Australia's Privacy Act 1988 treats an AI agent as part of the organisation that runs it. Whatever personal information the agent reads, infers, writes or sends overseas is handled by that organisation, so the privacy principles on collection, use, quality, security, cross-border disclosure and access apply to each step the agent takes, and data breaches it causes may be notifiable.
What does the Privacy Act require, in plain words?
The Privacy Act 1988 sets out privacy principles that govern how covered organisations handle personal information. The OAIC regulates it. In substance, an organisation must:
- manage personal information openly, with a clear privacy policy;
- collect only what is reasonably necessary for its functions, by lawful and fair means, and tell people about the collection;
- use and disclose information for the purpose it was collected, or a related purpose the person would reasonably expect, unless an exception applies;
- meet higher standards for sensitive information, such as health information, usually with consent;
- take reasonable steps to keep information accurate, up to date and complete;
- take reasonable steps to protect it from misuse, loss and unauthorised access;
- remain accountable for information it sends overseas;
- give people access to their information and correct it on request.
A notifiable data breach scheme requires notice to the OAIC and affected individuals when a breach is likely to cause serious harm. Amendments to the Act have also added transparency expectations about substantially automated decisions that significantly affect individuals; check when those apply to you.
Who does it apply to?
Australian Government agencies and private sector organisations above a turnover threshold, along with certain smaller businesses, including health service providers and those that trade in personal information. Most banks, insurers, health providers and larger retailers are covered.
Coverage follows the organisation, not the technology. An agent run by a covered organisation is covered. The healthcare and e-commerce sector briefs for Australia, on compliance teams in healthcare and compliance teams in e-commerce, show how this plays out in practice.
What changes when the system handling the information is an agent?
Agents create privacy events that a database never did. Four stand out.
Inference is collection. When an agent works out something new about a person, such as that a caller is likely in financial hardship or that a patient's symptoms suggest a condition, it has generated personal information. The OAIC has treated generating personal information this way as a form of collection. It must be reasonably necessary, and if it is health information it is sensitive.
Purpose creep happens at tool level. An agent handling a delivery query that can also read a customer's full purchase and payment history will, sooner or later, use it. Use limitation needs enforcing in the agent's permissions, not only in policy.
Accuracy is now an output problem. An agent that writes a wrong summary into a customer record has created inaccurate personal information. The duty to keep information accurate covers what agents write.
Cross-border disclosure is often invisible. If the model behind an agent runs outside Australia, sending personal information to it may be a cross-border disclosure for which the organisation stays accountable.
How do you meet the Act when agents handle personal information?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Collect only what is necessary | Agent sources limited per capability; inference capabilities approved explicitly | Capability register listing sources and permitted inferences |
| Use for the collected purpose | Agent tools scoped to the task; out-of-scope retrieval refused | Refusal records naming scope as the deciding bound |
| Sensitive information | Health and other sensitive data accessible only to capabilities with a recorded basis | Access approvals and consent records |
| Accuracy | Agent-written records marked as such and reviewed where they affect decisions | Edit history of agent-written entries |
| Security | Agent credentials held in a secrets store; least-privilege access | Access logs and credential reviews |
| Cross-border accountability | Model and hosting locations known; contracts bind overseas recipients | Provider data-location records and contract terms |
| Access and correction | Agent-generated notes included in access requests | Access request responses showing agent content |
| Breach notification | Agent incidents assessed under the notifiable breach scheme | Breach assessments and notifications |
What will the OAIC or an auditor ask?
- What personal information does each agent read, and why does it need it?
- Does any agent infer sensitive information, such as health or financial hardship?
- Where is the model run, and what personal information goes there?
- If a customer asks for their records, do you include what agents wrote about them?
- What stops a customer service agent pulling information it does not need?
- Has an agent ever disclosed information to the wrong person, and what did you do?
What does the Privacy Act not cover?
It does not regulate AI as such, and Australia has no AI-specific statute. The government's voluntary AI safety standard offers guidance but no binding duties. Small businesses below the threshold are mostly outside the Act unless an exception brings them in. It does not address conduct issues such as misleading statements, which sit with consumer and financial services law. And it does not classify systems by risk tier the way the EU does; for that contrast, see EU AI Act risk tiers explained for AI agents. For the wider control map, see AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
In Pryme each capability has its own sources, so an agent answering delivery questions cannot read payment history it was not granted. Every fact reaching the agent carries its provenance, and each action is filed, which helps answer access requests and breach assessments with a record rather than a reconstruction.
Questions people ask
Is putting customer data into an AI agent a disclosure?
If the agent runs inside your organisation's control, it is usually a use. If the data goes to a third-party model provider, it may be a disclosure, and if that provider is overseas the cross-border principle applies.
Do agent-generated notes count as personal information?
Yes, if they are about an identified or reasonably identifiable person. They are subject to accuracy, security and access duties like any other record.
Does the Act require telling customers an AI agent is used?
The Act requires clear privacy policies and collection notices, which should describe how agents handle information. Amendments on automated decisions add specific transparency expectations where decisions significantly affect people.