Woolworths has become the first Australian retailer to run a consumer assistant on Google’s agentic Gemini platform, switching its Olive chatbot from a scripted refund tool into an AI that can read a handwritten recipe, build a basket and, with a shopper’s consent, handle checkout.
The grocer put the upgraded assistant in front of more than 200,000 staff in April, ahead of a consumer rollout that began in July to app pick-up and delivery customers. iTnews reported that Olive already handles more than 70 per cent of the retailer’s contact-centre interactions, making the shift one of the largest live agentic deployments in Australian retail.
Woolworths is an early user of Gemini Enterprise for Customer Experience, a Google Cloud product unveiled at the NRF 2026 retail conference in January. According to Google Cloud’s announcement, the platform folds shopping and customer service into one agent, with US grocer Kroger, hardware chain Lowe’s and Papa Johns named alongside Woolworths as launch customers.
From refund bot to shopping companion
The old Olive was, in Woolworths’ own words, a “deterministic chat and voice bot” that answered set questions and processed refunds. The agentic version reasons across a task. Technology director of digital experiences Venky Erode Sivasubramaniyam demonstrated it at Google Cloud Next ’26, showing Olive identify a spaghetti carbonara from a photo, suggest organic and budget swaps, calculate the savings and assemble the ingredients into a cart.
iTnews reported the assistant can also read a photo of a handwritten shopping list or recipe, apply eligible rewards discounts and complete a purchase once the customer agrees. Woolworths says it moved Olive from scripted bot to agent in roughly six months.
Sivasubramaniyam pointed to where the value sits. He noted that about 80 per cent of groceries are repeat purchases, and said the aim is proactive basket suggestions that nudge on-demand shoppers toward larger weekly shops. That framing matters: for a supermarket, an assistant that lifts average basket size is a revenue engine, not just a support cost saver.
Eight ‘agentic judges’ as the safety layer
The most distinctive part of the build is a home-grown control layer. Woolworths constructed eight proprietary “agentic judges” that sit behind Olive and vet every response before a customer sees it.
Sivasubramaniyam described three. A “number cruncher” recalculates every mathematical claim in a reply, from unit prices and ingredient quantities to serving sizes and savings estimates. A “product detective” checks each response against legal, food-safety and compliance rules. A “goal judge” confirms the task was actually completed, so that a request to build a dinner basket under $20 is rejected if the agent lands at $25.
This addresses the two failure modes that make large enterprises nervous about agentic AI: confident arithmetic errors and unsafe or non-compliant claims about food. Rather than trusting a single model to police itself, Woolworths added independent checkers between the model and the customer. It is a pragmatic answer to a governance problem, and one built in-house rather than bought off the shelf.
Why it matters for Australia
Woolworths is not moving alone. At Google Cloud’s Sydney summit in late June, the company sat alongside Bunnings, Transurban, Macquarie and the Hawthorn Football Club as local proof points for agentic AI. Bunnings built its “Buddy” shopping agent on the same Gemini platform in about six weeks and, per Google Cloud, more than doubled online conversion. Google Cloud’s vice president for Australia and New Zealand, Paul Migliorini, framed the shift as moving from tools that wait for a prompt to ones that “take action on your behalf.”
The stakes are concrete. Woolworths and Bunnings reach millions of Australian shoppers each week, so an assistant that reads photos, applies loyalty discounts and completes checkout touches ordinary consumers directly, not just enterprise IT teams. It also lands amid live scrutiny of grocery pricing and rewards programs, which raises the bar on the accuracy of any automated savings claim. The “number cruncher” and “product detective” judges read less like engineering flourishes and more like risk controls for a sector under the microscope.
Woolworths’ judges also offer a template. Google supplies auto-scoring and quality-assurance features inside Gemini Enterprise, but Woolworths chose to add its own vetting layer on top. For other large Australian organisations weighing agentic deployments in banking, insurance or government, the lesson is that responsible scale may depend less on the base model and more on the checks a company wraps around it.
The next test is the consumer rollout now underway. Staff have used Olive in a controlled setting; putting an agent that can spend money in front of the public is a harder trial of both the technology and the safeguards. If the judges hold up in the wild, expect more Australian retailers to follow Woolworths and Bunnings onto agentic platforms before Christmas. If they slip, the same deployment will become the cautionary case study.
Sources: iTnews, Google Cloud Press Corner, Computer Weekly, iTWire, iTnews (Bunnings).
