Woolworths AI Glitch Highlights Risks of Hybrid Legacy-LLM Retail Systems

The recent erratic behavior of Woolworths’ AI assistant, Olive, serves as a stark reminder that the "move fast and break things" ethos of Silicon Valley can have embarrassing, and potentially costly, consequences when applied to the high-stakes world of retail customer service. Australian shoppers interacting with Olive were met with responses that veered from the helpful to the surreal, with the bot claiming to have a "mother" and providing incorrect pricing for staple goods. While these interactions might seem like harmless digital quirks, they expose a deeper structural flaw in how major corporations are currently deploying generative artificial intelligence: the messy collision of cutting-edge Large Language Models (LLMs) with aging legacy infrastructure.

According to Woolworths, the references to Olive’s "mother" were not a case of the AI developing a persona or "hallucinating" in the traditional sense. Instead, they were the result of old decision-tree scripts—some dating back several years—that remained buried within the system’s logic. When a user entered data that the system interpreted as a birthdate, it triggered a "fun fact" script from a previous era of chatbot technology. This highlights a significant challenge for enterprise-level AI rollouts: technical debt. Many retailers are not building AI from scratch but are layering LLMs over existing customer service frameworks. When these layers aren't perfectly synchronized, the result is a "Frankenstein’s monster" of a bot that oscillates between sophisticated language generation and rigid, outdated pre-programmed responses.

“

According to Woolworths, the references to Olive’s "mother" were not a case of the AI developing a persona or "hallucinating" in the traditional sense.

The pricing errors reported by users point to an even more critical failure known as a lack of "grounding." LLMs are probabilistic engines; they predict the next most likely word in a sequence based on training data, not real-time facts. For a retail AI to be useful, it must be tethered to a live database of inventory and pricing—a process often achieved through Retrieval-Augmented Generation (RAG). If this connection is weak or intermittent, the AI will rely on its training data, which may be months or years out of date, or it will simply guess based on patterns. In a retail environment where price sensitivity is at an all-time high due to inflation, providing a customer with an incorrect price isn't just a technical glitch; it’s a breach of trust that can lead to accusations of price gouging or bait-and-switch tactics.