Australia's Voluntary AI Safety Standard, Explained for Ordinary Businesses
Australia has a Voluntary AI Safety Standard with ten guardrails, published by the Australian Government through the National AI Centre. It is voluntary. It is not law, and nobody will fine you for ignoring it. What it does give you is the clearest published answer to the question every owner, board and insurer is starting to ask: how do you know your AI is not quietly hurting someone? This page walks through all ten guardrails in plain English and, for each one, what a business of 10 to 200 people actually has to do.
Written for the business that is already using AI, not the one thinking about it. If your team drafts customer emails with an AI tool, your phones are answered by one, your CRM has AI features switched on, or someone is pasting client documents into a chat window, the guardrails apply to you today. Australian context throughout: Privacy Act, Australian Privacy Principles, Australian Consumer Law, Fair Work and the industry rules you already live under.
Realistic ROI
Why Work Through the Guardrails With Yes AI
The ten guardrails are short, sensible and free to read. The hard part is not understanding them, it is applying them honestly to a real business that already has AI scattered through it, without buying a compliance apparatus built for a bank. Four things shape how Yes AI does this work.
We start with what you are actually running
A governance project that starts with a document is already wrong. We start by finding the AI in your business, which is nearly always more than the leadership team thinks: the writing tool the sales team pays for personally, the AI features quietly switched on inside your CRM, the transcription service sitting on every client call, the assistant built into your office suite that nobody counts as AI. Until that list is honest, every policy you write describes a business that does not exist.
Sized for a real Australian SME, not a bank
The guardrails were written to be applied in proportion to risk, and that is the part people miss. A 25 person accounting firm using AI to draft correspondence does not need the apparatus of a lender scoring loan applications, and pretending otherwise is the surest way to have the whole thing quietly abandoned within a month. We go heavy where a decision could genuinely hurt someone and light where the worst case is a clumsy internal summary. You end up with the smallest set of controls that would stand up if a client, an insurer or a regulator asked.
We are honest about legal status
There is a lot of AI compliance being sold in Australia right now on the implication that a law already exists. It does not. We will tell you clearly what is voluntary, what is proposed and still under consideration, and what is already binding on you through the Privacy Act, the Australian Consumer Law, Fair Work and your industry obligations. We will not dress guidance up as regulation to make a sale, and if the honest answer is that a particular guardrail barely touches your business, that is what we will say.
We build the controls, not just describe them
We are a consultancy that also builds. So when the answer is an approval step before an AI drafted quote reaches a customer, a log that captures which version of a system made which decision, a disclosure line at the right point in a call flow, or a complaint route that actually reaches a person, we implement it inside the systems you already run rather than handing over a list of recommendations and leaving. Governance that lives only in a document decays quietly. Governance built into the workflow keeps working after everyone has forgotten the project.
All Ten Guardrails, and What You Actually Have To Do
The Voluntary AI Safety Standard sets out ten guardrails. Grouped below into six pieces of work, with the doing-list for each. Read them as tasks with an owner and a date, not as principles to agree with.
Accountability you can point to
Name one person who owns AI in the business and put it in writing where staff can see it. In a business of this size that is usually the general manager or the operations lead, not an outside consultant and not the vague answer of IT. Three things need to exist on paper: who decides whether a new AI tool may be used, what the business is trying to achieve with AI, and how you will keep up with the laws that already apply to you. Add short training so people know the rules exist and where to find them. One page that is true beats a manual nobody opens.
A risk process, not a risk feeling
List every AI use, then rate each one on two questions: how badly could this go for a person, and how quickly would we notice if it went wrong. A tool drafting internal meeting notes is low. A tool that screens job applicants, sets credit terms, triages a health or safety issue, or speaks to a distressed customer is high, and the standard expects the controls to scale with that. Write the rating down with a date and the name of whoever did it. The score itself matters far less than the fact that someone thought about it before go live and can show they did.
Protect the system, govern the data
Two halves. Protect the system with real accounts, multi-factor authentication, no shared logins, least privilege, and a written rule about which tools staff may put customer information into. Then govern the data: know where it came from, whether you had the right to use it, whether it holds personal information under the Privacy Act, which country it is hosted in, and how long it is kept. The question that catches most businesses out is whether their customer records are being used to train someone else's model. Get the vendor answer in writing and keep it with the register.
Test it, then keep a human genuinely in charge
Before go live, run the system against real cases where you already know the right answer, including the awkward and unusual ones, and record how it did. After go live keep watching, because performance drifts as your products, prices, people and customers change. Then make sure a person can see what the system decided, override it, and stop it entirely. Meaningful oversight means that reviewer has the reason, the time and the authority to disagree. Someone clicking approve on two hundred items an hour is not oversight, it is decoration.
Tell people, and let them push back
If a customer is talking to an AI, or an AI helped decide something about them, say so in plain words at the moment it matters, not buried on a terms page. Label AI generated material a reasonable person would otherwise take as human work. Then give people a real route to challenge an outcome: a named contact, a stated timeframe, and a human with the authority to reverse it. Log every challenge and what happened. That log is your evidence you handled it properly, and it is the fastest way to find the cases where the system is quietly wrong.
Supply chain, records, and the people affected
Most Australian SMEs buy AI rather than build it, so guardrail 8 mostly means asking suppliers the right questions and passing what you learn to anyone downstream who depends on your output. Guardrail 9 is records: keep enough of a trail that an auditor, an insurer or a customer could check what you did, including versions, decisions, test results and who approved what. Guardrail 10 is engaging the people affected before you deploy, your own staff first, then customers and anyone in a vulnerable position, with a real eye on fairness. That last one is the guardrail most businesses skip and the one that surfaces problems earliest.
Six Everyday Situations and What the Guardrails Ask Of You
| Task | Traditional | With Yes AI | Notes |
|---|---|---|---|
| AI answering the phone at a 40 person clinic | Callers assume they are speaking to a receptionist and nobody tells them otherwise | The system says it is an AI in the opening line and any caller can ask for a person | Guardrail 6 in one sentence of script. Add a hard rule that anything clinical, distressed or urgent transfers to a human immediately, and keep the transcript against the record. |
| AI shortlisting job applications at a labour hire firm | The tool ranks applicants and the recruiter works down the list | Rated high risk, tested against your own past hiring, and every rejection reviewed by a person | This is where anti-discrimination law and Fair Work obligations bite hardest. Never let the system auto-reject, and keep criteria you could defend out loud to the applicant. |
| AI drafting quotes from inbound emails at a building supplier | The draft goes out as written if it looks about right | A person checks price, scope and terms before anything reaches the customer | The Australian Consumer Law does not care that an AI wrote it. A wrong quoted price is still your representation, so the oversight step sits before send, never after. |
| Staff pasting client files into a free AI tool | Nobody knows it is happening until something leaks or a client asks | An approved tool list, a written rule on what may be pasted, and business plans with training switched off | Shadow AI is the most common gap we find. Guardrail 3 starts with an honest amnesty: ask the team what they already use before you write a single rule. |
| AI chatbot handling refunds on a retail website | The bot decides, and unhappy customers end up arguing with the bot | The bot is labelled, the refund rule is disclosed, and a named person handles disputes within a stated timeframe | Guardrail 7. A contest route that loops back to the same bot is not a contest route, and the ACCC position on misleading conduct applies to what your bot says. |
| AI scoring payment risk to set a wholesaler's credit terms | A score appears and the account goes on hold with no stated reason | The inputs and reason are recorded, the customer can ask why, and a person can lift the hold | Keep the model version, the inputs and the decision together. Credit and payment decisions are exactly the sort of use the proposed mandatory guardrails describe as higher risk. |
Six Honest Warnings Before You Start
Voluntary does not mean unregulated
The standard is guidance, not law, and there is no penalty attached to ignoring it. What will bite you is the law that already applies to whatever the AI is doing: the Privacy Act and the Australian Privacy Principles wherever personal information is involved, the Australian Consumer Law and the ACCC where you make claims or handle customers, Fair Work and anti-discrimination law in anything touching employment, and your own professional or industry obligations. Working through the guardrails is a sound way to meet those existing duties. It is not a shield in itself, it is not a certification, and there is nothing to wave at anyone at the end of it. Treat this page as general information about the standard rather than legal advice, and take your own advice on how those laws land on your particular business.
The mandatory guardrails are proposed, not passed
The Australian Government consulted on mandatory guardrails for AI in high-risk settings, and that work remains under consideration rather than in force. Be very careful with anyone who tells you Australia has an AI Act, that you are legally required to comply, or that buying their product makes you compliant. None of that is true today. Check the current position with the Department of Industry, Science and Resources rather than relying on any vendor, or on this page. What is genuinely worth doing now is building the habits, because a business already keeping a register and a record barely notices if a mandatory regime does arrive.
A policy nobody reads is worse than no policy
It is easy to buy a thirty page AI policy, file it, and feel finished. That document becomes a liability the moment your actual practice differs from it, because you have written down a standard you are demonstrably not meeting. Aim for the shortest thing your team will genuinely follow: an approved tool list, a rule about what may and may not be put into an AI tool, a named owner, and a register that is current. Then check twice a year whether the business still works that way. Short and true beats long and aspirational every single time.
Human oversight fails quietly
Oversight is the guardrail that looks fine on paper and rots in practice. If the reviewer is pushed through two hundred items an hour, has no visibility of why the system decided what it decided, or gets pulled up for slowing the queue, they will approve everything, and you have automated the decision without ever deciding to. Design it so the person can genuinely say no: give them the reason, give them the time, sample what they approved, and track how often they override. An override rate sitting at zero for months is a warning sign, not a success.
You cannot see inside your vendor's model, so ask for what you can get
Guardrail 8 asks for transparency across the AI supply chain, and for most SMEs that means depending on suppliers who will not, and often genuinely cannot, explain how their model works. Do not chase the impossible. Ask for what a vendor can answer in writing: where the data is hosted, who their subprocessors are, whether your data trains their model, how they handle a security incident, how much notice you get before the model changes underneath you, and what testing they have done. A vendor who will not answer those in writing has told you something useful.
AI should not govern itself
Do not ask a model to write your risk assessment, rate its own risk level, or generate the register and then file it unread. The output will read beautifully and mean nothing, and you will have manufactured exactly the documentation theatre the guardrails exist to prevent. Use AI to draft or to summarise if it helps, then have a person who knows the business argue with every line. The judgement calls, which uses are high risk, what fairness means for your customers, who is accountable when it goes wrong, are business decisions and they belong to people.
How Yes AI Helps You Meet the Guardrails
The honest AI inventory
We find every AI tool actually in use, including the ones bought on a personal card, the features switched on inside software you already pay for, and the ones nobody thought to mention. Then we map what data each touches, who relies on its output, and what would happen if it were wrong. Most owners are surprised by the length of that list, and it is the only sensible foundation for everything that follows.
Risk rating and a gap review against the ten guardrails
We rate each use for real world harm and for how quickly you would detect a failure, then walk all ten guardrails against your reality and tell you plainly where you already comply, where you are close, and where there is nothing at all. You get a prioritised list with an owner and a date beside each item, not a scored report card that changes nothing on Monday.
Documents your team will actually use
We write the short versions: the acceptable use rule, the AI register, the disclosure wording for customers, the challenge and complaint route, and the questions to put to your vendors. Plain Australian English, sized for a business of 10 to 200 people, and shaped so an owner or board can read and sign them off without a lawyer translating first.
Controls built into the workflow, not bolted beside it
Governance only survives if it happens without anyone remembering to do it. We build the approval step, the audit trail, the disclosure line, the complaint capture and the review reminder into the systems your team already uses, so the record accumulates as a by-product of doing the work rather than becoming a chore someone has to catch up on before an audit.
Getting to Working Governance in Five Steps
Most Australian businesses of 10 to 200 people can go from nothing written down to a genuinely defensible position in two to six weeks. Nobody needs a compliance department. What they need is an accurate list, an honest rating, a short set of rules and the habit of writing things down as they happen.
Find every AI already in use
We talk to the teams, review the tools and subscriptions, and build the register: what it is, who uses it, what data goes in, what comes out, which decisions depend on it, and who the vendor is. This is the step people want to skip and the step that changes the answer, because the risky uses are almost never the ones on the executive slide.
Rate the risk and find the gaps
Each use gets rated for how much harm it could do and how fast you would notice a failure. We then run all ten guardrails against your reality and produce a short list of gaps, each with an owner and a due date. Low risk uses get a deliberately light touch so the effort lands where it can actually prevent harm.
Write the short documents and name the owner
One accountable person, one acceptable use rule, one register, one disclosure line, one challenge route. We draft them in plain English, test them with the people who have to live with them, and get them signed off. If a document cannot be explained to a new starter in five minutes, it is too long and it will be ignored.
Build the controls into the work
We wire the approvals, logs, disclosure wording, complaint capture and review reminders into your existing systems so the evidence builds itself. Testing gets scheduled rather than promised, oversight gets a real queue with real time allowed, and the register updates when someone adds a tool instead of at audit time.
Review, record and keep it current
We set a review rhythm, usually quarterly for higher risk uses and twice a year for the rest, with a standing agenda: what changed, what went wrong, what the override and complaint logs show, and what new tools appeared. You finish with a dated trail that answers guardrail 9 without anyone reconstructing history from memory.
Related Reading
AI Governance Policy
Turn the guardrails into a policy your board can actually sign off.
AI Risk Assessment
Work out which of your AI uses are genuinely higher risk.
ISO/IEC 42001
The certifiable management system that formalises the same disciplines.
AI Acceptable Use Policy
What your staff may and may not put into an AI tool.
Disclosing AI to Customers
When you should tell a customer they are dealing with AI.
AI and Data Privacy
Privacy Act and APP obligations where AI touches personal information.
FAQ
Get Ahead of the Guardrails While They Are Still Voluntary
Book a free call and we will walk the ten guardrails against your business, tell you plainly where you already comply and where the real gaps are, and show you the shortest path to a position you could defend. No obligation, no scare tactics, and no pretending guidance is law.
All discussions held in confidence. Australian-based consultants.