The Agentic Commerce Shift: What AI Shopping Agents Mean for How Australian Businesses Sell Online

In January 2026, Mastercard completed what it called Australia’s first authenticated agentic transaction: an AI agent, running on the Sydney built Maincode model and authorised through a Commonwealth Bank debit card, bought movie tickets from Event Cinemas without a person clicking a single button on a checkout page. A month later, eBay updated its user agreement to explicitly ban “buy for me” agents, LLM driven bots, and any end to end flow that places an order without human review. Both companies are responding rationally to the same shift. They have simply reached opposite conclusions about what to do with it.

That contradiction is the actual state of agentic commerce in September 2026, not a settled outcome business leaders can wait out. Payment networks, banks, and a subset of retailers are actively building the rails for AI agents to browse, select, and pay on a customer’s behalf. A larger group of merchants is blocking them at the door, citing data scraping, fraud exposure, and disputes over who is liable when a bot buys the wrong size. Both positions will keep existing at once for the next several years, and a business’s website, product data, and checkout flow now sit at the centre of that argument whether its owner has thought about it or not.

Three Protocols, One Immature Standard

Three competing technical standards are attempting to define how an AI agent completes a purchase, and none of them has won yet. Google, Shopify, Etsy, Walmart, and Target announced the Universal Commerce Protocol (UCP) in January 2026, publishing product catalogues and checkout capability through a structured manifest file that an agent can read directly from a merchant’s domain. OpenAI and Stripe built the Agentic Commerce Protocol (ACP) to execute payment inside a chat interface using a scoped, single use payment token, so an agent never holds raw card details. Google’s own Agent Payments Protocol (AP2) takes a different approach again, using cryptographically signed authorisation records rather than tokens, and in April 2026 Google handed governance of it to the FIDO Alliance specifically to keep it neutral rather than Google owned, alongside a companion standard built with Mastercard called Verifiable Intent that creates a tamper proof log of what a user actually authorised an agent to do.

These are not mutually exclusive systems competing for the same job. A single transaction can use UCP to assemble a cart, ACP to move the payment, and AP2 to prove the customer actually authorised it, each protocol handling a different layer of the same handshake. That is a reasonable long term architecture, but it also means a business waiting for “the winner” to implement is waiting for something that was never going to happen on that timeline. The near term reality is fragmentation: a handful of major platforms each pushing their own manifest format, and a merchant’s product data needing to be legible to more than one of them at once.

Why the Retailer Response Split

The retailer backlash is not simply resistance to new technology. eBay’s ban, effective 20 February 2026, targets a specific failure mode: an autonomous agent placing an order with no human confirming the size, colour, address, or price before the payment clears, and no clear party to hold accountable when it gets one of those wrong. Amazon ran into a version of the same problem from the other direction, restricting third party data scraping in August 2025 and then facing criticism when its own “Buy for Me” agent was found scraping competitor product pages without the same permission it was denying to others. Security researchers tracking bot traffic across more than 1,800 major websites have found AI agents increasingly treated as unwelcome visitors, blocked at the same layer that has historically stopped price scrapers and inventory bots.

OpenAI’s own retreat from in-chat checkout supports the caution. The company launched Instant Checkout inside ChatGPT in September 2025 and scaled it back within six months, redirecting purchases to a retailer’s own website or app rather than completing them natively. CNBC’s reporting on the change points to the underlying problem plainly: product data across the web is too inconsistent and unstructured for an agent to complete a transaction reliably without a human somewhere in the loop to catch an error before it becomes a dispute. For a retailer, that is not an argument against agentic commerce. It is an argument that discovery and conversation can safely happen inside an AI interface today, while final checkout mostly still belongs on a website the merchant controls, at least until product data and authorisation standards mature.

Why Momentum Continues Regardless

None of that caution has slowed the institutions building the payment rails. A PayPal survey found 48 percent of Australians already use an AI assistant somewhere in their shopping search, and Adobe has tracked triple digit growth in traffic to retail sites referred from generative AI tools over the past two years. Mastercard is not treating the Event Cinemas transaction as a proof of concept it might revisit later. The company has confirmed Westpac as a second Australian banking partner, is establishing a regional AI centre of excellence, and is opening a dedicated agentic commerce team out of its Sydney technology hub to work directly with local financial institutions and merchants. Mastercard’s own announcement frames Australia as a deliberate early market for the technology, not an afterthought.

Inside Retail’s assessment for the Australian market lands on the same conclusion from the retail side: the infrastructure is being built with or without a given merchant’s participation, and the businesses that wait for consensus on a single standard will spend 2027 retrofitting product data and checkout flows under time pressure, while competitors who prepared incrementally will already have the option live. Blocking agent traffic outright, as eBay has done, is a legitimate short term risk control. It is a much weaker long term strategy for a business that depends on discovery traffic, because the same AI systems doing the shopping are increasingly also the ones doing the searching.

What Actually Changes on a Business’s Own Website

The practical starting point has little to do with picking a protocol and everything to do with the state of a business’s product and service data. UCP, ACP, and AP2 all depend on structured, accurate, machine readable information: current pricing, real stock levels, precise variant and specification data, clear shipping and return terms, and a checkout flow that can report its own state back to whatever is asking. A catalogue that lives correctly formatted in one system and gets copied by hand into a website is already a liability under the current SEO and answer engine environment. Under an agentic commerce model it becomes an active source of failed or disputed transactions, because an agent will act on whatever data it can read, accurate or not.

This is squarely an integration and data architecture problem before it is a marketing one. Product, pricing, and inventory data typically sit in several disconnected systems, a CMS, a point of sale platform, an accounting package, that were never built to expose a single consistent feed to an external agent. Businesses that have already invested in clean API level data integrations between those systems are in a materially stronger position than those relying on manual updates across platforms, because the same connective layer that keeps a CRM and an accounting system in sync is what a checkout manifest needs to stay accurate in real time. A website itself also needs to be built on infrastructure that can expose that data cleanly, which is one reason platform choice and disciplined website architecture are no longer purely a design decision.

The Questions a Board Should Actually Be Asking

For a chief executive or a chief financial officer, the more consequential questions sit above the technical layer. Liability for a transaction an authenticated agent completes on incorrect data the business itself published is not clearly assigned by existing law or contract. Most cyber insurance policies and merchant services agreements were drafted before a non human buyer was a realistic scenario, and a dispute resolution process built around a human cardholder does not automatically extend to a cardholder who authorised an agent rather than the purchase itself. AP2’s move toward Verifiable Intent and FIDO governed authorisation records exists precisely because these accountability questions do not yet have settled answers, and a business accepting agent initiated payments today is operating slightly ahead of the legal and insurance frameworks that will eventually catch up to it.

That is a reason for a considered rollout, not a reason for inaction. A services business or a professional firm with no direct e-commerce checkout is not exposed to the same transactional risk as a retailer, but it faces a related exposure: whether an AI agent researching vendors on a client’s behalf can find accurate information about pricing structures, service scope, and availability, and whether that information is structured well enough to be trusted rather than skipped in favour of a competitor whose data reads more cleanly.

A Practical Framework for the Next Two Quarters

Committing to a single protocol this year is premature for most businesses. Building the underlying discipline that any of them will eventually require is not. Five priorities are worth acting on now:

• Audit product and service data for structural accuracy, not just SEO completeness. Pricing, availability, specifications, and terms need to be correct at the source system and propagated automatically, not maintained by hand across the website, the accounting platform, and the point of sale system separately.
• Treat API integration as commerce infrastructure, not back office convenience. The connective layer between a CMS, inventory system, and payment processor is what will eventually feed any checkout manifest a protocol requires, and it is worth building correctly regardless of which standard prevails.
• Decide deliberately whether to allow, restrict, or block agent traffic, and document why. A blanket ban and an open door are both defensible positions depending on transaction risk and dispute exposure. An undocumented default, arrived at by accident, is not.
• Confirm what existing insurance, merchant agreements, and dispute processes actually say about a non human buyer. Most were written before this question existed and may not answer it the way a business assumes.
• Watch the standards without betting the business on one. UCP, ACP, and AP2 are converging toward complementary roles rather than a single winner, and the safest technical position is a data and integration layer clean enough to plug into whichever manifest format ends up mattering most in a given customer’s market.

Where This Leaves Business Leaders

Agentic commerce is not a hypothetical feature sitting a few years out. It has already processed a real payment on Australian banking rails, and it has already been banned outright by one of the internet’s largest marketplaces, sometimes for the same underlying reason: the industry has not finished deciding who is accountable when a machine makes a purchase decision on a person’s behalf. That disagreement will take years to resolve fully, and no business needs to resolve it alone or first.

What a business can control now is whether its own data, systems, and checkout infrastructure are clean enough to plug into whichever version of this future arrives in its market and its category. A digital project consultancy engagement that starts with an honest audit of product data, system integration, and platform architecture is a defensible use of a technology budget in 2026 regardless of which protocol eventually leads. Waiting for the standards war to resolve before doing that groundwork is not a neutral choice. It is a bet that a business can move fast enough later to catch up to competitors who started building the foundation now.