Australia's Privacy Act 1988 has no clause about AI agents, yet it shapes how staff who use them must be trained. Organisations must collect, use, disclose and secure personal information properly, and taking reasonable steps includes making sure people know how. An agent adds new ways to get it wrong, so staff need to know where.
What does the Privacy Act ask, in plain words?
The Act sets out privacy principles that govern the life of personal information. In substance: collect only what you reasonably need, by fair means, and tell people why. Use and disclose it for the purpose you collected it for, or a closely related one they would expect. Keep it accurate. Protect it from misuse, loss and unauthorised access. Let people see and correct what you hold. Take care before sending it overseas, because you remain accountable for what happens to it there. Certain data breaches must be notified to the OAIC and to the people affected.
The Act has also been amended to add transparency about some automated decisions that use personal information and significantly affect people. Check with counsel how and when that applies to your agents.
None of this mentions training directly. But "reasonable steps" to protect information has always included the people who handle it. A privacy policy no one in operations understands is not a reasonable step, which is why AI literacy and certification for people who work with agents belongs in the privacy programme rather than beside it.
Who does it apply to?
The Act covers Australian Government agencies and most private-sector organisations above a small-business turnover threshold, plus some smaller ones regardless of size, such as health service providers. Inside a covered organisation, it reaches anyone whose work touches personal information. With agents, that list widens:
- staff who instruct an agent or review its drafts, such as a customer service lead approving a reply that quotes an account history;
- process owners who decide which sources an agent may read;
- privacy officers who handle access and correction requests that now include agent output;
- anyone who reports or triages incidents, because an agent can disclose in seconds what a person would take an afternoon to leak.
For how the Act applies to the agents themselves, see Australia's Privacy Act and AI agents.
What does it mean when the system is an agent that acts?
An agent that only reads (depth 1) still uses personal information, so the purpose question applies from the first day. At depth 2 it drafts replies, letters or notes that a person sends, which makes the reviewer the last check before disclosure. At depth 3 it executes routine work within bounds, so the bounds themselves carry the privacy rules and someone must have set them correctly.
Four skills follow for the people involved:
- Spotting purpose drift. A reviewer should notice when a draft uses information collected for one purpose to serve another, such as pulling a hardship note into a marketing reply.
- Reading provenance. Facts reaching a governed agent arrive stamped as sealed, signed, declared or external. A reviewer who sees an external claim, such as an emailed request to change a contact address, should know it is the weakest class and treat it accordingly.
- Handling a refusal. When the agent refuses because a request falls outside its bounds, the right response is to route it, not to rephrase the request until it passes.
- Recognising a breach. Misdirected agent output is a potential data breach. Staff must know the clock starts when it is discovered and who to tell.
How do you meet the Act's expectations through training?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Use and disclose for the collected purpose | Reviewers trained to check purpose before approving any held draft that quotes personal details | Review decisions in the agent's record, with decline reasons citing purpose |
| Protect personal information | Agent sources limited by role scope; staff trained not to paste extra data into prompts | Source configuration history; training completions dated before access was granted |
| Keep information accurate | Staff trained to correct the source, not just the draft | Correction entries in the system of record linked to the review |
| Notify eligible breaches | Incident drill covering agent misdisclosure | Drill records, escalation times, the incident log |
| Respond to access and correction requests | Privacy team trained to retrieve what an agent read and produced about a person | Request files showing the agent's record was searched |
Two points make this work. Train before access, not after: the order of the dates matters. And make declines useful, because a reviewer's reason for declining a draft is evidence that the control operated. If your team also answers requests from California residents, the training overlaps with the approach in CCPA and training teams that answer privacy requests with AI agents, though the rules differ.
What will the OAIC or an auditor ask?
After a complaint or a breach, the questions are practical. What personal information can the agent reach, and why does it need it? Who reviewed the output that went out, and were they trained to check it? How was the breach found, and how long did escalation take? Can you show the individual everything the agent produced about them?
An auditor will add a sampling question: pick ten approved drafts, then show the approver's training record on each approval date. If the record has gaps, the control has gaps. A wider view of what Australian organisations need is in what AI agent training organisations in Australia need.
What does the Privacy Act not cover?
It does not regulate how accurate or fair an agent's reasoning is, except where that affects the accuracy of personal information. It does not set operational resilience rules; APRA-regulated entities answer to CPS 230 for that, and conduct towards consumers sits with ASIC. It does not set a training syllabus or require certification. The government's voluntary AI safety standard is guidance, not law. And the Act does not stop at your walls: if a provider processes the data, your obligations follow it.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Pryme Intelligence Academy sets every module test in the learner's own industry and role, so a privacy officer and a contact-centre reviewer each practise on cases that look like their own work. For teams, an admin view shows progress, results, certificates and expiry dates, which gives you the dated training record that sits beside the agent's record when someone asks who approved what.
Questions people ask
Does an AI agent count as a disclosure of personal information?
Sending agent output to someone outside the organisation is a disclosure in the ordinary sense, and the usual rules apply. Passing data to a provider that runs the agent may also raise questions about use, disclosure and overseas transfer, which is why the provider arrangement belongs in your privacy assessment.
Do small businesses need to train staff under the Act?
Many small businesses fall outside the Act, but not all, and those handling health information usually fall inside it. Even where the Act does not apply, training people to review agent output protects customers and the business.
Should training cover the agent's refusals?
Yes. A refusal often means the agent was asked to use information outside its bounds. Staff who treat refusals as obstacles to rephrase around defeat the control the bounds were built to provide.