Privacy Act Reform and What It Means When Your Business Uses AI
Australian privacy law is being rewritten in stages, and the parts already through Parliament land squarely on businesses that have started using AI. A statutory tort now lets a person sue over a serious invasion of privacy. Privacy policies will have to set out where automated decisions that significantly affect people are made. The security and destruction expectations have been tightened, children are getting their own code, and a broader fair and reasonable handling test is still on the table. None of that stops you using AI. It changes what you have to be able to show.
This is a working guide written for owners and managers rather than lawyers: what has passed, what is still proposed, and the practical changes worth making in your business this quarter. It is general information and not legal advice, so take anything that matters to your own solicitor before you act on it.
Realistic ROI
Why Work Through This With Yes AI
Privacy reform is being treated in most businesses as a document problem, which is why so many privacy policies now describe a company that does not exist. The obligations that are actually landing are about what happens in the systems: where the data goes, who can reach it, how long it stays, and who is accountable when a decision affects someone. Four reasons an Australian AI consultancy is the right partner for that part of the work, sitting alongside your solicitor rather than in place of them.
We start with what your business actually does
Privacy work goes wrong when it starts with a template. We start with the walk around: who collects what, which tool someone reaches for when they are behind, what the website chatbot stores, where the call recordings sit, which spreadsheet quietly holds the entire customer list. That produces an inventory of every point where personal information meets an AI system, with a named owner against each one. Everything afterwards, the policy wording, the retention rules, the access controls, is writing down what is true and fixing what is not.
We know where the data really moves, because we build the plumbing
Most privacy advice stops at the edge of the systems. We build the integrations between the systems, so we can tell you where a customer record actually goes once it leaves your CRM, what a vendor retains, which copies exist in a search index or a log file, and what a deletion request would genuinely have to reach. That matters now, because the tightened security and destruction expectations are about what happens in the plumbing, not what the policy page promises.
Australian consultancy, Australian obligations
We are based in Melbourne and we work to Australian rules: the Privacy Act and the Australian Privacy Principles, the state by state position on recording a phone call, the Notifiable Data Breaches scheme, and the voluntary AI safety guardrails that expect much the same discipline. We are not translating a European or American compliance product and hoping it fits. When you ask where your data sits and who can reach it, you get a straight answer about your setup, not a generic one.
We change the system, not just the paperwork
A privacy policy that describes an ideal version of your business is worse than no policy, because the gap is now written down in your own words. So we do the build as well as the review: retention actually set on the transcript store, the approved AI tool configured so it does not train on your content, the chatbot disclosing itself, a human review step wired into the workflow instead of printed on a poster. The team that maps the problem is the team that closes it, and stays on to keep it closed.
What the Reforms Change for a Business Using AI
Six moving parts, some already passed and some still proposed, and each one touches AI use somewhere. Read the badge on each card as its legal status. Where something has a transition period or is still a policy commitment rather than law, we say so, because getting that distinction wrong is how businesses either panic early or get caught late.
A person can now sue you directly
The statutory tort for serious invasions of privacy came in through the 2024 amendments and gives an individual a direct cause of action, covering both intrusion into someone's private affairs and misuse of information about them. Two things matter for an AI project. A court can award damages without the person having to prove financial loss, so "no real harm done" is not the defence people assume it is, although the invasion still has to be serious and the conduct intentional or reckless rather than merely careless. And it sits alongside the Australian Privacy Principles rather than inside them, so a small business that has always relied on the turnover exemption should stop treating itself as untouchable. The trigger in an AI context is almost always the same: personal information ended up somewhere nobody expected it to be.
Your privacy policy has to name automated decisions
The reforms add a requirement to set out in your privacy policy the kinds of personal information used in computer programs that make, or substantially help make, decisions that significantly affect an individual's rights or interests, and the kinds of decisions involved. There is a transition period before it bites, so confirm the current timing rather than assuming it applies today. The hard part is not the wording, it is the honesty test behind it. Someone has to walk the business and answer whether an AI scoring, ranking, screening or prioritising step genuinely influences an outcome for a customer, an applicant, a tenant or a patient. Most businesses find at least one they had not thought of, and it is often sitting in recruitment or in how enquiries get triaged.
Children are getting their own code
The reforms require a Children's Online Privacy Code covering services likely to be accessed by children, developed by the regulator rather than written line by line in the legislation. If you run anything a person under eighteen might use, a tutoring service, a sports club portal, a paediatric clinic booking form, a retailer with a teenage customer base, assume the expectations around consent, default settings, profiling and targeting will end up stricter than your current practice. The AI angle is age unknown data. If your chatbot, recommender or marketing model cannot tell whether it is dealing with a child, the safe design is the one you would be comfortable with if it were.
Security and destruction expectations tightened
APP 11 has been clarified so that the reasonable steps you take to protect personal information expressly include technical and organisational measures, which reads as an invitation to show your working. It pairs badly with the way AI creeps into a business: a customer list pasted into a chat window, a document store indexed for search, transcripts of every call sitting in a folder with no end date, an export on somebody's laptop. The destruction limb is the one everyone forgets. If information no longer serves the purpose it was collected for, it should not still be sitting in a prompt log, a search index, a summarisation output or an eighteen month backup because nobody ever set a rule.
Enforcement got cheaper for the regulator
The amendments added lower tiers of civil penalty and infringement notice powers, so the regulator no longer has to run a full Federal Court case to act on the smaller things, including a privacy policy that does not contain what it is required to contain. Courts were also given broader scope on remedies. In plain terms, the cost of a sloppy policy page and a vague breach response has gone down for them and up for you. Separately, doxxing was made a criminal offence, which is worth knowing if your business handles disputes, debt recovery, or any situation where a frustrated staff member might be tempted to publish someone's details.
Fair and reasonable is the direction of travel
A broader test, requiring that collection, use and disclosure be fair and reasonable in the circumstances whether or not the person consented, was agreed in principle in the response to the Privacy Act Review but has not been legislated. Nor has the removal of the small business turnover exemption, though it sits on the same list. Treat both as planning assumptions rather than current obligations. The reason to design for them now is that a fair and reasonable test would be judged on what an ordinary person would expect, and "they clicked accept on a privacy policy nobody reads" has been a weak answer for years. Build to the standard you would be comfortable explaining out loud to the customer it concerns.
Six Situations Australian Businesses Are Already In
| Task | Traditional | With Yes AI | Notes |
|---|---|---|---|
| Staff pasting customer emails into a free AI chatbot to draft replies | Nobody knows what went in, and the vendor may retain and train on it | An approved business tier tool with training off, or an internal path that is easier than the shortcut | The record of what left the business is the one thing you cannot reconstruct afterwards, so the control has to exist before the incident, not after it. |
| AI ranking or scoring job applicants, tenants, patients or credit risk | The score quietly becomes the decision and no document mentions it exists | A named human decision maker, a written reason, and the privacy policy updated to match | The reforms ask you to disclose significant automated decisions. Keeping a person accountable for the outcome is also how you survive being asked to explain one. |
| Call recordings and transcripts fed into an AI summariser | Recordings kept forever in a shared folder with no retention rule at all | Consent handled at the start of the call, retention set, transcripts destroyed on schedule | Recording law is state by state and sits separately from the Privacy Act, so both have to be right. The transcript is still personal information after the audio is gone. |
| A website chatbot collecting names, numbers and problem detail | Conversation logs pile up in a vendor account nobody has ever audited | Disclosed as AI, minimal collection, logs retained for a stated period and no longer | People tell a chatbot things they would never put in a form. Design for the health condition or the financial trouble that will inevitably turn up in a transcript. |
| Answering an access or correction request now AI is in the stack | A frantic search across inboxes, drives, the CRM and three AI accounts | An inventory that says where personal information lives, so the search is bounded | The inventory is not compliance theatre. It is the only reason the request can be answered in a sensible timeframe, and it is the same list a breach response needs. |
| Sending customer data to an AI service that processes it offshore | Assumed to be fine because the vendor is a well known name | Data location checked, contract read, cross border disclosure handled deliberately | APP 8 is its own obligation and the reforms add a mechanism to recognise countries with comparable protections. The default answer is not automatically yes. |
Six Things We Will Tell You Straight
This is general information, not legal advice
We are an AI consultancy, not your solicitor, and the Privacy Act is being reformed in stages with commencement timing that keeps moving. Everything on this page is written to help you ask better questions and get your house in order before someone else asks them. Where a specific obligation, a contract, a data breach or a threatened claim is involved, take it to a lawyer who can look at your actual circumstances. We work happily alongside your legal adviser and would far rather build to their position than guess at it.
AI cannot tell you whether you are compliant
It is tempting to point a language model at your privacy policy and ask whether it meets the Act. You will get something confident and plausible back, and it will miss the thing that matters, because compliance turns on what your business actually does rather than what the document claims. The useful work is the walk around the business: who collects what, where it goes, who can see it, how long it stays. AI is genuinely good at drafting, comparing and spotting gaps once you have that map. It is not a substitute for knowing.
Consumer AI tools are still the most common leak
The realistic risk in most Australian SMEs is not a hacker. It is a good employee under time pressure pasting a customer list, a resume, a clinical note or a full email thread into a free personal AI account to save twenty minutes. Free and personal tiers often reserve the right to retain and train on what you submit, and you cannot pull it back. Decide which tools are approved, switch on the business tier controls, put it in writing, and give people a sanctioned path that is easier than the shortcut. A ban with no alternative just moves the behaviour somewhere you cannot see it.
Deleting personal information from AI systems is harder than it sounds
Setting a retention rule in your CRM is the easy half. The same information is usually also sitting in a document index, an embedding store used for search, prompt and response logs, a summary someone saved, an email trail, and a backup on its own schedule. If a person asks you to delete their data, or the information stops serving the purpose it was collected for, the deletion has to reach all of it. Build the AI stack so every copy has a known home and an owner from day one, because retrofitting deletion into a system that sprawled quietly is expensive and never quite complete.
Disclosing an automated decision does not make the decision acceptable
Adding a line to the privacy policy is a transparency obligation, not a permission slip. A model that screens people out of a job, a tenancy, a loan or a service can be discriminatory, wrong, or simply unable to explain itself, and the disclosure cures none of that. Keep a human who can see the inputs, override the output and write a reason an ordinary person would accept. Where a decision meaningfully affects someone, treat AI as the thing that prepares and prioritises the work, not the thing that delivers the verdict.
De-identified is not the same as anonymous
Stripping a name off a record feels like it takes the record outside the Act, and it usually does not. Combinations of postcode, date of birth, appointment time, job title and purchase history re-identify people with uncomfortable ease, and the more context an AI system has around them, the easier it gets. If you plan to use customer data to build or tune something, get advice on whether it is genuinely de-identified before you rely on that, and be honest about whether your original collection notice covered this new purpose. Secondary use is where good intentions meet the Act head on.
How Yes AI Helps
A plain review of where AI meets personal information
We map every flow where customer, staff or patient information reaches an AI system, including the tools nobody formally approved, and rank them by how much would hurt if they went wrong. You get a short, readable inventory naming the data, the system, the purpose, where it is processed and who owns it. That single document is what your solicitor, your insurer and your own breach response all end up asking for.
Privacy policy and disclosure wording that matches reality
We redraft the sections the reforms actually touch, including how automated decisions that significantly affect people get described, plus your collection notices, chatbot disclosure, call recording script and staff AI use policy. Your lawyer receives an accurate description of what the business really does to review and sign off, rather than a template they have to reverse engineer from scratch.
Retention, deletion and access built into the tooling
Rules that live in a document get ignored, so we set them where they take effect: retention on the transcript and log stores, deletion that reaches every copy including search indexes and backups, least privilege access so a junior account cannot pull the whole customer table, and training switched off on the tools that offer it. Boring, verifiable, and the part that actually survives an incident.
Human review points and ongoing support
We design the workflow so a named person owns any decision that materially affects someone, sees enough to disagree with the model, and leaves a record. Then we stay on. The reforms are staged, your tooling will change faster than the law does, and a new AI feature can create a new obligation quietly. We review on a rhythm so this does not go stale six weeks after the workshop.
How We Run a Privacy and AI Review
Five steps from a first conversation to controls that are actually in place and monitored. Most Australian SMEs get through it in two to six weeks depending on how many systems hold personal information and how much AI has already spread through the business.
Map where personal information meets AI
We sit with the people doing the work and list every point where customer, staff or patient information reaches an AI system: the approved tools, the ones nobody approved, the website chatbot, the call transcripts, the document search, the marketing model. What comes out is a short inventory naming the data, the system, the purpose, the processing location and the owner. It is common for that walk to surface a flow nobody in the room knew was running.
Classify the data and agree the rules
Not all of it carries the same weight. Health information, information about children, financial detail and anything else sensitive gets stricter handling than a business email address. We agree what may go into which tool, what never leaves the building, how long each category is kept and how it is destroyed. Those rules then become retention settings and access controls rather than a document nobody opens twice.
Update the documents so they match the business
We redraft the privacy policy sections the reforms touch, including automated decisions that significantly affect people, alongside collection notices, the chatbot disclosure, the call recording script and your staff AI use policy. Your solicitor gets a clean, accurate account of what the business does to review, which is a much cheaper hour than asking them to work it out from a generic template.
Put a named human on the consequential decisions
For every decision that materially affects a person, we identify who owns the outcome, what they see before they decide, how they override the model, and what gets recorded. In practice this usually means AI prepares, ranks and drafts while a person signs. It is the design that leaves you with something to show when someone asks how a decision was reached, and, quietly, it tends to produce better decisions than a model or a rushed human working alone.
Go live, monitor and review on a schedule
We put the controls in place, train the team on what changed and why, and set a review rhythm. The reforms are staged, so obligations arrive over time, and your own tooling will move faster than the law does. New tool, new flow, new obligation: the inventory gets updated and the policy follows. We stay on hand so the work holds up rather than decaying into a folder nobody opens.
Related Reading
AI and Data Privacy
The everyday privacy practices behind an AI rollout.
Privacy Impact Assessment
The assessment that documents an AI project against the APPs.
Disclosing AI to Customers
Transparency obligations where a customer is dealing with AI.
AI Governance Policy
Where these obligations get written down and owned.
Call Recording Laws
The state by state rules on recording customer calls.
AU AI Safety Guardrails
The voluntary standard that expects exactly this discipline.
FAQ
Get Your AI and Privacy Position in Order
Book a free call and we will walk through where personal information meets AI in your business, what the reforms already passed will expect of you, and the short list of changes worth making first. No jargon, no scare campaign, and a clear view of what to put in front of your solicitor.
All discussions held in confidence. Australian-based consultants.