Can you put client information into ChatGPT? A guide for accountants
Short answer: not as-is. Client names, SINs, SSNs and account numbers are confidential, and pasting them into a consumer AI tool is a disclosure your professional rules still cover. You can still use AI on the work if the identifying details never reach it.
Your confidentiality rules already apply to AI
There's no special carve-out for chatbots. CPA Ontario says plainly that the Code of Professional Conduct's rules on confidential information apply to AI use, with the same accountabilities as any other disclosure. Its example: a staff member uploads confidential client data to an AI tool without the firm knowing, errors reach the audit file, and the client files a complaint with the regulator.
The same piece notes that a lack of a firm policy doesn't mean AI isn't being used. It cites a KPMG survey in which 48% of Canadians said they'd used AI at work in ways that may not match their employer's guidelines.
US preparers: Section 7216 covers tax return information
If you prepare US returns, Internal Revenue Code Section 7216 restricts disclosing or using tax return information, which includes identifying, financial, employment and investment details a client gives you for the return. Knowing or reckless violations can bring a fine of up to $1,000 and up to a year in prison, with a parallel civil penalty under Section 6713.
Disclosures to providers that help with preparation, processing or e-filing can be allowed without consent, but disclosures for other purposes or outside the US generally need the client's written consent first. A general-purpose chatbot doesn't fit neatly into the allowed categories, which is why most firms treat it as off-limits for identifiable return data.
What happens to what you paste
| ChatGPT plan | Used to train models? | What you can do |
|---|---|---|
| Free, Plus, Pro (personal) | Can be, unless you turn it off | Settings › Data controls › turn off "Improve the model for everyone" |
| Temporary Chat | Check OpenAI's current terms | Not saved to your history, but may be kept up to 30 days for safety review |
| Business, Enterprise, Edu | Not by default | Still stored by OpenAI under your workspace's terms |
Even with training off, the text is still sent to and stored by a third party. Turning off training reduces one risk; it doesn't make the disclosure go away.
The safe way: send the work, not the identities
AI doesn't need to know who your client is to explain a deduction, summarize a statement or draft a letter. Replace each identifying detail with a consistent placeholder before you paste, then swap the real values back into the answer.
SIN: 046 454 286 → SIN: [ID_NUMBER_1]
RRSP to account 4471-302-118 → RRSP to account [ACCOUNT_1]
Consistent tags matter. If every name becomes "XXXX," the AI can't tell the client from the spouse and the answer gets vague. With [PERSON_1] and [PERSON_2] it keeps them straight.
A checklist for your firm
- Write a one-page AI policy: which tools are allowed, and that identifiable client data never goes into them unmasked.
- Turn off model training on any personal AI accounts staff use for work, or move to a business plan.
- Mask names, SIN/SSNs, account numbers, addresses and dates of birth before anything is pasted or uploaded.
- Review the AI's output like a junior's work: check every figure and reference.
- Keep a short record of what was masked and sent, in case a client or regulator asks.
Do this in one step with Stand-In
Drop in a PDF, scan, Word or Excel file. Stand-In swaps every name, SIN/SSN, account number and address for a tag like [PERSON_1], then puts the real details back in the AI's answer. It runs on your device; nothing is uploaded.
Sources: CPA Ontario, The hidden risk of shadow AI · The Tax Adviser, The many implications of Sec. 7216 · IRS, Section 7216 FAQs · OpenAI, Data controls FAQ