AI Privacy for Accountants: Client Data in AI Tools
Staff already paste client details into AI tools. See what each provider keeps, which duties apply to tax practices and where to set your rules.
- Published
- Updated
Client onboarding, simplified
Less time chasing client details, more time preparing returns.
Made for Australian accounting firms
AI privacy depends on the plan your practice uses and the information it sends. Approve the account’s data terms before staff enter client details.
Use the TPB AI guidance to decide what your professional duties require. An AI policy template turns those decisions into staff rules.
For staff using OpenAI, the ChatGPT privacy guide explains its personal and business accounts. A paid personal subscription can still have consumer data terms.
What Happens to Client Data You Enter Into AI
An AI provider can use a prompt to train models, retain it, let people review it or store it overseas. Each question needs its own answer.
- Training uses submitted content to change how a model answers future users. A business agreement can exclude this by default.
- Retention is how long the provider keeps prompts, responses and uploaded files. Saved history, safety records and backups can have different lifetimes.
- Staff review means people at the provider or its contractors can read content for permitted purposes, such as support or abuse investigations.
- Storage location is where copies sit. The place a model processes a request can differ from the place it stores the conversation.
Deleting a conversation removes it from your history first. Backend deletion can take longer, and legal holds or safety investigations can preserve copies.
OpenAI’s chat retention policy describes that distinction. Files saved separately in its Library or Space need their own deletion.
Turning training off does not rule out staff access. OpenAI’s enterprise privacy terms permit limited human access, with different conditions for Business and Enterprise.
Compare AI Tools on Client Data
Compare the business service your practice would actually use. A provider’s consumer chatbot and its managed business workspace can have different promises under the same brand.
OpenAI’s model-training policy allows training on eligible personal-account content. Business and Enterprise content is excluded by default.
Business AI Data Terms
| Provider | Service compared | Training default | Retention after deletion | Storage location |
|---|---|---|---|---|
| ChatGPT | Business or Enterprise workspace | Off by default | Normally within 30 days, with stated legal or security exceptions | Business region selection where available, with safety copies in the US. Enterprise has separate regional options |
| Claude | Team or Enterprise under commercial terms | Off by default. Submitted feedback can permit training | Deleted chats removed from backend within 30 days, with policy and legal exceptions | Stored in the US. Processing can happen elsewhere |
| Gemini | Qualifying Google Workspace edition | No model training outside your organisation without permission | Workspace customer-data deletion can take up to 180 days after an unrecoverable deletion instruction, subject to legal retention | Eligible data-region settings can select US or EU storage and processing |
| Copilot | Microsoft 365 Copilot with a work account | Prompts and responses do not train foundation models | Practice retention policies and legal holds can preserve deleted interactions | Stored under Microsoft 365 commitments. Processing can cross regions |
| DeepSeek | DeepSeek’s hosted chatbot | Policy permits training and improvement using submitted data | Purpose-based retention, without a fixed post-deletion period in the privacy policy | People’s Republic of China |
| Perplexity | Enterprise service | Enterprise terms prohibit training on customer content | DPA retains input until customer deletion or up to 30 days after service ends, subject to legal retention | Enterprise terms do not promise an Australian storage region |
The safer starting point for client work is a practice-managed business workspace with training excluded, limited access and agreed retention. No brand makes every client upload acceptable.
If your practice already uses Microsoft 365 or Google Workspace, its managed AI can simplify staff administration. Review existing folder permissions before giving either service access to client documents.
ChatGPT Business and Claude Team also exclude training by default. Their different rules for human access, feedback and overseas copies still affect what your practice can approve.
OpenAI’s Business storage explanation separates regional storage from US safety copies. Choosing a region does not fix where requests are processed.
Anthropic’s commercial training policy treats submitted feedback separately. Disable chat-rating feedback for client work when using that service.
Google’s Workspace privacy hub distinguishes app histories and retention settings. Microsoft’s Copilot retention documentation explains why deleted chats can remain discoverable under a retention policy.
AI Inside Existing Software
Acrobat AI Assistant works with document content. Grammarly works with writing. Include both in your approved-tool review even when staff never open a chatbot website.
Windows Recall needs a separate device decision because it saves screen snapshots locally. Apple Intelligence has both on-device processing and cloud features, so approve the feature as well as the device.
Meeting Recorders
Meeting recorders such as Otter handle conversations involving clients and other participants. Set rules for permission, recording access and deletion before staff invite a recorder to a client meeting.
The Duties That Decide What AI May Receive
Permission to use an AI account does not itself authorise disclosure of client information. Apply the professional duty and the relevant privacy rule to the proposed upload.
The Tax Practitioners Board’s AI guidance statement applies Code item 6 to AI. Obtain client permission before disclosing their affairs to a third party, unless a legal duty requires disclosure.
Explain the proposed disclosure, who receives it and where information will be stored. The TPB also expects you to consider how the tool is configured and used.
For practices covered by the Australian Privacy Principles, assess permitted use and disclosure under APP 6 and overseas disclosure under APP 8. APP 11 requires reasonable steps to protect personal information.
The OAIC’s AI privacy guidance recommends keeping personal information out of publicly available generative AI tools. Its advice also covers staff training and review of output accuracy.
The Privacy (Tax File Number) Rule 2015 restricts TFN use and disclosure to authorised purposes. Client consent alone does not authorise any use of a TFN.
For that reason, draw the practice’s first boundary at TFNs. Remove them before general AI review, including from attachments and scanned pages. Tax Genie removes every TFN before its AI review.
APES 110 requires members to protect confidential information throughout its handling. Disclosure needs an applicable legal or professional basis, or client authorisation that the law permits.
De-identifying a document means removing details that identify someone directly or through the remaining context. Remove names, contact details, identifiers and identifying file metadata, then assess whether the remaining facts still identify the client.
Insufficient redaction leaves text beneath a black shape that someone can copy. Permanent redaction removes the underlying content. Inspect the exported copy and its searchable text before sending it anywhere.
AI Security Risks for a Small Practice
AI can help with practice work when staff use approved accounts and restricted information. Match each risk to a specific staff rule.
| Risk | Practice rule |
|---|---|
| Client information enters a personal account | Give staff practice-managed accounts and prohibit personal accounts for identifiable client work |
| Staff use an unapproved chatbot, extension or recorder | Name approved tools and plans. Require approval before connecting another service |
| AI invents an answer or a source | Have a qualified person check the source and conclusion before advice reaches a client |
| A prompt or document reaches an unintended recipient | Remove client identifiers. Restrict shared links and connectors, then limit access to client folders |
| An uploaded document contains instructions aimed at the AI | Treat document instructions as untrusted. Require staff approval before an AI sends data or changes records |
A privacy rule must cover browser extensions and connected drives as well as chat windows. An account with training excluded can still disclose information through a shared conversation or another connected service.
Choose Your Next AI Privacy Guide
Choose the provider’s page in the directory for its account terms and settings. Use the practice policy to record the approved plan, permitted inputs and who checks AI output.
If information must stay within infrastructure your practice manages, use the private AI guide. Local processing still needs protected devices and access limited to the right staff.
For documents that contain client identifiers, start with de-identification. Remove the identifying content before it reaches the AI service, then keep the cleaned copy separate from the original.
Explore this topic
11 guides
Start here
- Is ChatGPT Safe? Data Privacy for Accountants
Is ChatGPT confidential enough for client tax records? What OpenAI keeps on each plan, the settings to change and where TFNs must never go.
- Is Adobe Acrobat AI Assistant Safe? Client PDFs
Acrobat now offers to summarise every client PDF. Learn what its AI Assistant sends to Adobe, whether it trains on files and how to switch it off.
- Is Claude Safe? Claude Privacy for Accountants
Does Claude train on your data or keep client files? Check Anthropic’s plan terms, the training setting and a safe rule for practice staff.
More guides
- Is Copilot Safe? Copilot Privacy for Accountants
Is Copilot safe for confidential client files? See what enterprise data protection covers, what Copilot can see in Microsoft 365 and what to lock down.
- Is Otter AI Safe? Recording Client Meetings
An Otter bot joined your client meeting. Learn what Otter keeps, the consent rules for recording in each state and how to delete an Otter account.
- Is Gemini Safe? Gemini Privacy for Accountants
Using Google Workspace? Learn what Gemini keeps, whether people review your prompts and which settings protect client files in Gmail and Drive.
- Is Windows Recall Safe? Screenshots of Client Data
New Copilot+ laptops can snapshot every screen, including client portals. See what Windows Recall stores, how to switch it off and when it matters.
- Is Grammarly Safe? What It Reads in Client Emails
Grammarly reads what you type in client emails and portals. See what it keeps, whether it trains on text and how to remove it from staff devices.
- Is Apple Intelligence Safe? Privacy on Practice Macs
Run the practice on Macs? See what Apple Intelligence processes on the device, what goes to Private Cloud Compute and when ChatGPT gets client text.
- Is DeepSeek Safe? Privacy and Government Bans
Why did Australia ban DeepSeek on government devices? What its privacy policy says about storage in China and whether staff may use it on client work.
- Is Perplexity Safe? Privacy for Tax Research
Researching tax questions with Perplexity? Learn what it keeps from your searches and uploads, and how to ask without client details.