Under the Australian Consumer Law (ACL), a business is responsible for what its AI tells customers, exactly as it is for what its staff say. AI customer-facing systems create exposure under sections 18, 29 and 60 to 62, and section 64 means consumer guarantees cannot be excluded, restricted or modified by a disclaimer. The fix is controls in place before launch.
This guide is for general counsel, CIOs, COOs and customer operations leaders at mid-sized, enterprise and multi-site Australian organisations deploying AI that talks to customers. It maps where AI and Australian Consumer Law intersect and sets out a practical control stack. It is general information, not legal advice.
Key takeaways
- The ACL (Schedule 2 of the Competition and Consumer Act 2010) is technology-neutral. AI output that reaches a customer is treated as the business's statement.
- Four representation traps cause most problems: pricing, refunds and warranties, capability claims, and service commitments.
- Section 64 makes the consumer guarantees in sections 60 to 62 non-excludable, so "the chatbot's answers are not binding" terms do not release you.
- There is no formal ACL safe harbour for AI, but scope documents, guardrails, full logging, human approval for material commitments and disclosure create a defensible record.
- From 10 December 2026, Privacy Act automated decision-making transparency rules add disclosure obligations for some AI uses.
Why does an AI agent's statement bind the business?
In February 2024, the British Columbia Civil Resolution Tribunal decided Moffatt v Air Canada (2024 BCCRT 149). The airline's website chatbot told Jake Moffatt he could book a flight and claim a bereavement discount within 90 days. That was not Air Canada's policy. The tribunal rejected the airline's argument that the chatbot was a separate entity responsible for its own actions and held the airline liable.
The case was Canadian and the amount small, but the logic is the same one an Australian court or tribunal would apply under the ACL: a business is responsible for representations made on its behalf, whether by staff, contractors or a language model.
How does the Australian Consumer Law apply to AI?
The ACL does not say "by a human employee" anywhere. A representation in trade or commerce is a representation whether it comes from a contact centre agent, a brochure or an AI model. The ACCC's compliance and enforcement priorities regularly include consumer and fair trading issues in the digital economy, and AI-generated conduct falls squarely within that.
The provisions every CIO, COO and general counsel should know:
Section 18: misleading or deceptive conduct. A person must not, in trade or commerce, engage in conduct that is misleading or deceptive or likely to mislead or deceive. Intent is not required. A confidently wrong AI answer is enough.
Section 29: false or misleading representations. Covers representations about the standard, quality, value, performance characteristics or sponsorship of goods and services, plus price, testimonials and warranty terms. For a body corporate, the maximum penalty is the greater of $50 million, three times the benefit obtained, or 30 per cent of adjusted turnover during the breach period.
Sections 60 to 62: consumer guarantees for services. Services must be supplied with due care and skill (s60), be reasonably fit for any disclosed purpose (s61) and be supplied within a reasonable time where none is agreed (s62). When AI helps deliver a service, the guarantees attach to that service.
Section 64: guarantees cannot be excluded. Consumer guarantees cannot be excluded, restricted or modified by contract. A clause saying AI outputs are not binding does not release the business from its s60 to 62 obligations.
Directors should also note their duties of care and diligence under sections 180 and 181 of the Corporations Act 2001. Foreseeable AI-amplified consumer law risk is the kind of operational risk a reasonable board is expected to govern.
What are the four representation traps in customer-facing AI?
| Common failure pattern | What good looks like | Impact | |
|---|---|---|---|
| Pricing (s18, s29) | Agent confirms a stale price, or is manipulated into accepting an absurd offer | Prices come from one source-of-truth system with a hard refusal outside an authorised range, and material transactions need human confirmation | ACL s18, s29 |
| Refund / warranty / return (s18, s29, s60) | Agent invents or misstates the refund window, return conditions or warranty coverage | Agent retrieves the live published policy, quotes it verbatim, and cannot extend it without a human approver | ACL s18, s29, s60 |
| Capability or performance claims (s18, s29, s61) | Agent affirms a use case the product does not support, or overstates compatibility or certifications | Answers constrained to a curated knowledge base; unsupported claims trigger refusal or handover | ACL s18, s29, s61 |
| Service commitments (s18, s62) | Agent promises a delivery date, callback time or service level the business cannot meet | Time commitments calculated from live capacity data and never asserted beyond what the business has approved | ACL s18, s62 |
The pricing trap has a well-known example. In December 2023, a user manipulated a Chevrolet dealership's chatbot in Watsonville, California into "agreeing" to sell a new Tahoe for one dollar. The dealer did not honour it, and probably did not have to, but the episode shows how cheaply a customer-facing agent can be pushed into statements the business never authorised.
How does ACL liability flow from an AI conversation?
- Customer queryConsumer interacts with the AI agent in trade or commerce
- AI processesRetrieval, generation, or both, shaped by prompts and tool calls
- Representation madeThe output reaches the consumer as the business's statement
- Customer relianceConsumer acts on it: books, buys, claims, cancels
- Enforceable representationAssessed under ACL s18, s29 and s60-62
- Exposure crystallisesComplaint, ACCC action, tribunal claim, class action, or AFCA referral in regulated sectors
Notice what is missing. No step lets the AI vendor absorb the liability. No disclaimer breaks the chain. "It was a hallucination" does not reduce the obligation. See operating AI agents in production for how this risk shows up after launch.
What do vendor pitches get wrong about AI and consumer law?
"It is the AI vendor's responsibility." The ACL attaches to the business supplying goods or services in trade or commerce, which is the business deploying the agent. A vendor indemnity may give you a commercial claim against the vendor, but it does not stop the ACCC acting against you or a consumer claiming against you in a state tribunal or federal court.
"Our terms and conditions disclaim it." Section 64 makes the consumer guarantees non-excludable. You can describe the scope of a service, but you cannot contract out of due care and skill or fitness for purpose. The ACCC's guidance on unfair contract terms further limits what disclaimers can do in standard form contracts.
"It is just a chatbot." What matters is whether a reasonable consumer would treat the output as the business's statement. They do, and Moffatt is the clearest illustration.
Our AI vendor selection questions include a section for getting these claims out of the room during procurement.
What is the safe-harbour pathway for AI under the ACL?
There is no formal safe harbour for AI-mediated representations. A control stack materially reduces exposure and creates a record a regulator, court or tribunal can review.
- Pre-deploymentLegal scope reviewCounsel maps every category of representation the agent may make and every category it must refuse, in a written scope-of-authority document.
- Pre-deploymentGuardrails and refusal patternsSystem instructions, refusal patterns for out-of-scope queries, and escalation triggers for high-value or high-risk interactions, built before any customer traffic.
- Pre-deploymentFull-content logging and retentionEvery interaction logged with prompt, output, timestamp, user identifier and model version. Retention set against warranty and limitation periods, and at least the ATO's five-year minimum where transactions are involved.
- Pre-deploymentHuman approval for material commitmentsRefunds above an agreed threshold, pricing exceptions, warranty extensions and service promises go to a person before the customer hears yes.
- Go-liveAI-interaction disclosureCustomers are told in plain language at the start that they are dealing with AI.
- Steady stateIncident response runbookContainment, customer remediation, escalation, regulator posture for systemic issues, board reporting.
- Steady stateInsurance reviewConfirm in writing that professional indemnity, cyber and management liability policies cover AI-related consumer law exposure, and check for AI exclusions or sub-limits.
The privacy overlay matters too. The Privacy and Other Legislation Amendment Act 2024 introduces automated decision-making transparency obligations from 10 December 2026. Where a computer program is used to make, or substantially help make, decisions that could significantly affect an individual's rights or interests, the privacy policy must say so, including the kinds of personal information used. The OAIC's guidance on commercially available AI products sets further expectations, and maximum penalties for serious interferences with privacy reach $50 million for a body corporate. See our guide to Privacy Act automated decisions.
For the wider governance overlay, see AI agent governance, data access and human override and our data sovereignty guide.
What ACL exposure level is your AI deployment?
What controls are in place on your customer-facing AI?
- Internal only; staff review outputs before any customer communicationLow exposure: standard governance and accuracy monitoring
- Talks to customers on narrow informational topics only, no pricing or refund authority, human escalation availableLow to medium: confirm logging, disclosure and escalation work
- Quotes pricing, policies or capabilities with logging and disclosure, but no legal scope reviewMedium: commission a scope-of-authority review and tighten refusals on s29 categories
- Can commit to refunds, pricing or service levels without human review, incomplete logging, no executive sign-offHigh: pause, add human approval, complete scope review and logging before resuming
- Regulated sector (financial services, health, legal, energy retail) with direct customer interactionHigh regardless: sector obligations stack on the ACL; review against ASIC, AHPRA, TGA or AER guidance
High-exposure deployments are common. An operations or marketing team stands up an AI tool quickly, legal is not consulted, and the tool makes representations nobody has reviewed. Our piece on why not to DIY AI agents without understanding them covers this pattern.
Pre-launch ACL readiness checklist
If the answer to any question is "I do not know", that is your next two weeks of work.
- Authorised representations: have we documented what the AI may say and what it must refuse?
- Logging: where is every input and output logged, for how long, and can we retrieve a specific interaction within hours of a complaint?
- Change control: who signs off a new prompt, knowledge source or model upgrade, and is there a change log?
- Customer disclosure: where do we tell customers they are dealing with AI, and has it been reviewed against OAIC guidance?
- ACL mapping: is every AI touchpoint mapped to s18, s29 and s60 to s62, with failure modes identified?
- Human approval thresholds: above what value or risk does a representation need human approval?
- Out-of-scope handling: does the AI refuse, hand over or guess, and can we evidence it with test cases?
- Regulator path: if a misleading representation reaches many customers, who decides our response and who speaks to the ACCC?
- Insurance: have we confirmed cover in writing for AI-related consumer law exposure?
- Vendor liability: does our vendor agreement allocate liability for misleading outputs in a commercially meaningful way?
- Sector obligations: have we checked the relevant regulator's AI guidance?
- Board reporting: is AI deployment risk reported to the board or executive on a set cadence?
A simple test: sit your most senior non-technical person in front of the agent and ask them to break it. If they produce a statement the business would not want to honour, it is not ready.
How Solve8 can help
Our consultants and engineers bring legal scope, guardrails, logging and human approval together before a customer-facing AI goes live, and we build systems that run on your infrastructure in Australia with a person approving material commitments. We built RootCauseAI, our own on-premise investigation product, with audit-grade logging from day one. See our AI strategy and managed AI services pages, or talk to us about a pre-launch review.
