It depends: Claude is suitable for client work only when your brokerage approves the plan’s data terms and limits what Claude can access. Keep identifiable client files out of a personal Claude account. A paid individual subscription and incognito mode do not establish approval to upload a client’s financial records.
The Short Answer
Use Claude for generic drafting and genuinely de-identified questions unless your brokerage has approved a controlled client-data workflow. Claude Team and Enterprise have different training terms from consumer plans, but the brokerage still needs to assess retention and overseas handling.
The Office of the Australian Information Commissioner (OAIC) recommends that organisations avoid entering personal information into publicly available generative AI tools. Its guidance on commercially available AI applies that advice particularly to sensitive information.
In a hypothetical broker task, ask Claude to simplify a generic document-request email using fictional details. Uploading a real client’s payslips and bank statements introduces a different risk. Removing the name alone leaves transaction details and other clues that can identify the client.
Does Claude Train on Your Data?
Claude’s consumer training use depends on your model-improvement setting and the policy’s exceptions. Free, Pro and Max are consumer plans. Paying for Pro does not give the account Team or Enterprise data terms.
Anthropic changed its consumer training arrangements in 2025. Its consumer privacy policy permits training on inputs and outputs unless you opt out. Feedback submissions and conversations flagged for safety review have separate exceptions.
To opt out on a browser or desktop account, follow Anthropic’s model-improvement settings instructions:
- Open your account menu and select Settings.
- Select Privacy.
- Turn off Help Improve our AI models.
- Return to Privacy and confirm the setting remains off before starting work.
With model improvement enabled, Anthropic can retain eligible new or resumed chats in de-identified training pipelines for up to five years. Turning it off stops use of previous and new chats in future training runs. It cannot reverse training already started or completed. The consumer retention policy separates this training retention from deleting saved conversations.
Claude for business teams uses commercial terms. Anthropic says Team and Enterprise chats are not used for training by default. Explicit feedback and training opt-ins can change that treatment. Keep client information out of feedback reports too.
Incognito Chats, Deletion and Retention
Claude incognito chats stay out of your chat history and memory, and Anthropic says they are not used for model improvement. They are still retained for 30 days by default. Enterprise retention settings can require longer retention.
Start a new chat outside a project and select the ghost icon. Confirm the Incognito chat label appears before entering information. Anthropic’s incognito instructions also explain that these chats remain available in Team and Enterprise organisational exports. Incognito does not make a conversation invisible to the organisation that manages the account.
Deleting a regular consumer chat removes it from history immediately and schedules backend deletion within 30 days. Saved commercial chats also remain in the product until deleted or removed under the applicable retention arrangements. Anthropic’s commercial retention policy explains the difference from API retention. Safety investigations and legal obligations can extend retention.
Zero data retention is a separate approved arrangement for eligible API uses. It does not follow automatically from an Enterprise chat subscription or incognito mode. Anthropic’s zero data retention coverage excludes some model and feature combinations and preserves safety-classification records. A brokerage buying an AI service through another supplier needs that supplier’s retention terms as well.
Connectors, Cowork and Gmail Access
Connectors let Claude retrieve information from accounts you authorise, so their scope can exceed the document you intended to discuss. For a client mailbox, leave Gmail and broad document access disconnected until the brokerage approves the specific connection.
Anthropic’s Google Workspace connector guide describes these permissions:
- Gmail can search and read messages, draft replies and send emails. Attachment metadata is available, while attachment content is not directly accessible through Gmail.
- Google Drive can retrieve accessible files and perform actions such as sharing, moving or trashing them. The connected Google account’s existing access limits apply.
- Retrieved information is stored with the chat. Anthropic excludes raw connector data from training, but copied content and responses containing that information can fall under consumer training settings.
Sending and certain file actions require approval by default. Team and Enterprise owners can allow some actions without asking each time. Keep those approvals enabled for client work.
Cowork can read and change files in connected desktop folders. Anthropic’s Cowork safety guide explains that cloud sessions process opened local files on its servers. Desktop access does not keep the processing on your computer.
Use a dedicated folder containing only approved working material. Keep the client-file archive and signed-in client systems outside its access. Documents and emails can contain malicious instructions that try to redirect Claude, a risk called prompt injection. Manual approvals and reviewing outputs reduce exposure, but neither makes broad client-mailbox access safe by itself.
Where Claude Stores Data
Standard Claude commercial service data is stored in the US, with processing able to occur in several regions. Anthropic’s server-location guidance distinguishes traffic routing from storage. Australian routing alone does not establish Australian storage.
Anthropic also lists Australia among regional deployment options through cloud platforms. Those deployments have their own regional endpoints and configuration. They do not establish that an ordinary Australian Claude web account keeps its files in Australia. Treat a contractual storage location and a contractual processing location as separate requirements.
For a brokerage covered by the Privacy Act, Australian Privacy Principle 8 (APP 8) can apply when personal information is disclosed overseas. The OAIC’s APP 8 guidance explains that giving an overseas contractor personal information is usually a disclosure. Limited arrangements that preserve effective control can be a use instead.
Before approving client files, record who can access them overseas and which contract controls their handling. A no-training setting answers only one part of that assessment.
Claude or ChatGPT for Client Work
Choose an approved business workspace on its contractual protections and enabled features, rather than assuming either brand is safe for every client file. Both providers separate consumer training controls from business workspace defaults.
Privacy Settings Compared
| Control | Claude | ChatGPT |
|---|---|---|
| Individual-account training control | Settings, then Privacy, then Help Improve our AI models | Settings, then Data controls, then Improve the model for everyone |
| Business training default | Team and Enterprise content is excluded by default | Business and Enterprise content is excluded by default |
| Temporary conversation | Incognito excludes history, memory and model improvement | Temporary chat excludes history and model improvement while it remains temporary |
| Temporary retention | 30 days by default, potentially longer under Enterprise policy | A copy can remain for up to 30 days for safety |
| Memory in temporary mode | Incognito does not use existing memory | Personalised temporary chat can use existing memory. Choose Unpersonalized to disable that use |
OpenAI’s data controls instructions explain the training switch and business defaults. Its temporary chat guidance distinguishes temporary conversations from ones you save. Saving a temporary ChatGPT chat makes the normal account settings apply.
To move useful work from ChatGPT to Claude without carrying client data across, rebuild the reusable instructions instead of importing a complete chat export:
- Identify generic prompts and templates you want to retain.
- Create a clean local copy. Remove client facts from examples, quoted messages and attached documents.
- Replace real scenarios with fictional ones, including unusual circumstances that could identify a person.
- Put the clean instructions into the approved Claude workspace. Keep the original export out of Claude.
- Run a fictional example and inspect both the instructions and response for client identifiers before sharing the template with staff.
A memory transfer is another transfer of information. Review any proposed memory summary before pasting it into Claude. Keep client-case records in the brokerage’s approved record system, and transfer only the reusable material the new assistant needs.
