Voluntary AI Safety Standard: A Guide for Retailers

The Voluntary AI Safety Standard gives Australian retailers ten guardrails for using AI responsibly, from assigning an owner to handling shopper complaints. A store deploying a supplied shopping agent still owns the risks of how it uses that agent. Your supplier’s controls support your work, but they don’t transfer that responsibility.

The government evolved the standard into six practices in its Guidance for AI Adoption in 2025. Use that newer guidance for your ongoing work, while the ten-guardrail map below helps you explain an existing safety assessment to your board.

What the Voluntary AI Safety Standard Sets Out

The Voluntary AI Safety Standard is voluntary guidance published in 2024 by the National Artificial Intelligence Centre within the Department of Industry, Science and Resources. It covers organisations that develop or deploy AI across the supply chain. Following it creates no exemption from existing Australian law.

You’re a deployer when you put a supplied agent on your site to help customers shop. You choose its purpose, the catalogue it reads and the actions it can take. Those choices affect the risks you must manage, even when another organisation builds the underlying AI.

The newer Guidance for AI Adoption combines the earlier guidance into six practices:

1. Decide who is accountable.
2. Understand impacts and plan accordingly.
3. Measure and manage risks.
4. Share essential information.
5. Test and monitor.
6. Maintain human control.

For a small store, the foundations guidance gives starting actions. The accompanying implementation guidance supplies more detail for each practice. Keep your record about the actual shopping service you run, including any purchasing or order-support actions.

The Ten Guardrails Applied to a Shopping Agent

The ten guardrails cover the responsibilities below, in their original order. These retail actions are practical applications of the official guardrails, not a government certification checklist. The store remains responsible wherever supplier input is needed.

Guardrail	Retail action	Who does the work?
1. Set accountability and governance	Name a senior owner who approves launch and has authority to pause the agent. Publish how your store governs AI use.	Store leads; supplier identifies its own accountable contacts.
2. Assess and control risks	Record what happens if the agent invents a returns promise or buys an unwanted item. Assign a control to each risk.	Store assesses its use; supplier explains known system risks.
3. Govern and protect data	Decide which catalogue and conversation data the agent needs. Restrict staff access to shopper conversations.	Shared; supplier explains processing, retention and security.
4. Test before launch and monitor afterwards	Test a sold-out variant, a refund request and an ambiguous delivery question. Save the actual replies and review failures after launch.	Store tests its site; supplier provides relevant system test evidence.
5. Keep humans able to intervene	Rehearse pausing purchasing actions and routing a shopper to staff when an answer or order goes wrong.	Store operates the procedure; supplier provides the controls.
6. Tell people when AI is involved	Label the chat as AI before a shopper starts. Explain which actions it can take on their behalf.	Store owns the notice; supplier explains agent behaviour.
7. Let people challenge outcomes	Give shoppers a clear route to a staff member who can review a disputed recommendation or transaction.	Store handles the complaint; supplier helps investigate system faults.
8. Exchange information across suppliers	Tell the supplier your intended use and request details of the underlying models and known limitations.	Shared; each party supplies information the other needs.
9. Keep evidence of your controls	Save the approved settings, test results and change history so an assessor can trace a release.	Store keeps its record; supplier provides version and system records.
10. Involve affected people	Ask customers with access needs and customer-service staff where the chat causes confusion. Change the service when their feedback exposes harm.	Store leads engagement; supplier helps resolve technical barriers.

Treat these as ongoing responsibilities. Reassess the controls when you add an action, change the supplier or connect a different source of shopper data. For a detailed reply-testing workflow, use the guide to chatbot testing.

Australian AI Ethics Principles

Australia’s AI Ethics Principles are eight voluntary principles published by the Australian Government’s industry department in 2019 as part of its AI Ethics Framework. The Department of Industry, Science and Resources maintains the official principles page. It also directs organisations to the newer Guidance for AI Adoption.

The principles describe the outcomes responsible AI aims for. The guardrails turn those aims into organisational work. The overlaps below are a retail interpretation, not a separate official scoring system.

Official principle	What it means in shopper conversations	Relevant guardrails
Human, societal and environmental wellbeing	Consider harm to customers alongside the commercial benefit of recommendations.	2 and 10
Human-centred values	Keep the shopper in control of buying decisions and avoid pressure that undermines their choices.	2, 5 and 10
Fairness	Test whether different customers can access assistance and receive fair treatment.	2 and 10
Privacy protection and security	Limit conversation data to what the service needs and protect access to it.	3 and 8
Reliability and safety	Check that answers match product details and store policies.	2 and 4
Transparency and explainability	Tell shoppers they’re talking to AI and explain why a product suits their stated needs.	6 and 8
Contestability	Let a shopper dispute an outcome and obtain a human review.	7
Accountability	Make it clear who owns the service and keep evidence of decisions.	1, 5 and 9

For fairness, compare answers to the same shopping need expressed in different ways. Check that the agent doesn’t make unsupported assumptions about a person. For transparency, explain the recommendation through catalogue facts, such as fit or delivery availability.

A complaint route must lead to someone who can act. A notice saying the chat uses AI doesn’t resolve a wrong promise or an unauthorised purchase. Staff need enough conversation context to investigate without exposing unrelated shopper data.

Mandatory Guardrails: Where the Proposal Stands

The government’s 2024 proposal considered mandatory guardrails for AI development and deployment in high-risk settings. The official consultation outcome says the government will not proceed at this time with that proposal. Feedback informed the National AI Plan in 2025.

There is separate legislative work in 2026. The Department of the Prime Minister and Cabinet’s AI standards consultation paper covers large data centres and frontier AI training. Andrew Charlton’s Sydney Trust and Safety Festival speech confirms that AI standards legislation is being developed and discusses regulation of serious frontier-model risks.

Those announcements don’t establish an enacted, general high-risk classification for a retailer’s shopping agent under the earlier mandatory-guardrails proposal. A store can’t claim a settled exemption under that proposal either. Assess your actual use, especially actions involving purchases or personal information, while continuing to meet applicable law.

Use the guide to AI regulation in Australia for the wider legal position. The voluntary safety record below helps you manage the service. It doesn’t determine your legal obligations.

Questions to Put to Your AI Supplier

Request evidence you can use for your store’s deployment, with a named supplier contact for follow-up. A general responsible-AI statement won’t explain what happens in your customers’ conversations.

• Ask for test methods and results covering catalogue accuracy, policy answers and any purchasing actions you enable.
• Request a data-flow explanation covering conversation storage, subprocessors, retention, deletion and any use for model training.
• Ask which records identify the model, settings and catalogue version behind a disputed reply.
• Get instructions for pausing the agent, disabling an action and restoring a previous configuration. Confirm what happens to orders already started.
• Agree how the supplier reports incidents and behaviour changes, including who contacts your store and what investigation evidence you receive.
• Request known limitations and the division of responsibilities for complaints and corrective action.

Keep the answers beside your own tests and launch decision. For a wider procurement assessment, read the AI vendor risk assessment guide.

A One-Page Record for Your Store

A one-page record can capture the agent’s purpose and your operating decisions, with links to supporting evidence. It’s a summary you maintain, not a substitute for the underlying records.

This hypothetical store record shows the level of detail to start with:

Field	Example entry
Purpose and limits	Help customers choose products and answer delivery questions. Purchasing actions are disabled for this deployment.
Accountable owner	Ecommerce manager, with authority to pause the agent and approve changes.
Known risks	Invented delivery promises, wrong variant recommendations and unnecessary personal information in chat.
Controls	Answers use approved catalogue and policy sources. Staff review disputed replies through the published complaint route.
Launch tests	Sold-out variant, unclear delivery destination, conflicting policy documents and a request containing personal information. Link each result and corrective action.
Supplier evidence	Link the data-handling explanation, system limitations and incident contact details.
Human intervention	Record who pauses the agent, how they do it and the result of a rehearsal.
Approval and review	Record the launch decision, last review date and next review date. Reopen the assessment when actions or data sources change.

Keep failed tests as well as successful ones. When a problem appears, record the action you took and whether the same shopper request now produces an acceptable result. Use the AI governance framework guide when your store needs a process covering several AI systems.

Before launch, have the accountable owner follow the evidence links and rehearse the pause procedure. They must be able to explain which risks the store controls itself and which controls rely on the supplier.