When AI Becomes the Consumer: The Economics of Agentic Commerce

The shift from interface to agent

For thirty years, digital commerce has been organized around a single assumption: a human being sits at the end of every transaction. Search engines rank results for human eyes. Product pages are designed to persuade human attention. Checkout flows are built around human hesitation, and fraud systems are tuned to human behaviour. Even the economics of online advertising rest on the scarcity of human attention.

Agentic commerce breaks that assumption. An AI agent can be told to "find a flight to Accra under $1,400 with one stop and a refundable fare," "reorder the household basics we usually buy, but switch to whatever is cheapest this week," or "find three suppliers for this component and negotiate the best delivery terms." The agent searches, compares, shortlists and, increasingly, completes the purchase. The consumer delegates the work and keeps only the decision rights they choose to keep.

This is no longer speculative. In 2025, the payment networks and the major AI platforms began building the infrastructure for it. Mastercard introduced Agent Pay, built on a new class of "agentic tokens" that let registered, verified agents transact on a consumer's behalf within limits the consumer sets. Visa announced its own agent-ready credentials. OpenAI and Stripe published the Agentic Commerce Protocol to let agents complete purchases inside conversational interfaces, and Google introduced the Agent Payments Protocol (AP2) to standardize how agents prove they are authorized to pay. When the networks that move the world's card payments start issuing credentials to software, the question is no longer whether agents will buy things. It is what happens to markets when they do.

This report argues that agentic commerce is best understood not as a new channel but as a change in who the economic actor is. That shift has five consequences: search costs collapse, brand value is re-priced, intermediary power migrates, payments become the trust layer, and pricing becomes algorithmic on both sides of the market. Each creates winners, losers and strategic choices for incumbents, particularly in financial services.

What actually changes when the buyer is software

The economics of information offers the cleanest lens. George Stigler's foundational insight was that searching is costly, so rational buyers stop searching before they find the best price. Price dispersion persists not because markets are broken but because the marginal cost of one more comparison eventually exceeds its expected benefit. Peter Diamond went further: if every buyer faces even a small search cost, sellers can charge well above competitive prices without losing customers.

Human search costs are mostly cognitive and temporal. Comparing twelve insurance policies, reading fine print and tracking price changes over a week are tedious, so most people don't. An agent's search cost is close to zero. It can query dozens of merchants, parse terms and conditions, monitor prices continuously and apply exactly the criteria it was given.

The internet was supposed to deliver this once before. Early research on "frictionless commerce" found that online prices were lower and more frequently adjusted than offline prices, but dispersion did not disappear. Retailers differentiated on trust, service and brand, and consumers continued to pay premiums to sellers they recognized. The lesson is important: lowering search costs does not automatically produce perfect competition, because buyers care about more than price, and much of what they care about is hard to specify.

That is where agents differ from search engines. A search engine lowers the cost of finding options; a human still evaluates them. An agent lowers the cost of evaluating options as well, but only against criteria that can be articulated. The economics of agentic commerce therefore turn on a single question: which attributes of a purchase can be made explicit enough for software to act on, and which cannot?

Attributes that are measurable, comparable and verifiable, such as price, delivery time, return terms, specifications, carbon intensity and verified ratings, become intensely competitive. Attributes that are ambiguous or experiential, such as taste, status, aesthetic fit and emotional resonance, remain with the human or with whoever shapes the agent's interpretation of them. Everything that follows comes from that division.

Five economic consequences

1. Search costs collapse, but only for what can be specified

Where agents operate, price dispersion on specifiable goods should narrow sharply. Commodity-like categories such as household staples, electronics with standardized specifications, travel, insurance and basic financial products are the most exposed. Merchants that relied on consumer inattention to sustain margins will find that inattention is no longer available to them. An agent does not forget to cancel a subscription, does not overlook a cheaper equivalent and does not accept a renewal price without checking the market.

The less obvious effect is on product design. If agents compare along explicit dimensions, sellers have an incentive to compete on those dimensions and to obscure or bundle the rest. Expect more effort to make offers look superior on the attributes agents rank, and more creativity in fees, conditions and add-ons that are harder to parse. The contest moves from capturing human attention to shaping machine-readable comparisons.

2. Brand moves from persuasion to machine legibility

In human commerce, brand does two jobs. It persuades, by creating emotional preference, and it signals, by reducing uncertainty about quality. Agentic commerce weakens the first and strengthens the second.

Persuasion depends on attention, and agents do not have attention to capture. A beautifully shot campaign does little for software choosing between two equivalent products. Signalling, however, becomes more valuable. When an agent must decide whether a seller is reliable, it needs evidence: verified reviews, return rates, delivery performance, dispute history, certifications and consistent, structured product data. In George Akerlof's terms, agents operating in a market with hidden quality differences will either find credible signals or default to the safest-looking option.

The strategic implication is that brands must become legible to machines. That means structured, accurate product information; verifiable claims; consistent performance data that agents can query; and a reputation that survives algorithmic scrutiny rather than human impression. Brands with strong emotional equity will not disappear. Consumers will still instruct agents to "only buy from brands I already use," but that equity will need to be expressed as an explicit preference, not a moment of persuasion at the shelf.

3. Intermediary power migrates to whoever controls the agent

The most important strategic question in agentic commerce is not who sells, but who sits between the consumer and the market.

Digital marketplaces are classic two-sided platforms. Their power comes from controlling access to buyers, which lets them charge sellers for visibility and set the terms of competition. When a consumer delegates purchasing to an agent, the agent becomes the new point of access. The agent decides which merchants are considered, how options are ranked and which offer is selected. Whoever controls that layer, whether an AI platform, an operating system, a super-app, a bank or a retailer's own assistant, inherits the gatekeeping position that search engines and marketplaces hold today.

This creates three risks that regulators and incumbents should take seriously:

• Self-preferencing. An agent operated by a company that also sells products has an incentive to favour its own inventory or its paying partners. Unlike a ranked list of search results, an agent's choice can be opaque. The consumer may never see the alternatives it rejected.
• Sponsored influence. If merchants can pay to be preferred by an agent, "agent optimization" becomes the new advertising, but with less visibility to the consumer than a labelled ad.
• Concentration. Agents improve with data and scale. A small number of general-purpose agents could end up mediating a large share of consumer spending, creating bottlenecks with significant pricing power over merchants.

The counterweight is that agents can also be built by the consumer's side of the market. A bank, a card issuer or an independent provider could offer an agent whose explicit obligation is to the consumer, much as a fiduciary has obligations to a client. Whether agentic commerce concentrates or distributes power will depend heavily on whether consumers can choose, switch and audit their agents.

4. Payments become the trust layer

When a human clicks "buy," the payment system can rely on a set of assumptions: the person is who they claim to be, they intended to make this purchase and, if something goes wrong, there is a clear party to hold responsible. Agentic transactions strain every one of those assumptions.

Agent-initiated payments raise four questions that the existing system was not designed to answer:

• Identity. Is this agent who it claims to be, and is it acting for the person it claims to represent?
• Authority. Did the consumer actually authorize this purchase, at this price, from this merchant? What were the limits of the mandate?
• Liability. If the agent buys the wrong thing, overpays or is manipulated by a malicious merchant, who absorbs the loss: the consumer, the agent provider, the merchant or the issuer?
• Recourse. How should disputes and chargebacks work when the "customer" who made the decision is software?

This is why the networks moved early. Tokenization, which replaces card numbers with limited-use credentials, already solved a version of this problem for mobile wallets and stored cards. Agentic tokens extend the same logic: a credential bound to a specific agent, carrying the consumer's permissions and limits, and verifiable by the merchant and issuer. The emerging protocols aim to capture the consumer's intent in a verifiable form, so that every agent purchase can be traced back to an explicit, auditable mandate.

The strategic significance is larger than fraud prevention. In agentic commerce, trust becomes infrastructure. The parties that can verify agents, encode consumer mandates and resolve disputes will sit at the centre of the transaction, even if they never see the consumer's screen. For payment networks and issuers, this is an opportunity to move from being the rail underneath commerce to being the authority that makes delegated commerce safe enough to scale.

5. Pricing becomes algorithmic on both sides

Merchants already use algorithms to set prices. Agentic commerce adds algorithms on the buying side. The result is a market in which software negotiates with software, potentially millions of times a day.

Two dynamics are worth watching. The first is personalized pricing. If merchants can identify an agent and infer its principal's willingness to pay, they may try to price discriminate. Agents, in turn, may conceal identifying signals or shop through anonymizing intermediaries. The contest over information about the buyer becomes a contest between algorithms.

The second is algorithmic coordination. Experimental research has shown that pricing algorithms using reinforcement learning can learn to sustain supra-competitive prices without explicit communication, effectively colluding by observing and responding to one another. Buyer agents may counter this by spreading demand and rewarding price cuts, but the outcome is not guaranteed. Competition authorities will need to understand markets where neither side's behaviour is set directly by a human.

The principal-agent problem, revisited

Economists have studied delegation for decades. The principal-agent problem describes what happens when one party acts on behalf of another but has different incentives or better information. Shareholders and managers, patients and doctors, and clients and brokers are the classic examples. The standard tools for managing it are monitoring, incentive alignment and contractual limits.

AI agents create a new version of this problem with an unusual feature: the agent may serve two principals at once. It acts for the consumer, but it is built, operated and monetized by a company. If the provider earns revenue from merchants, the agent's objectives may quietly diverge from the consumer's. Even without bad intent, an agent may optimize the wrong thing, such as the cheapest option when the consumer valued reliability, because preferences were imperfectly specified.

Three design principles follow, and they are as much commercial as ethical:

1. Explicit mandates. Consumers should be able to set and see the rules their agent follows: budgets, preferred and excluded merchants, trade-offs between price and speed, and which decisions require their approval.
2. Auditable choices. An agent should be able to explain why it chose one option over another and show what it rejected. Without this, consumers cannot detect misaligned incentives.
3. Disclosed conflicts. If an agent's ranking is influenced by commercial relationships, that influence should be disclosed in the same way sponsored search results are labelled today.

Agents that meet these standards will earn the trust needed for consumers to delegate higher-value purchases. Agents that do not will be confined to low-stakes, easily reversible tasks.

Who wins and who loses

The effects will not be evenly distributed. The table below summarizes how the main actors are exposed.

Actor	What changes	Strategic response
Consumers	Lower search costs and better prices on specifiable goods; new exposure to opaque agent decisions	Demand control, transparency and the ability to switch agents
Merchants with commodity offers	Margin pressure as agents compare relentlessly	Compete on verifiable reliability, speed and service; make data machine-readable
Brands built on persuasion	Advertising loses reach when the buyer has no attention	Convert emotional equity into explicit consumer preferences and verifiable quality signals
Marketplaces and search platforms	Risk of disintermediation by agents that bypass their interfaces	Become the agent, or become the infrastructure agents rely on
AI platform providers	Opportunity to become the new gatekeeper	Scale fast; manage regulatory risk around self-preferencing
Payment networks and issuers	Agent identity, mandates and liability become central	Own agent verification, tokenized mandates and dispute resolution for delegated transactions
Regulators	Competition and consumer protection frameworks assume human decision-makers	Develop standards for agent transparency, liability and interoperability

What this means for banks and payment networks

Financial institutions are unusually well placed in agentic commerce because the hardest problems are problems they already solve: identity, authorization, risk and dispute resolution. Five priorities stand out.

Treat agent identity as a product. Verifying that an agent is legitimate, and that it represents a specific customer, will become a core service. Institutions that can issue and validate agent credentials at scale will hold a structural position in the transaction.

Encode the mandate, not just the payment. The valuable data in an agentic transaction is not only the amount and the merchant but the instruction behind it: what the customer asked for, the limits they set and the trade-offs they accepted. Capturing that mandate in a verifiable form makes disputes resolvable and fraud easier to detect.

Redesign disputes for delegated decisions. Chargeback rules assume a human made a mistake or was deceived. Agentic disputes will include cases where the agent misread an instruction or was manipulated by a merchant. Clear liability frameworks, agreed across networks, issuers, merchants and agent providers, will determine how quickly consumers are willing to delegate.

Consider offering the agent. Banks hold the richest picture of a customer's financial life and already carry a duty of care. A bank-provided agent that manages bills, subscriptions, everyday purchases and savings, with explicit obligations to the customer, could become a significant differentiator. It would also keep the bank in the decision loop rather than being reduced to the rail underneath someone else's agent.

Instrument risk models for non-human behaviour. Fraud models trained on human patterns will misread agent behaviour, which is faster, more consistent and more frequent. Institutions will need separate behavioural baselines for agents and new signals for detecting compromised or malicious ones.

What to watch

The pace of agentic commerce is uncertain, and the honest position is that adoption will be uneven across categories. Several indicators will show whether it is moving from demonstration to scale:

• The share of transactions carrying agent credentials or agent-specific tokens.
• Whether consumers delegate higher-value and less reversible purchases over time, beyond routine replenishment.
• The emergence of interoperable standards, so agents and merchants are not locked into a single platform's protocol.
• Regulatory action on agent transparency, self-preferencing and liability for agent errors.
• Merchant investment in structured, machine-readable product and service data.
• Whether dispute rates on agent-initiated transactions converge with, or diverge from, human-initiated ones.

Conclusion

Agentic commerce changes the identity of the buyer. When software does the searching, comparing and paying, the scarce resource in commerce shifts from human attention to machine trust. Search costs fall, brands must become legible to machines, gatekeeping power moves to whoever controls the agent, and payments evolve from a settlement mechanism into the layer that makes delegation safe.

None of this guarantees better outcomes for consumers. Agents can intensify competition, or they can concentrate power in a few platforms with opaque incentives. Which path dominates will depend on decisions being made now about standards, liability, transparency and who is allowed to build agents.

For incumbents, particularly in financial services, the opportunity is significant but time-bound. The institutions that define how agents are verified, how mandates are expressed and how disputes are resolved will shape the economics of delegated commerce. Those that wait will find that the most important customer relationship in their market is no longer with a person, but with the software acting on that person's behalf.

References

1. Akerlof, G. A. (1970). The market for "lemons": Quality uncertainty and the market mechanism. Quarterly Journal of Economics, 84(3), 488–500.
2. Brynjolfsson, E., & Smith, M. D. (2000). Frictionless commerce? A comparison of Internet and conventional retailers. Management Science, 46(4), 563–585.
3. Calvano, E., Calzolari, G., Denicolò, V., & Pastorello, S. (2020). Artificial intelligence, algorithmic pricing, and collusion. American Economic Review, 110(10), 3267–3297.
4. Coase, R. H. (1937). The nature of the firm. Economica, 4(16), 386–405.
5. Diamond, P. A. (1971). A model of price adjustment. Journal of Economic Theory, 3(2), 156–168.
6. Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs and ownership structure. Journal of Financial Economics, 3(4), 305–360.
7. Mastercard. (2025, April 29). Mastercard unveils Agent Pay, pioneering agentic payments technology to power commerce in the age of AI [Press release].
8. Rochet, J.-C., & Tirole, J. (2003). Platform competition in two-sided markets. Journal of the European Economic Association, 1(4), 990–1029.
9. Stigler, G. J. (1961). The economics of information. Journal of Political Economy, 69(3), 213–225.