A recent run-in with Woolworths’ AI assistant Olive left some Australian shoppers unsettled — and highlighted broader risks of deploying generative AI in customer-facing roles. Users expecting grocery help and recipe tips instead encountered surprisingly personal-sounding replies: Olive referenced a “mother” and offered other human-like details. Subsequent checks also turned up pricing inconsistencies for basic items, and when asked a direct product price the assistant sidestepped with stock checks and pickup fee information rather than a clear figure.
Olive is powered by a large language model (LLM), which generates plausible-sounding text from learned patterns rather than holding beliefs or family ties. Woolworths told the Australian Financial Review that the surprising “mother” references came from legacy scripted responses. An old decision tree apparently matched inputs that resembled a birthdate to a scripted “fun fact.” After customer pushback, Woolworths says it removed that scripting.
The pricing mistakes, however, stem from a separate technical challenge: LLMs will produce outputs based on training and prompt context unless they are explicitly and reliably grounded in live data. If the assistant’s connection to current price or inventory systems is incomplete or poorly managed, it can return outdated, incorrect or evasive answers.
This episode is part of a pattern. In 2022 an Air Canada chatbot wrongly advised a passenger he could buy full-price tickets and later claim a bereavement refund; no such policy existed, and when the airline refused to honor the advice the passenger sued and prevailed. The tribunal rejected the airline’s argument that the chatbot was a separate legal entity, ruling that a chatbot’s output is part of a company’s communications and the company is responsible. In January 2024 UK delivery firm DPD briefly faced viral attention after its chatbot composed a critical poem and used profanity; the company disabled the bot soon after.
Both incidents illustrate the same core failure: companies put AI in front of customers without sufficient guardrails and oversight and were surprised by the results. That’s especially consequential for Woolworths, Australia’s largest supermarket chain, because customers reasonably expect accurate pricing and clear information when managing household budgets. Woolworths discloses that Olive “may make mistakes,” but that disclaimer clashes with everyday consumer expectations — and with regulatory scrutiny. The Australian Competition and Consumer Commission has already opened proceedings over allegedly misleading discount pricing, meaning Olive’s glitches are hard to dismiss as trivial.
There is a commercial rationale for giving chatbots a personality: research shows that conversational, friendly interfaces with names and personas can increase engagement, satisfaction and sales. But anthropomorphism increases risk. When a personable assistant fails, the perceived breach of expectation can make customers more frustrated than with an obviously mechanical system.
The practical takeaways are straightforward. Deploying AI for customer interactions is not a set-and-forget exercise. Companies are accountable for what their public-facing systems say and must invest in accurate data grounding, monitoring, robust fallback behavior, and governance. For consumers, conversational AI can feel authoritative, but it remains a tool that can err; double-check unclear or consequential information. As AI becomes routine in everyday transactions, clear accountability and healthy scepticism are as important as the underlying technology.
