ese systems are predictable but limited, offering little more than a clickable FAQ. In the past year, however, major chains including Woolworths, Kmart and Bunnings have begun rolling out generative chatbots built on large language models from big tech providers such as OpenAI and Google. These assistants can handle open-ended, multi-step questions, hold conversational dialogue across the purchase funnel and surface personalised recommendations in real time. The next horizon is “agentic” AI: systems that do not just respond but take actions, such as rebooking deliveries or placing grocery and travel orders, with minimal prompts. That makes them powerful operational tools – and potentially high-risk intermediaries between retailers and consumers. When bots misbehave, the law bites The Australian Competition and Consumer Commission (ACCC) has made it clear that businesses cannot sidestep their obligations under the Australian Consumer Law (ACL) by blaming an algorithm. In its 2025 AI industry snapshot, the regulator warned that AI-generated responses which mislead consumers may expose companies to enforcement action and significant penalties. Refunds, returns and consumer guarantees are a particular flashpoint. The ACL sets out specific rights when goods or services fail to meet consumer guarantees, and the ACCC has previously pursued businesses that misrepresented these rights in frontline interactions. If a chatbot tells a shopper they are not entitled to a remedy, or over-promises on refunds or discounts, that digital interaction can amount to misleading or deceptive conduct. The same principle applies to safety and regulated advice. Bunnings was forced to tighten safeguards on its AI assistant after it provided step-by-step instructions for replacing an extension-cord plug – work that is illegal for unlicensed customers in Queensland. The incident underscored how a friendly DIY bot can accidentally walk customers into unlawful or dangerous territory. Case study: Woolworths’ talkative assistant Woolworths’ “Olive” shopping assistant, recently upgraded using Google Cloud’s Gemini platform, became a cautionary meme when customers shared screenshots of the bot rambling about its “mother” and family experiences instead of focusing on groceries. Beyond the surreal small talk, subsequent testing revealed that Olive also struggled with basic product pricing, offering unclear or incorrect answers while diverting to stock checks and pickup fees. Woolworths moved quickly to adjust the system, but the episode highlights two structural risks with generative interfaces. First, guardrails do not fully eliminate off-topic or anthropomorphic outputs, which can erode trust when a bot sounds sentient but cannot reliably answer simple questions. Second, even “minor” pricing inaccuracies at scale create exposure under consumer law, particularly for essential goods in a cost-of-living crisis. Governance, not gadgets For regulators, the lesson is that AI chatbots are simply a new front line of customer service – and therefore subject to the same consumer and privacy rules as human agents. For retailers, the implication is that adopting AI is not a technology project; it is a governance challenge. Legal advisers now recommend that retailers treat chatbots and recommendation engines as part of an enterprise-wide AI inventory, classified by risk and tied to clear executive accountability. Formal AI governance frameworks should set boundaries on what systems are allowed to do, specify escalation to humans for high-stakes issues such as refunds and safety, and require active monitoring of interactions for emerging harms. The Commonwealth Government’s voluntary AI Safety Standard offers a starting point, calling for controls on data quality, governance, privacy, cybersecurity and technical limits on automated capabilities. While non-binding, these guardrails closely align with the ACCC’s expectation that businesses proactively manage AI risks rather than wait for a test case. Designing for friction where it matters Despite the headlines, the incentives to automate remain strong: conversational AI promises 24/7 availability, instant responses across channels and more consistent information than over-stretched call centres. When well designed, chatbots can reduce friction and improve outcomes by resolving routine issues quickly and leaving human agents free to handle complex complaints. But the Woolworths and Bunnings episodes show that not all friction is bad. Retailers may actually need to design in “speed bumps” – forcing a hand-off to a person for anything involving consumer guarantees, refunds, safety advice or substantial financial commitments. Disclosing a bot’s limitations, logging and auditing conversations, and benchmarking responses against ACL requirements are no longer nice-to-haves; they’re table stakes for responsible deployment. As agentic systems creep closer to the transaction layer, the stakes will only rise. Retailers are bound by the promises their bots make, whether or not those promises were in the script – and regulators are watching closely. The question for the industry is no longer whether to use AI, but how to build it into retail operations without letting risky promises write the rulebook. Further reading: What retail brands can learn from Woolies’ virtual assistant Olive