Consumer AI has mainly helped people search, compare and decide. It is now starting to choose and pay. These agents could save us time, but they could also steer us towards the firms that fund or control them. Any service acting for consumers should owe its users loyalty, show its work and remain under their control.

In January 2026, an AI agent used a Commonwealth Bank debit card to buy cinema tickets, and a Westpac credit card to book a stay in Thredbo. Mastercard called these Australia's first fully authenticated agentic transactions. In each case, software completed the payment.

They were controlled trials, but they show AI moving from helping with purchases to completing payments. This is agentic commerce. J.P. Morgan estimates that this form of commerce could reach US$2 trillion to US$5 trillion by 2030, equal to 3 to 7 per cent of global payment volumes.

An agent may watch airfares, compare power plans, reorder groceries and pay within limits set by its user. A good agent could read hundreds of offers, catch hidden fees and switch providers in seconds. That could give ordinary people some of the search power large firms already have. It could also give the firms behind the agent new power to steer consumer choices.

Picture asking an agent to feed a family of four for a week on $180, avoiding peanuts and favouring Australian-made products when the price is close. Minutes later, a full basket appears for $176. But did it check more than one supermarket? Did it favour products whose sellers paid for placement, or products its software found easier to buy? Did it treat "Australian-made" as an instruction or a vague hint? Once the agent pays, its interpretation becomes a choice made with your money.

From selling attention to selling choice

An agent may look like a personal assistant even when it sits inside a system built by a retailer, bank, payment provider or marketplace. The firm may benefit when the agent picks a partner or completes a sale. Commercial funding is not itself the problem. The problem begins when someone else's interest overrides the user's instructions.

The last generation of platforms learned to sell our attention. This one may learn to sell the choices we hand over. Repeated across millions of purchases, small tilts can shape which products get seen and which firms reach customers. At scale, agents can redirect demand as well as respond to it.

Australia has already seen how commercial incentives can distort digital rankings. In 2022, the Federal Court ordered Trivago to pay $44.7 million for misleading consumers about hotel rates. Its ranking gave significant weight to what booking sites paid it. Trivago's own data showed a higher-priced offer taking the top spot over a cheaper one in 66.8 per cent of listings. Trivago's algorithm decided which offer consumers saw first. It sold attention but left the final choice to the consumer. An agent could go further and make the purchase for them.

Agents could also reshape competition among sellers. As Assistant Minister Andrew Leigh has argued, firms may soon optimise for agents as well as search engines. A product may win because it is better, or because the merchant has cleaner data and a checkout the agent can finish in one step. A small Australian retailer may offer the better product and still never be considered by the agent.

Business models are still forming. OpenAI says its shopping results are organic and unsponsored, and that merchant fees do not affect rankings. Google is piloting sponsored Direct Offers inside AI Mode. We should not assume a friendly answer is a neutral search of the market.

From disclosure to duty

Leigh also set out four tests for what he called invisible agents: loyalty, contestability, verifiability and responsibility. His loyalty test asks firms to disclose the incentives shaping an agent. Australia's financial advice reforms suggest that disclosure should be only the first step.

Shopping agents and financial advisers deal with different products and stakes. Both can appear to act for a person while being paid to favour a product.

Under the Future of Financial Advice reforms, an adviser giving personal advice to a retail client must act in that client's best interests and, where interests conflict, give priority to the client. The reforms also banned many commissions and volume-based payments relating to retail investment products.

Disclosure tells a person that a conflict exists; a duty changes what the provider may do. That distinction matters when an AI searches a market, narrows the options to one and can spend the user's money. The steering may be hard to see.

The Australian Consumer Law already bans misleading or deceptive conduct. A Treasury review found that the Australian Consumer Law, together with other laws, can generally deal with harms from AI goods and services. From 1 July 2027, a new ban on unfair trading practices can catch conduct that manipulates consumers or unreasonably distorts their choices, if it causes or is likely to cause harm. Failing to disclose key facts can be one example. Even so, the new law does not require a consumer agent to put the user's goals first, as a loyalty duty would. The government has also named consumer protection in agentic commerce as an AI safety priority.

Hidden steering is wrong, but it can be hard to detect and prove. An agent's answer is personal and hard to reproduce. Rejected options may disappear, and a provider can blame the model or merchant. A right without a record is hard to enforce. A record without someone responsible is of little use.

Australia should guarantee three things.

One: know whose side it is on

Every AI shopping service should state its role before it recommends, chooses or buys. A seller assistant works for a retailer or brand. A marketplace assistant searches inside one platform or group of partners and should say where that search stops.

A consumer agent claims or appears to act for the shopper. That role should carry a higher duty. It should take reasonable steps to follow the user's stated goals and consider a fair range within its stated scope. It should not allow merchant payments or internal targets to override what the user asked for.

Disclosure is the floor, not the duty. Disclosing a conflict in a footer does not make it fair for that conflict to shape the choice. Loyalty does not mean finding one perfect deal. The best choice depends on the person: one shopper wants the lowest price, while another values privacy or local production. The duty is to follow the user's goals through an honest process.

Two: know what it did and who answers

A normal receipt tells us what we bought and what it cost. An agent receipt should also state:

• what the user asked for
• what part of the market it searched, and the main limits
• the main reasons for its choice
• whether a payment, partnership or ownership link shaped the options or their order
• what authority it used, and whether the user confirmed the purchase

Providers would not need to publish their source code. Consumers need a short account in plain language, while providers should keep a fuller record for regulators. The Consumer Data Right offers a precedent for standard digital records: accredited providers must identify themselves, disclose sponsor arrangements and give consumers a consent receipt.

The firm offering the agent should answer for what the agent does, rather than treat the model as a separate party when something goes wrong. If an agent acts outside the user's authority, consumers need a simple way to stop or unwind the transaction and put the problem right.

Three: keep the power to stop and leave

People should control how much power they hand over. An agent might have standing approval to reorder groceries below a set amount, but it should ask again before starting a subscription or making an expensive purchase. Users should also be able to pause it, change its instructions and withdraw its authority.

People must also be able to leave. At their request, consumers should be able to move their preferences, standing instructions and purchase history to another agent. The Consumer Data Right offers a starting point because it already lets people share banking and energy data with accredited services.

Keep the rules narrow

Safety rules can become barriers that only the largest firms can afford. Open standards would help smaller retailers reach agents without paying each large platform for access.

The higher duty should apply when a service has meaningful power to select, recommend or buy for a consumer and a reasonable consumer would believe it acts for them. A change of label should not remove the duty. Seller and marketplace agents could still advertise and earn commissions, subject to existing consumer law, but they should not present themselves as independent consumer agents.

A consumer agent could accept some merchant payments, but not those that can reasonably be expected to shape its choice. The more power an agent has to act without approval, the stronger its records and controls should be.

Loyalty must not become a luxury

Protection and access must be designed together. The Quality of Advice Review noted that the average cost of advice put it beyond the reach of many Australians. The review recorded several barriers, including a falling number of advisers, professional standards and compliance costs. Removing a bad funding model does not automatically make good advice affordable.

AI services can be cheap to scale. Access may therefore depend less on price than on which agent is built into the services people already use. A loyal agent may cost a little. A conflicted one may come free with a bank app, phone or shopping platform. The easiest agent to reach may not be the one that best represents the user.

Private services may provide loyal agents at low cost. But in markets that are hard to navigate and where staying put can be costly, at least one trustworthy option should be easy to reach. Australia should develop public-interest and cooperative options alongside the duty.

That need not mean one vast government chatbot. Public-interest services could start in markets where households already struggle to compare offers. The Australian Energy Regulator's Energy Made Easy service is free, independent and compares offers from every retailer in the markets it covers.

In December 2025, the ACCC found that households on the same electricity plan for more than three years were paying $221 more a year on average than customers on new plans. Across New South Wales, south-east Queensland, South Australia and Victoria, nearly 2.5 million customers were paying prices at or above the relevant default offer.

A public-interest energy agent could use household data, with consent, to check plans, explain possible savings and prepare a switch for approval. Public bodies, consumer groups, cooperatives and private firms could build such services to shared standards.

A trustworthy option could set a standard for other providers without replacing them.

Australia should welcome agents that save time and find better deals, but their benefits will not be shared fairly by accident. A seller assistant should say whose side it is on. A consumer agent should follow the user's instructions, show its work, stay under the user's control and make it easy to switch.

Trustworthy representation should not depend on wealth, technical skill or which platform a person happens to use. The technology is new, but the principle is not: power exercised on our behalf must answer to us.

Blaise Fox, September 2026

Blaise Fox studies economics and philosophy at the University of Sydney and writes about technology, political economy and public policy.