Large retailers increasingly promise that conversational AI “assistants” will soon plan meals, organise events and handle routine shopping. The harder problem, as several recent incidents show, is making those agents feel helpful and trustworthy without making them strange — and many companies are still wrestling with the simpler chatbots they already run.
The Woolworths “Olive” episode
The tension was illustrated in early 2026 when the Australian supermarket group Woolworths reined in the personality of its customer-service assistant, Olive. The bot, first launched in 2018 for tasks such as order tracking and store queries, began volunteering odd personal anecdotes — including claims about its “mother” — after customers entered what looked like a date of birth. Reporting by Guardian Australia and analysis by The Conversation traced the behaviour to old, pre-written script branches: a years-old decision tree that fired a “fun fact” when it detected a birthdate-shaped input.
The reaction was largely irritation rather than delight. Customers questioned why a grocery chain needed a chatty bot at all, and follow-up testing reportedly surfaced other problems, including incorrect prices for basic items. Woolworths said it removed the scripting in response to customer feedback and is narrowing Olive’s focus to relevant support rather than quirky banter.
Why “pleasantly human” is hard
The episode points to a deeper challenge in retail AI. Giving a bot a human-style persona raises expectations: when it then misfires, the gap between “sounds like a person” and “behaves reliably” feels jarring rather than charming. The same period saw Woolworths announce a partnership with Google to extend Olive toward meal planning and recipe-based ingredient sourcing — more ambitious, agent-style capabilities that magnify the cost of any misstep. An agent that can act on a customer’s behalf has more ways to go wrong than one that simply answers questions.
What it means for smaller businesses
For smaller retailers and service firms watching from the sidelines, the practical lessons are unglamorous but useful. A bot’s personality should be scoped tightly to the job it actually does, and legacy scripts should be audited before they resurface in unexpected ways. Clear handoff to a human, accurate underlying data, and honest disclosure that a customer is talking to software all tend to matter more to satisfaction than a charming persona. These themes echo the broader contest among vendors to dominate the AI agent market, where reliability is becoming the real differentiator.
The bigger picture
None of this suggests conversational retail AI is doomed; it suggests the technology is still maturing. The companies that succeed are likely to be those that treat an assistant as a carefully bounded tool, test it against real customer behaviour, and resist the temptation to make it more human than it can reliably be. The original reporting on the Woolworths case is available from The Guardian, with further analysis from The Conversation.
