The Hon Andrew Leigh MP
Assistant Minister for Productivity, Competition, Charities and Treasury
Competition Policy for Invisible Agents
Economic Society of Australia (NSW Branch) Conference on ‘The Invisible Mind – The Economics of AI’,
Sydney
Wednesday, 17 June 2026
I acknowledge the Gadigal people of the Eora Nation, on whose lands we meet today. I pay respect to their elders, and all First Nations people present. Thanks to Professor David Orsmond and your Economic Society of Australia (NSW Branch) colleagues for organising today’s conference.
The title of today’s conference is The Invisible Mind, which is wonderfully evocative. It sounds like a lost Hitchcock film, or perhaps the name of a popular philosophy podcast.
Since Adam Smith, economists have talked about invisible hands. Today I want to talk about invisible agents.
An invisible agent is an AI system that acts for someone in a market. It may search, compare, recommend and buy. It may set prices, design bids, screen suppliers and negotiate contracts. It may work for a consumer, a firm, a platform or a payment provider. Often it acts at machine speed, across many transactions, with a human setting broad instructions rather than watching each step.
Economics has always cared about agency. We ask who has information. We ask who has incentives. We ask who bears risk. We ask who captures the surplus.
AI agents place a mysterious intermediary into the mix.
When a person walks into a shop, we know what roles each person is playing. The buyer wants a good product at a good price. The seller wants a sale. The shop assistant may be helpful, although over-enthusiastic about the extended warranty if they get a bonus for selling it.
In an AI-mediated market, the picture becomes murkier. A shopping agent may say it is helping the consumer. Yet it may be trained by a platform, paid by a merchant, optimised for engagement or tuned to protect an ecosystem.
So the central question is simple: when an AI agent acts, whose interests does it serve?
Three years ago, I spoke at the McKell Institute in Sydney about artificial intelligence and competition policy. The argument had two sides.
The optimistic side was that AI could make markets more competitive. A startup could use AI to code a website, produce marketing material or improve customer service. A challenger firm could move faster and put pressure on incumbents.
The cautionary side was that the market for foundation models might itself become concentrated. I identified five risks.
First, costly chips. Frontier AI requires extraordinary computing power. That gives an advantage to companies with access to advanced semiconductors, cloud infrastructure and the capital to train large models. NVIDIA’s stock has grown tenfold in the past three years. The ‘magnificent seven’ have spent hundreds of billions on data centres, with Morgan Stanley projecting a further US$3 trillion in global data centre spending through 2028.
Second, private data. The best models depend on vast quantities of high-quality data. Firms with large proprietary datasets have an advantage, while publishers, platforms and content owners are increasingly bargaining over access. Reddit’s licensing deal with Google, reportedly worth around US$60 million a year, showed how training data has become a strategic input.
Third, network effects. AI systems may improve with use. More users can mean more feedback, better tuning and stronger distribution. That creates a familiar digital-market dynamic: the leading model attracts more users, and those users may help the leading model pull further ahead. On one estimate, ChatGPT has three-quarters of the AI chatbot market. That provides a vast amount of feedback with which to improve its model.
Fourth, immobile talent. Building frontier models requires scarce engineering and research skill. If the best people are locked inside a small number of firms, or pulled into the orbit of cloud-model partnerships, challengers face a steeper climb. The Australia Competition and Consumer Commission’s December 2025 AI industry snapshot notes media reports estimating that the specialised expertise needed to develop frontier models may be limited to several hundred people globally, with intense competition for that talent helping to drive licensing arrangements and ‘acquihires’ between large digital platforms and AI start-ups.
Fifth, an ‘open-first, closed-later’ model. A firm may begin by encouraging developers, customers and complementors to build on its system, then later restrict access once dependence has grown. That pattern has appeared before in digital markets. In foundation models, it could arise through APIs, plugins, model marketplaces or agent ecosystems. The more your favourite AI model knows about you, the more useful it becomes – and the more reluctant you may be to switch to a rival.
Those five competition risks remain: chips, data, network effects, talent and ecosystems that may narrow over time. My concern is that foundation models may follow the path of internet search: a crowded field in the late 1990s, a dominant monopolist now.
But today I want to focus on another dimension of competition in the AI economy. My earlier analysis asked whether foundation-model markets would be competitive. The next question is whether AI agents will make other markets more competitive.
For most of the internet era, online shopping still involved a human in the loop. We typed a query. We clicked a link. We compared options. We cursed the pop-up asking us to subscribe to a newsletter. We abandoned the cart. We came back three days later after being chased around the internet by mattress ads.
Agentic commerce changes the sequence.
A consumer can ask for ‘a birthday present for a ten-year-old who likes science and hates tidying her room’. The agent can search, compare, check reviews and purchase.
Agents are already in the market. Microsoft’s Copilot Checkout is beginning to let United States users complete purchases directly within Copilot, with PayPal, Shopify and Stripe integrations. Mastercard is developing Agent Pay, built around registered agents, verification, tokenisation and consumer-set controls. Visa is deploying Intelligent Commerce, designed to let approved AI agents find and buy on behalf of users within spending limits.
This is commerce moving from browsing to delegation.
That could be valuable, reducing search costs and saving time. A small merchant in Wagga Wagga, Wollongong or Warrnambool may gain customers if an agent can read product data better than a human can navigate a website.
It could also create a new and unpredictable gatekeeper.
For two decades, firms worried about search engine optimisation. The next contest may be agent optimisation. The commercial question becomes: how do I make my product legible to the machine that chooses on behalf of the human?
A human can see when the cereal is at eye level on a supermarket shelf. A human can see a sponsored link, at least when site is playing fair. A human can suspect that a hotel website saying ‘two rooms left’ has discovered behavioural psychology and developed a theatrical streak.
With AI agents, steering can become harder to see.
A product might be chosen because it is better. It might be chosen because the merchant has cleaner data. It might be chosen because checkout is smoother. It might be chosen because the platform has a commercial relationship with the seller.
Better is good. Cleaner and smoother will tend to tilt the market against small players. Undisclosed kickbacks are a serious problem.
Here are four examples of agentic misbehaviour that competition policy should take seriously.
First, hidden steering.
A consumer asks an agent for running shoes under $100. The agent favours products that support a preferred checkout protocol, because the purchase can be completed in-chat. The consumer sees convenience. Rivals see a tollbooth. A small shop with a better product but a clunkier integration slides from ‘best option’ to ‘buried option’.
Supermarkets have shelf space. Search engines have rankings. AI agents have recommendations that can feel simple, personal, trusted and conversational.
A paid placement in a list looks like advertising. A paid preference inside a friendly answer can look like advice. That’s why your financial adviser has to tell you if they’re getting a commission. For AI agents, the task is to ensure that existing consumer protections keep working in a market where influence is harder to see and easier to personalise.
Second, personalised extraction.
Suppose a seller-side agent works out that one customer is calmly researching a fridge, while another needs one today because theirs failed overnight. The price changes accordingly. The market moves from ‘what is the price?’ to ‘what is my price?’
A common price gives consumers a benchmark. It lets people compare. It makes unfairness visible. A personalised price can turn every buyer into a private auction, and risks less tech-savvy consumers being exploited.
There are benign versions of price discrimination. Student discounts expand access. Off-peak pricing can spread demand. But agentic pricing could become finer, faster, less visible and more problematic than anything consumers have previously faced.
To deploy personalised pricing, the new system may infer urgency from calendar entries, browser behaviour, location data and the fact that your agent has asked repeatedly about fridges.
Third, agent gaming.
Imagine a buyer-side agent trained to rank products using prices, product descriptions, reviews and delivery terms. A seller-side agent studies its behaviour. It learns that certain phrases, image styles, checkout paths or warranty wording move products up the ranking. It rewrites the listing to please the buyer’s agent.
The consumer thinks the agent has found the best product. In reality, two machines have fought a tiny battle over the ordering of a list.
This already echoes the history of search engines. Search created search-engine optimisation. Social media created engagement optimisation. Agentic commerce will create agent optimisation.
Some of that will be harmless. Clean product data helps buyers. Accurate inventories help markets. But optimisation can slide into manipulation. If the buyer’s agent can be gamed, the consumer’s delegated choice may become another advertising surface.
Fourth, machine coordination.
Competition law has always had a mental picture of collusion. Men in suits in a smoke-filled room. A quiet conversation on a golf course. A spreadsheet passed around with market shares and a wink. As Adam Smith wrote 250 years ago, ‘People of the same trade seldom meet together, even for merriment and diversion, but the conversation ends in a conspiracy against the public.’
Modern cartels are rarely that cinematic. AI makes the picture truly weird.
What if sellers use pricing agents that learn from the same data? What if the agents discover that price wars are painful? What if they settle into a pattern of higher prices without a meeting, an email, a phone call or a lunch at which anyone says the forbidden words?
Recent research suggests that large-language-model pricing agents can reach supracompetitive prices in simulated oligopoly settings. Small changes in prompts can affect the tendency to coordinate.
Agent design, model diversity and market structure can shape outcomes. Different models, different laws, different data and different strategies may serve as grit in the gears of tacit coordination.
In each case, the service feels smooth. The competition problem is under the bonnet.
Simply put, we want our agents to be our agents.
If a system claims to act for the consumer, consumer law should hold it to that representation. If it is acting for a merchant, that should be clear. If commercial arrangements shape what appears, consumers should know.
A travel agent who secretly took payments to steer families into worse hotels would face questions. A digital agent should face the same discipline.
We already have a useful precedent from outside agentic commerce. In 2024, Air Canada was ordered by a Canadian tribunal to compensate a customer who relied on incorrect information from its chatbot about bereavement fares. The airline argued that the chatbot was a separate legal entity. The tribunal treated that argument with the warmth it deserved.
Businesses should answer for the systems they deploy.
The Air Canada principle should apply to agents as it does to chatbots. When you give an agent your credit card number, you should know if it’s getting kickbacks through the sale process.
There are also implications for consumer law.
Last year, Treasury’s Review of AI and the Australian Consumer Law concluded that the Australian Consumer Law is well placed to respond to many of the harms that may arise from AI-enabled goods and services, including agentic systems. That reflects the strength of a law that applies across the economy and is designed to adapt to technological change. It does not need to know whether the sales assistant is human, a chatbot, or an invisible agent operating behind a checkout button.
A misleading statement remains misleading when produced by a model. A hidden fee remains hidden when an agent clicks through it. A fake review remains fake when written in perfect prose. A subscription trap remains a trap when the button is pressed by software.
The Australian Consumer Law already has a range of tools for this world. The prohibition on misleading or deceptive conduct can reach artificial intelligence-generated representations and omissions. The Australian Government’s work with states and territories on unfair trading practices is relevant to hidden steering, manipulative choice architecture, personalised pressure and other forms of digital manipulation. The unfair contract terms framework may also have work to do where standard-form terms govern how an agent uses data, ranks products, delegates authority or allocates risk. The Treasury Review’s message was reassuring without being complacent: Australia starts with a sturdy kit of tools, and government must keep watching the technology as it moves.
The enforcement challenge is observation. In the past, a regulator could compare advertised prices or inspect terms or review documents. In agentic markets, conduct may be personalised and hard to reproduce.
Regulators may need to become machine shoppers.
A competition agency could run controlled tests using synthetic consumers. Ask the same market for the same product at the same time, varying the buyer’s inferred income, urgency or loyalty. Observe whether price, ranking or disclosure shifts. Test whether affiliated products climb the ladder. Check whether a merchant using the preferred protocol receives extra prominence.
Mystery shopping has long been part of consumer protection. The mystery shopper of the future may be a fleet of test agents. In this, our Government’s new AI Safety Institute might work with the Australian Competition and Consumer Commission, providing technical expertise on AI models.
Competition agencies will also need audit capacity. If a firm deploys a pricing agent, it should keep records of instructions, data inputs, model changes and pricing outcomes. If a platform deploys a shopping agent, it should be able to explain ranking factors to regulators. If a payment system registers agents, it should preserve a trail of mandate and authorisation.
A firm that can log every click for advertising can keep records when an algorithm sets prices for consumers.
Where might the problem arise? Invisible agents depend on visible infrastructure: chips, data centres, cloud contracts and payment rails. Market power can live at any of those layers.
At the compute layer, scarce chips and specialised infrastructure shape who can train models and who can deploy agents at scale.
At the cloud layer, strategic partnerships can give one firm preferential access, stronger information and lower switching costs. The Federal Trade Commission’s AI partnership study described cloud commitments worth billions of dollars, discounted computing resources, sensitive information flows and constraints affecting the ability of AI developers to work with other cloud providers.
At the model layer, default integration can count as much as raw capability. A slightly weaker model built into every browser may beat a better model that users must seek out.
At the commerce layer, protocols, wallets and checkout systems can become the rails on which agents move. Open standards can help. Captured standards can harm. The key governance questions are: who writes the rules, who pays to connect, who can appeal exclusion and who sees the data?
This is where Australia has a specific interest.
We are a user economy in AI more than a frontier-model economy. Our households, firms, public services and charities will use models built elsewhere, often through platforms whose centre of gravity sits overseas. That makes contestable access crucial.
A small Australian retailer should have a fair shot at being found by an agent. A local software firm should have a fair shot at plugging into the stack. An Australian consumer should know when a recommendation reflects their interests, rather than a commercial arrangement buried in the machinery.
Australia’s new merger regime is relevant here. Some AI competition issues will arise through acquisitions. Others will arise through partnerships, exclusive access, compute commitments and talent arrangements that sit near the edge of merger law.
There is a conduct agenda too: scrutinise self-preferencing, tying, bundling, exclusionary access and discriminatory terms. In agentic markets, one verb deserves special attention: steer.
Can a cloud provider steer customers to its own model? Can a platform steer buyers to affiliated merchants? Can an assistant steer users away from rivals? Can a payment provider steer transactions towards preferred rails?
Good steering helps people find what they want. Anticompetitive steering hides better options or operates on a pay-to-play basis.
So what should competition policy ask of invisible agents?
I suggest four tests.
First, loyalty. When firms make AI agents that claim to act for users, system incentives should be clear. Payments, commissions, affiliations and protocol advantages should be disclosed in a way humans can grasp and machines can read.
Second, contestability. Agent-mediated markets should remain open to rivals on fair terms. A business should avoid invisibility merely because it lacks the favoured integration. A rival model should have a path to users where it competes on quality.
Third, verifiability. Regulators and consumers need ways to test agent behaviour. That means audit trails, data retention, controlled experiments, reproducible tests and access to the information needed to separate convenience from foreclosure.
Fourth, responsibility. Firms should answer for the agents they deploy. More autonomy should bring stronger governance. More sensitive decisions should bring clearer records. As our Government’s National AI Plan makes clear, even as AI agents become more widespread, human oversight remains important.
These tests are straightforward and based on existing consumer law principles. They avoid the fantasy of predicting every future use of AI. They ask whether the system preserves rivalry and informed choice, backed by accountability.
They also avoid the common policy mistakes of complacency and overreach. The error of complacency is to assume that once digital markets are consolidated and gatekeepers are charging a toll, it’s too late to do anything.
The other error is overreach. Compliance systems can become moats when incumbent-sized firms can afford them and challengers cannot. A rule written in the name of safety can become a gift to incumbency if it ignores competition. Since George Stigler, economists have warned about large firms that favour complex regulations because they know that their smaller rivals will find it harder to comply.
The better path is targeted, empirical, evidence-based and competition-aware. Use existing laws where they fit. That includes the Australian Consumer Law’s prohibitions on misleading or deceptive conduct and false or misleading representations, the unfair contract terms framework and the proposed unfair trading practices prohibition. Strengthen tools where gaps are clear. Keep monitoring technological developments so the Australian Consumer Law and competition laws remain fit for purpose. Build technical capability inside regulators. Work internationally, because an artificial intelligence agent selling in Sydney may be trained in San Francisco, hosted in Virginia and paid through a global payment network.
Let me finish with the consumer.
Economists sometimes talk about consumers as if they are calculating machines. Real people are busy, tired, hopeful and occasionally trying to buy concert tickets while making dinner.
AI agents could help them. They could find cheaper products, compare complex offers and save people from paperwork. They could help a person with limited English deal with a bank. They could help a time-poor parent find the safest product at the best price.
That is the promise.
The risk is that agents become a digital version of the charming salesperson who is getting a cash bonus for selling you an extended warranty. An easy-to-use interface can make steering feel like service. A seamless checkout can make a choice feel already made.
Competition policy for invisible agents should aim for this goal: when technology acts for people, markets should work for people.
The invisible hand was never magic. It depended on rules, institutions, rivalry and trust.
The invisible agent will too.
The future of AI in markets will be shaped by who controls the gateways, who writes the protocols, who sees the data and who bears responsibility when agents act.
So as AI moves from answering questions to making choices, competition policy needs to move with it.
We should welcome the efficiency. We should use the tools. We should keep the gateways open and the agents honest.
Because when agents do the choosing, competition policy must keep asking: who is the customer, and who is being sold?
Acknowledgements
My thanks to several people, including officials from the Australian Treasury, for valuable comments on earlier drafts.
ENDS