AI privacy for online stores: is your data safe?

Learn which AI privacy risks matter for an online store, what AI tools keep and train on, and how to use them without exposing customer data.

AI privacy depends on what your team sends, which account they use and what the provider does with it. Keep customer details out of consumer assistants. Use an approved business service only after its terms and controls meet your store’s obligations.

The tool pages below explain the settings for each assistant. For the legal duties, start with the Privacy Act in Australia and the OAIC AI guidance. If staff use accounts your store hasn’t approved, address shadow AI alongside the data settings.

Is AI safe to use with customer data?

Yes, AI can be safe for store work when personal information stays out, or an approved service protects its use. Pasting identifiable customer records into a consumer account creates avoidable exposure. Paying for an individual subscription doesn’t establish permission to upload a customer list.

The Office of the Australian Information Commissioner (OAIC) recommends keeping personal information, especially sensitive information, out of publicly available generative AI tools. That is the starting rule for staff drafting replies or rewriting product copy.

The main AI privacy concerns affect different parts of a conversation:

  • Saved chats and uploaded files remain with the provider under its retention rules.
  • Training can use submitted text to improve later models when the account’s terms or settings allow it.
  • Safety checks can involve authorised human reviewers, even when training is off.
  • Overseas processing puts information under another country’s laws and adds cross-border privacy questions.
  • Shared links, compromised accounts and software faults can expose information beyond the intended readers.

In a hypothetical returns query, ask for a reply about a delayed refund using the policy and an anonymous summary. Leave out the customer’s name, email, order number and payment details. Removing a name alone is insufficient if the remaining facts identify the customer.

What AI tools keep and train on

Chat history, model training, safety review and API logs are separate uses of data. Chat history lets you return to a conversation. Training changes a model using examples. Safety review checks for misuse or harmful content.

An application programming interface (API) lets another application send requests to a model. Its provider can keep logs under terms that differ from the consumer chat app. OpenAI’s API data controls distinguish abuse-monitoring logs from stored application data. A no-training promise alone says nothing about whether those records exist.

AssistantConsumer accountBusiness distinction
ChatGPTTraining can be switched off while chats remain in history. Temporary Chat has separate safety retention.OpenAI says business data isn’t used for training by default. Retention controls depend on the service.
ClaudeConsumer training depends on permission. Anthropic says allowing model improvement extends the retention period for eligible conversations.Commercial terms differ from consumer terms. A staff member’s personal subscription remains a consumer account.
GeminiGoogle retains some chats for safety even when Keep Activity is off. Human review can still occur.Workspace use has separate protections. A personal Google account doesn’t inherit them.
CopilotConsumer Copilot has its own training and privacy settings.Microsoft says prompts and responses in protected work or school Copilot Chat aren’t used to train foundation models.

These distinctions follow OpenAI’s business privacy commitments, Anthropic’s consumer data policy, the Gemini Apps Privacy Hub and Microsoft’s work-account data protection.

AI data privacy rules for Australian stores

The Privacy Act applies to covered stores when AI handles personal information. Some small businesses are exempt, while exceptions bring others within the Act. The OAIC’s small business guidance explains that coverage.

For covered stores, the Australian Privacy Principles (APPs) still apply when an assistant does the work. APP 6 limits use and disclosure beyond the purpose for which information was collected, unless an exception applies. Using an email to resolve a customer’s complaint differs from allowing a provider to train on it.

APP 1 and APP 5 govern transparency through privacy policies and collection notices. APP 11 requires reasonable security steps. The OAIC’s guidance on commercial AI applies these duties to AI selection and use.

Under APP 8 on overseas disclosure, covered stores generally need reasonable steps to protect information disclosed to an overseas recipient. They can remain accountable for that recipient’s handling, subject to exceptions. An Australian billing address doesn’t establish where processing occurs.

AI data breaches and leaks

AI data breaches include service faults that expose another user’s information and public sharing that sends a private conversation beyond its intended audience. OpenAI’s account of a ChatGPT incident describes exposed chat titles and payment-related information affecting some subscribers. Shared Grok conversations also appeared in search results, as TechCrunch reported. A public share link creates a different exposure from a provider’s internal security breach.

If customer information leaks, stop further sharing and restrict the affected account or integration. Preserve the incident evidence, identify the information involved and assess who can access it. Ask the provider to contain access and explain which copies remain.

For stores covered by the Notifiable Data Breaches scheme, suspicion of an eligible breach requires a prompt assessment. Take all reasonable steps to complete it within 30 calendar days. When serious harm is likely and remedial action hasn’t prevented it, notify the OAIC and individuals at risk as soon as practicable, unless an exception applies. The OAIC breach response guide explains the threshold and response. Deleting a chat alone doesn’t establish that the exposure is contained.

Choosing the safer option

Choose the service by its data terms and your ability to enforce them across staff accounts.

OptionWhat to establish before customer data enters it
Consumer accountKeep identifiable customer information out. A training opt-out doesn’t remove storage or safety review.
Business planEstablish the permitted purpose, training terms, retention and administrator access. ChatGPT Business is one example of a separate business service.
Private AIEstablish who operates it and who receives prompts. The term private AI can describe different hosting and access arrangements.
Local modelA local LLM, or large language model, can process prompts on equipment you control. Secure its chat files and backups, and account for any cloud connections.

Give staff a named approved account and a rule for which information it can receive. Use anonymised material when the task doesn’t need identifiable records. Restrict connected files to the information required for the job.

Is each AI tool safe?

Use the guide for the assistant and account your team actually uses. The same brand can have consumer, business and API terms with different protections.

  • Is ChatGPT safe? Consumer training controls reduce one use of data. Business terms and retention still need separate assessment.
  • Is Claude safe? Consumer training permission changes data use. Confidential store files need an approved account and suitable terms.
  • Is Gemini safe? A personal account can retain chats for safety even with activity saving off. Workspace has separate protections.
  • Is Copilot secure? Protected work-account use differs from consumer Copilot. File permissions still decide which store documents staff can reach.
  • Is DeepSeek safe? Its app policy says personal data is processed and stored in China. Keep customer records out of the app.
  • Is Meta AI safe? Sending material to an AI feature changes who receives it. Keep customer details out of those interactions.
  • Is Perplexity safe? Its data-use control governs model improvement. Turning it off doesn’t by itself approve uploading confidential sales files.
  • Is Grok safe? Training controls and public sharing are separate. A shared conversation can become accessible beyond the person you send it to.
  • Is Grammarly safe? It processes text where it is active. Approve its account controls before staff use it in customer replies.

AI privacy and a store’s own shopping agent

A shopping agent on your site needs the same assessment as an assistant your staff use. Customers can put personal information into a conversation even when your catalogue contains only public product data.

shoppa is an AI shopping agent that helps customers find, compare and buy products in the chat using your catalogue and store documents. It holds shopper conversations, so your assessment must cover those conversations as well as the product feed.

Before connecting any shopping agent to customer information, get written answers to these questions:

  • Where are conversations stored and processed, including backups?
  • Which retailer staff and provider personnel can read them?
  • How long does each party keep chats, logs and uploaded files, and how does deletion work?
  • Which model provider receives the conversation, and what are its training and retention terms?
  • Which customer information do connected order or support systems expose to the agent?

Name the AI clearly to shoppers and explain its handling of personal information in the relevant notices. Give the agent access to the minimum customer information its job requires.

Explore this topic

12 guides

More guides