Hub). | Applies only to qualifying editions. The page does not cover personal Google accounts, so treat Gemini on a personal account like a personal ChatGPT account. |
Microsoft renamed Microsoft 365 Copilot to Microsoft Copilot, and Microsoft 365 Copilot Chat to Microsoft Copilot Chat. Licences and screens may still show the old names for a while.
A “no” in the training column is a starting point, not a yes. It tells you the account is worth considering. It does not tell you that a given client’s files may go into it.
How a yes gets decided
The data decision owner answers four questions and writes the answer in the approved tools register:
- Which tool, plan and work account? Name the exact product and plan, not the brand. ChatGPT Plus and ChatGPT Business are different answers.
- Which data types? Use the shared groups in the acceptable use policy: fine, ask first, never unless approved, and absolute never. When groups overlap, the stricter one wins. Passwords, access codes, private keys, sign-in links and MFA codes never go in, with no exception.
- Which tasks? For example “summarise a client’s emails for an internal note” is a different yes from “draft a reply to the client.”
- What do clients allow? Check engagement letters, contracts and written client instructions. A client restriction beats any internal approval. If you cannot find an answer, ask the person who manages the client relationship. Silence is not a yes.
Then staff use the minimum information the task needs. Removing names is not enough on its own: figures, file names, logos and context can still identify a client. Prefer invented details where they do the job. A person checks the AI-assisted work before it goes to a client.
Professional rules can decide it for you
Some professions have their own rules on sharing client information. Two examples:
- Australian tax and BAS agents. Xero’s summary of the Tax Practitioners Board guidance on AI says AI tools may be treated as third parties, so practitioners need client permission before using them with client information. That permission can be a signed engagement letter or other signed consent.
- US tax preparers. Federal rules on using and disclosing tax return information (IRC section 7216) may apply. Check with your adviser before any client tax information goes into an AI tool.
Already pasted it?
If client information went into a tool or account that is not approved for it:
- Stop. Do not carry on with the task or paste anything more. If a password or code went in, change it straight away. Keep a record: do not delete the chat unless the data decision owner says so.
- Tell the data decision owner (or their backup) through [channel]. Say which tool and account, roughly when, and what kind of information was involved. Do not copy the sensitive content into your message.
- Follow their instructions. They decide whether to delete the chat, tell the client or check the contract. Do not assume deleting the chat fixes it.
The acceptable use policy has the full incident steps.
AI built into tools you already use
Client information can reach AI without anyone pasting it.
- Accounting and practice software. Accounting and practice apps can include AI features that read the records already stored there. Check whether the feature is switched on, what it reads, and whether it is in your approved tools register. Treat it like any other tool.
- Copilot reading SharePoint and email. Microsoft Copilot works across files and emails each person can open. If a client folder is shared too widely, Copilot can surface it to anyone it is shared with, including people who should not see it. Fix the sharing first. The Microsoft 365 Copilot checklist walks through it.
- Connected apps. Meeting note-takers, mailbox or file summarisers, browser extensions and any app given access to your accounts need approval like any tool. Meeting recording follows your firm’s recording procedure, including notice and any consent the law requires. List each one in your AI inventory.
Fictional examples
These scenarios use invented firms and people.
| Scenario | Group | Decision |
|---|---|---|
| A bookkeeper at Harbor Accounts drafts a newsletter in the firm’s ChatGPT Business workspace from articles already on the firm’s website. | Fine | Allowed. The partner checks it before it goes out. |
| A consultant at Northgate Advisory wants to paste a client email into Copilot Chat to draft a reply. The email ends with the client’s full signature block: name, direct line, mobile and office address. | Ask first | Copilot Chat with a work login is approved for client emails, and the client contract allows AI help. The consultant deletes the signature block first, since the task does not need it. |
| A recruiter at Lantern Talent takes a screenshot of a candidate spreadsheet, blacks out the names and wants to paste it into Gemini on her personal Google account. | Never, unless approved in writing | Not allowed. A personal account is never approved, and the screenshot still shows salaries, job titles and towns that could identify people. Salaries put it in the stricter group. She uses invented rows to get help with the formula. |
| An office manager wants AI to check a payroll export for errors. | Never, unless approved in writing | Default no. The data decision owner has not approved a named tool for this task. |
| A project coordinator wants an AI browser extension to fill in a supplier portal using a saved password. | Absolute never | Not allowed. Sign-in details never go into an AI tool, and the extension itself is not approved. |
What this guide does not cover
This guide helps you set a practical business rule. It does not decide your legal obligations under privacy law, professional rules or client contracts. Have a qualified adviser review your rules for regulated work or sensitive personal information.
02Evidence
Sources
- OpenAI, ChatGPT pricing (privacy comparison: Free, Go, Plus, Pro)Checked
- OpenAI, ChatGPT BusinessChecked
- OpenAI, ChatGPT EnterpriseChecked
- Microsoft Learn, Enterprise data protection in Microsoft Copilot and Microsoft Copilot ChatChecked
- Microsoft Learn, Microsoft Copilot Chat privacy and protectionsChecked
- Google Workspace Admin Help, Generative AI in Google Workspace Privacy HubChecked
- Xero, TPB guidance on AI: what accountants and bookkeepers in Australia need to knowChecked