AI shopping assistants explained for Australian retailers

Last reviewed: 13 October 2026.

What an AI shopping assistant is

An AI shopping assistant is a conversational agent on a retailer's own website or app: the customer describes what they need in plain language and gets products, answers or a filled basket back. It is different from being listed on a third-party AI agent such as ChatGPT or Google's AI Mode, where someone else's assistant reads the retailer's catalogue and the retailer does not control the conversation.

How it works

Most current assistants pair a large language model with the retailer's own systems: the product catalogue, search, price and stock services and, increasingly, the cart. The model interprets a request such as "a waterproof jacket for a nine-year-old under $80", calls those systems and writes the answer. Newer versions are described as agentic: they take multi-step actions, such as turning a typed or photographed shopping list into a cart, usually with the customer approving each item before checkout.

Retailers are now running both models at once. Gap Inc. is a recent example. On 5 October 2026 it announced conversational discovery for Old Navy and Banana Republic, where customers describe their needs in natural language, while also making products available on two third-party discovery services, Alta Daily and Daydream. Chief technology officer Sven Gerjets told Retail Dive that "AI should make shopping easier and more intuitive, not add complexity or take choice away from the customer." Gap has published no performance figures and has not named the technology behind its own agents.

The trade-off is control against reach. An owned assistant keeps the conversation, the customer data and the basket on the retailer's site. A third-party listing reaches customers who start somewhere else, on terms the platform sets.

Who has deployed it in Australia

• Woolworths, Olive. Woolworths is rebuilding its long-running Olive assistant on Google Cloud's Gemini Enterprise for Customer Experience, announced in January 2026 (AdNews). Its FY26 results (26 August 2026) list a "transformed digital shopping assistant, Olive, leveraging agentic AI". iTWire reported the rollout to app shoppers through July 2026, with customers approving every item before checkout.
• Bunnings, Buddy. Bunnings announced Buddy on 4 April 2026 for bunnings.com.au, replacing Ask Bunnings AI. Customers type or photograph a list and Buddy adds items to the cart after checking local price and stock; checkout stays on the Bunnings site. The rollout began with some web visitors, with voice planned for later in 2026.
• Kmart, Joy. Kmart's own page describes Ask Joy, which sits in the website search bar and is promoted in the app. The page says Joy cannot place orders, manage returns, access accounts or check real-time in-store stock.
• Coles. Coles' FY26 results describe giving store support team members access to ChatGPT Enterprise. No customer-facing shopping assistant had been announced at the last-reviewed date.

On the third-party side, Inside Retail reported in June 2026 that retailers including Bunnings, Kogan, Adore Beauty and The Iconic were named partners as Google launched agentic checkout in Australia, allowing single-item purchases inside Search and the Gemini app. Amazon's Rufus assistant had not been confirmed for Australia at the last-reviewed date.

What it costs

Pricing is usually per conversation or per outcome rather than per seat. Published list prices, in US dollars, at the last-reviewed date:

• Google's CX Agent Studio, the platform behind the Woolworths, Bunnings and Kmart assistants, lists $0.50 per chat session, or $0.45 to $0.40 on one- or three-year plans. A new session is billed every 50 turns or after 30 minutes of inactivity.
• Salesforce lists Agentforce at $2 per conversation, or through Flex Credits at $500 per 100,000 credits.
• Intercom prices its Fin agent from $0.99 per resolved outcome, plus seats.

As an illustration only: iTnews reported Olive handles roughly 500,000 conversations a month. At a $0.50 list rate that volume would cost about US$250,000 a month before any discount. The language model underneath is a small share of that; Google's Gemini API lists its lightest models at under US$0.30 per million input tokens.

The larger costs are rarely on a price list: connecting the assistant to catalogue, price, stock and cart services; cleaning product data so it can answer attribute questions; writing and testing guardrails; and a standing team to review conversations. iTnews reported that Woolworths' rebuild took about six months.

Where it breaks

Wrong answers the retailer owns. In Moffatt v Air Canada (2024), a Canadian tribunal held the airline liable for its chatbot's misstatement and called the argument that the chatbot was a separate legal entity "a remarkable submission". In Australia, the ACCC has said businesses cannot sidestep their Australian Consumer Law obligations by blaming an algorithm, Inside Retail reports. Olive itself drew attention in March 2026 for off-script remarks and pricing errors; Woolworths said the remarks came from old pre-written scripts and removed them (University of Sydney).

Manipulation. Prompt injection, where a user or a piece of content steers the model away from its instructions, is the first risk on the OWASP list for language-model applications. Any assistant that can change a cart or quote a price needs limits the model cannot talk its way around.

Product data. The assistant is only as good as the catalogue it reads. Missing attributes, stale prices and stock data that is not real-time produce confident wrong answers; Kmart's page says Joy cannot see live store stock, for example.

Low use. A Bizrate Insights survey for EMARKETER (May 2026) found 38.2 per cent of US digital shoppers had never used a retail AI chatbot and were not interested, the largest single response.

Measurement. The published results compare users with non-users. Amazon said Rufus users are 60 per cent more likely to complete a purchase (Fortune); Walmart said Sparky users' order values are about 35 per cent higher (Modern Retail). Customers who choose to use an assistant are often already more engaged, and none of these companies published a methodology. A holdout test is the only clean measure.

Privacy. The OAIC's guidance on commercially available AI products says chatbots should be clearly identified as such, limits secondary use of personal information, and notes that AI outputs, including incorrect ones, can be personal information.

The wrong tool. For a small range, or where customers mostly ask about orders and returns rather than products, better search and a service bot may do the job at lower cost and risk.

Questions to ask a vendor

• Which of our systems will the assistant read from and write to, and how fresh is the price and stock data it sees?
• What can it do without the customer's confirmation, and who sets those limits?
• How are wrong answers detected, logged and corrected, and who reviews conversations?
• How is the price calculated, what counts as a billable conversation, and what happens to cost at peak?
• Where is conversation data stored, is it used to train any model, and how do we meet the OAIC guidance?
• How will we measure incremental sales rather than comparing users with non-users?
• How does this sit alongside listing our catalogue on third-party agents?