Is customer data safe when your business uses AI?
Customer data is not automatically safe or unsafe because a product uses AI. The answer depends on the information you share, why you share it, where it goes, who can access it and the controls you actually configure.
For an Australian business, start with a specific task and the smallest amount of information needed. A tool answering opening-hours questions has different requirements from one handling patient records or participant information.
Map one task from beginning to end
Before uploading records or connecting an inbox, complete this worksheet for the proposed use:
- Input: what a staff member or customer will provide.
- Purpose: the task that requires each field.
- Route: the website, AI supplier, connected systems and support services it passes through.
- Storage: where inputs, outputs, logs and backups remain.
- Access: the staff and suppliers who can see each copy.
- End of use: when information is removed and how deletion is verified.
If the supplier cannot explain a stage, mark it unresolved. “Hosted in Australia” answers a location question; it does not answer every question about support access, subprocessors or copies sent elsewhere.
Ask for answers about your exact product and plan
Consumer chat tools, business subscriptions and API products can have different terms. Ask these questions in writing:
- Are our inputs or outputs used to train models, and is that configurable?
- What is retained by default, and what can we change?
- Which subprocessors receive data?
- Can access be limited by role and protected with multi-factor authentication?
- Can we export and delete our data, including after cancellation?
- How are incidents reported to us?
- What happens when the supplier changes its terms or processing arrangements?
The OAIC’s guidance addresses organisations’ privacy obligations when using commercial AI. It recommends, as best practice, avoiding personal and particularly sensitive information in publicly available AI chatbots. It also explains transparency and collection considerations for customer-facing systems. The guidance does not certify any individual product as safe. Read the OAIC guidance.
Remove information the task does not need
Consider an illustrative workflow that groups enquiries by suburb and requested service. It may not need a person’s full address, payment details or attachments to perform that classification. Keep original records in the system that needs them, and send only the fields required for the proposed action.
Replacing a name with a reference number can reduce exposure, but do not assume it makes the remaining content anonymous. A detailed story, address or combination of attributes may still identify someone.
Test access and deletion, not just the answer quality
Use synthetic records during evaluation. Try a staff account with limited permissions. Confirm that one customer cannot obtain another customer’s details through the assistant. Check that someone leaving the business loses access to the AI tool and connected accounts.
Create a test record, follow where it is stored and request its deletion. Ask what remains in logs or backups and for how long. Record the answer rather than assuming the visible delete button removes every copy.
The Australian Cyber Security Centre’s small-business AI guidance is a useful companion for assessing cloud AI security and staff preparation. ACSC guidance.
Give staff a short operating rule
“Use only approved AI tools for approved tasks. Share only the information the task needs. Check outputs before they affect customers. If you enter information into the wrong tool or see an unexpected disclosure, stop and tell [responsible person] using [reporting channel].”
Adapt this to your organisation’s actual policy. Staff need a named contact and a practical response process, rather than an instruction to “be careful”.
Make a documented decision
For each proposed use, record the owner, permitted information, safeguards and unresolved issues. Pause uses involving sensitive information until the relevant privacy or professional reviewer has assessed them. Legal duties vary by organisation and activity; this worksheet supports procurement and does not establish compliance.
RightLink’s AI strategy service can help map a proposed workflow and identify questions to resolve before connecting customer systems.