Who Gets to Be the Customer When AI Does the Shopping?

AI commerce is being built through competing systems that decide what an agent may see, choose, authorize, and pay for. But the first people to use these systems will not necessarily be ordinary shoppers handing over their entire purchasing lives.

The early market is more specific than that.

It includes businesses with structured catalogs, developers paying for software resources, frequent online shoppers looking for price comparisons, enterprise finance teams automating routine payments, and consumers who want assistance while still approving the final purchase.

The important question is not whether everyone will adopt an AI shopping agent.

It is who already wants one, who is using one now, and what those early users tell us about the market that may follow.

The adoption gap is the story

The current evidence shows a sharp difference between using AI to begin a shopping journey and allowing it to complete one.

A survey of 1,000 American adults by NMI, a payments infrastructure company, found that 33 percent had used AI for shopping or product discovery during the previous month. Only 11 percent had used AI to complete a transaction. Seventy percent said it was important to review or override an AI decision before purchase, while 76 percent were uncomfortable with AI completing purchases on their behalf.

The same survey found an interesting difference between households. Forty eight percent of respondents with children at home had used AI for shopping or product discovery, compared with 25 percent of respondents without children at home.

That does not prove parents want autonomous purchasing. It suggests that people managing repeated errands, price comparisons, household needs, and time pressure may be more willing to use AI for the early parts of shopping.

The boundary appears at the moment of payment.

Visa’s 2026 Business to AI research found that 58 percent of surveyed Americans were comfortable with AI comparing prices and 55 percent were comfortable with AI applying discounts. Thirty eight percent were comfortable with AI completing a purchase. Only 27 percent were comfortable allowing AI to spend money autonomously without limits, while 60 percent would not allow AI to spend any amount without approval.

The consumer is not rejecting AI. The consumer is assigning it a smaller role.

AI may search. It may compare. It may identify a discount. It may produce a shortlist. The person still wants the authority to approve the purchase.

That is a very different market from the one implied by the phrase AI becomes the customer.

Businesses are further ahead than consumers.

In the same Visa survey, 77 percent of surveyed business decision makers said their organization was already using or piloting AI. Seventy one percent said they expected to optimize for AI agents, and 53 percent said they would allow AI agents to negotiate prices or terms with other AI agents. Eighty eight percent were willing to provide pricing or inventory data to enterprise AI systems.

These figures come from a survey commissioned by Visa, not from a neutral census of businesses. They should not be treated as proof that most companies are already selling to agents.

They do show where the commercial pressure is coming from.

Businesses have a reason to prepare before consumers have made up their minds. A merchant does not need millions of autonomous purchases to justify structured product feeds, machine readable inventory, delegated payment tokens, and agent compatible checkout. It may simply want to avoid being invisible when a new distribution channel becomes important.

Consumers can wait. Businesses may feel they cannot.

Who is using it now?

The current users fall into several identifiable groups.

The first is the approved partner ecosystem around AI shopping interfaces.

OpenAI’s Agentic Commerce Protocol allows merchants to provide structured product feeds so that ChatGPT can understand product descriptions, prices, inventory, media, and fulfillment information. OpenAI’s documentation says Instant Checkout is currently available to approved partners. The initial public launch began with United States Etsy sellers, with additional merchants and marketplaces expected to follow.

This is not yet an open marketplace where every retailer is equally available. It is a controlled channel in which participating merchants agree to make their products legible to an AI platform.

Stripe has identified early businesses using its agentic commerce infrastructure, including Etsy and URBN, the parent company of Anthropologie, Free People, and Urban Outfitters. Stripe has also named Coach, Kate Spade, and Ashley Furniture among major retailers testing its agentic commerce systems.

These businesses share several traits. They have large product catalogs. They already operate substantial online stores. Their products can be described through attributes such as price, size, color, inventory, category, delivery time, and return policy. Their digital systems are mature enough to support structured integrations.

This is not the same as saying that their customers have broadly delegated purchasing to AI. It means these retailers are positioning themselves for a channel in which AI influences discovery and may eventually control more of the checkout process.

The second group is made up of AI shopping and recommendation companies.

Visa has reported hundreds of controlled agent initiated transactions with partners including Consumer Reports, Skyfire, Gensmo, Fabrique, Price.com, Honeylove, BeyondStyle, Jomashop, and Ramp. In one example, Consumer Reports used Skyfire to demonstrate a purchase of Bose headphones through browser automation. In another, Gensmo connected AI generated fashion recommendations to purchases from Fabrique. Ramp used Visa’s infrastructure for business to business payment automation.

These examples reveal different potential markets.

Consumer Reports represents recommendation and comparison. Gensmo represents fashion discovery. BeyondStyle and Jomashop represent marketplace and retail aggregation. Ramp represents corporate payment automation.

The phrase AI shopping hides these distinctions. A person asking an AI system to compare headphones is not performing the same activity as a finance department automating recurring bills.

The third group consists of consumers participating in tightly controlled demonstrations.

In January 2026, Mastercard announced two authenticated agentic transactions in Australia. A Commonwealth Bank of Australia debit card was used to purchase cinema tickets from Event Cinemas. A Westpac credit card was used to book accommodation in Thredbo. Both transactions were completed with cardholder consent, and the payment participants could identify that an agent had conducted the transaction.

These are real transactions, but they are not evidence of ordinary mass adoption. They are evidence that card networks, banks, merchants, and AI providers are testing how responsibility can be preserved when software enters the payment chain.

The fourth group is using machine payments where the customer is already software.

Stripe’s Machine Payments Protocol is being used by companies such as Browserbase, which provides browser infrastructure; PostalForm, which lets agents pay to print and send physical mail; Prospect Butcher Co., which accepts agent orders for food; and Parallel Web Systems, which allows agents to pay for web access.

These examples may be closer to the immediate center of machine commerce than the vision of an AI buying ordinary household goods.

When the customer is already a software process, the transaction is easier to define. An agent requests a browser session, an API call, a research resource, a piece of compute, or a clearly specified service. The merchant can price the resource. The system can verify delivery. The payment can happen per request.

The ambiguity is lower than it is with a gift, a meal, clothing, or a product whose value depends on taste.

What the early adopters have in common

The early users are not random representatives of the population. They cluster around a particular set of conditions.

They have repeated tasks.

They have structured information.

They already use digital payment credentials.

They can define success clearly.

They benefit from speed, comparison, or automation.

They can tolerate a narrow interpretation of the purchase.

This explains why software services, business payments, travel bookings, tickets, fashion catalogs, and large online retailers appear early in the market.

A browser session is easy to define. A cinema ticket is easy to define. A hotel booking has dates, location, price, availability, and cancellation rules. A pair of headphones has a product identifier, price, specifications, and delivery options.

A meaningful gift is harder.

So is a local service whose quality depends on reputation and conversation. So is a purchase where the person wants to wander, notice something unexpected, or make a choice that cannot be reduced to price and delivery speed.

One reasonable prediction is that agentic commerce will first spread through purchases that are frequent, legible, comparable, and reversible.

That is not a prediction that people will never delegate more meaningful decisions. It is an argument that the burden of proof will be higher when the product has emotional, social, or identity value.

What history says about adoption

New payment technologies rarely become ordinary because the technology exists. They become ordinary when several systems change together.

There must be a usable interface. Merchants must accept it. Financial institutions must process it. Regulators must define responsibility. Consumers must understand the benefit. Fraud and dispute systems must function when something goes wrong.

Debit cards illustrate this clearly.

The Federal Reserve’s historical payment study recorded 8.3 billion debit card payments in the United States in 2000. By 2012, that figure had reached approximately 47 billion. The latest Federal Reserve Payments Study recorded 120.6 billion debit card payments in 2024.

That is not a single adoption event. It is a long institutional buildout.

Debit required bank accounts, card issuance, merchant terminals, payment networks, authorization rules, fraud detection, consumer familiarity, and enough acceptance that carrying the card became useful.

The behavior changed because the surrounding system changed with it.

Contactless payment followed a different path. In the United Kingdom, contactless payments grew from 3 percent of all transactions in 2015 to 38 percent in 2023. By 2024, consumers made 18.9 billion contactless payments. UK Finance reported 19.2 billion in 2025, representing 39 percent of all payments.

Contactless adoption accelerated because several conditions aligned. Terminals became widely available. Banks issued contactless cards. Payment limits increased. Mobile wallets gave consumers another way to use the system. Public transport and everyday retail made the habit repeatable. Consumers learned that the payment was fast and usually safe.

But contactless did not eliminate cash. The European Central Bank found that cash still accounted for 52 percent of point of sale payments in the euro area in 2024.

Online shopping provides an even longer comparison. US ecommerce represented less than one percent of retail sales in the first quarter of 2000. By the second quarter of 2026, it represented 17.1 percent of total US retail sales.

It took more than two decades for online retail to become a major channel. The growth was not smooth. It depended on broadband, logistics, search, online identity, payment security, returns, mobile devices, marketplaces, and changing consumer expectations.

The lesson is not that AI shopping will take twenty six years to become important.

The lesson is that the adoption clock begins before the consumer habit is visible. The infrastructure, standards, merchant integrations, and risk systems may be built for years before the public treats the new behavior as ordinary.

The behavioral shift may be larger than the payment shift

Debit changed how money moved.

Contactless changed the physical gesture of payment.

AI shopping may change where comparison and judgment happen.

That distinction matters because payment friction and decision friction are not the same thing.

A person tapping a card still knows what they are buying. An agent may remove the act of comparison before the person sees the available options.

Research on cashless payment suggests that reducing the visibility of payment can affect behavior. A meta analysis covering more than forty years of research found a relationship between cashless payment methods and increased consumer spending, although the effect has become smaller as people have become more familiar with digital payments.

Other research has associated contactless payments with a lower perceived pain of paying.

These findings do not prove that AI shopping will make everyone spend more. They do suggest that when payment becomes less visible, self control may move away from the moment of purchase and into advance settings.

AI agents could create a similar shift around deliberation.

A person who normally compares five stores may see one recommendation. A person who normally notices the difference between a local merchant and a large marketplace may receive only the option that fits the agent’s structured data. A person who might have abandoned the purchase after seeing the total may never experience the same sequence of hesitation if the agent completes the transaction under a preapproved limit.

The behavior may change in several ways. Purchases may become more frequent because search costs fall. Consumers may accept more automated replenishment because the effort of repeating the purchase disappears. Merchants may face stronger price competition because agents can compare prices faster than people.

At the same time, merchants may become more dependent on the systems that decide which prices and products are shown.

The likely market effect is not simply more sales. It is a redistribution of attention, comparison, and bargaining power.

From search ranking to agent ranking

Search engines changed the market by deciding which pages became visible.

Marketplaces changed it again by deciding which sellers appeared within a commercial environment.

Agentic commerce may move the ranking function closer to the act of purchase itself.

A search result can still give a person ten options. An agent may provide three. It may explain the shortlist, apply a preference filter, and complete the order.

This concentrates importance in the data and rules that shape the shortlist.

A merchant may need accurate product descriptions, live inventory, delivery promises, return terms, reviews, identity information, and payment compatibility. Businesses that cannot provide those signals may become difficult for an agent to recommend even if their products are valuable.

That could help small businesses if machine readable data lowers the cost of discovery. A specialist shop might reach a customer who would never have found it through ordinary search.

It could also disadvantage small businesses if every new channel requires technical integration, feed maintenance, standardized policies, and ongoing payments to intermediaries.

The phrase open protocol does not automatically mean open competition.

A protocol may be open while the platform controls the audience. A merchant may retain ownership of the product and remain the merchant of record while losing control of discovery, recommendation, and customer context.

This is one of the most important potential financial shifts.

In ordinary ecommerce, the merchant often owns the storefront and tries to attract the customer.

In agentic commerce, the agent may own the explanation of the product, the shortlist, the recommendation order, the payment permission, and the post purchase conversation.

The merchant may still fulfill the order, but another company may control how the customer arrived there.

What the protocols are really competing over

WebMCP helps an agent interact with structured website tools. ACP helps connect merchants and agents. AP2 attempts to record what a person intended to authorize. MPP extends processor infrastructure to machine payments. x402 proposes a more internet native route for paying for digital resources.

These systems are not interchangeable.

They are competing answers to different questions.

WebMCP asks what an agent may do on a website.

ACP and UCP ask how an agent, platform, merchant, and order system coordinate.

AP2 asks how a person’s intention can be recorded and connected to an eventual payment.

MPP asks how machine payments can use established processor infrastructure.

x402 asks whether software can pay directly for access to an internet resource through a request and response flow.

The financial future may not divide cleanly between traditional payments and crypto payments. Stripe’s x402 integration shows how an on chain payment can also be recorded as a Stripe PaymentIntent. A machine payment can therefore move through a more internet native protocol while still entering an established accounting and reporting system.

The deeper competition concerns where trust and value accumulate.

Will the important institution be the AI platform that controls discovery? The payment processor that controls authorization and settlement? The wallet that stores the agent’s spending permissions? The merchant that owns the product and fulfillment relationship? The network that decides how agents are recognized? Or the person who sets the limits?

What small businesses need to prepare for

Small businesses should not assume they need to build a fully autonomous shopping system immediately.

The first requirement is basic machine legibility.

A small business should know whether its product names, prices, inventory, delivery information, return terms, and contact details are accurate and consistently represented across its website and commerce systems.

The next step is understanding how an agent would interpret the business.

Can the system distinguish available products from discontinued ones? Can it understand variations? Can it find the return policy? Can it identify shipping costs before the customer commits? Can a person still intervene when the agent misunderstands the order?

Small businesses should also avoid scattering their effort across every emerging protocol. A better approach may be to use a commerce platform or payment provider that can support several standards while preserving the merchant’s own customer records and fulfillment systems.

The strategic goal is not to become dependent on one AI marketplace. It is to become compatible enough to be discovered while retaining a direct relationship with the customer.

Small businesses should also track agent referrals separately from ordinary traffic. A visitor who arrives through a recommendation is not necessarily equivalent to a customer who searches for the business by name.

The business will eventually need to know whether agents produce higher conversion, lower loyalty, more returns, smaller margins, or more price comparison.

What medium sized businesses need to prepare for

Medium sized businesses should focus on integration and measurement.

They are large enough to have multiple systems but often do not have the resources of a global retailer. Their greatest risk is fragmentation.

Product data may live in one system. Inventory may live in another. Shipping information may be maintained elsewhere. Customer service may not share the same order state as the payment provider.

An agent cannot create a reliable purchase experience from inconsistent information.

Medium sized companies should establish a single source of truth for product, inventory, price, fulfillment, refunds, and customer support. They should test how their data appears in AI shopping environments and measure what information is lost between recommendation, checkout, fulfillment, and post purchase service.

They should also define an agent policy.

Which purchases require human approval? What price limits apply? Can an agent use discounts? Can it substitute a product? What happens when inventory changes after recommendation? Who handles a dispute when the customer says the agent misunderstood the instruction?

These are not merely technical questions. They are commercial policies.

Medium sized businesses should also negotiate access to attribution and customer data. If an AI platform sends a customer but keeps the entire relationship inside its own interface, the merchant may become dependent on a channel it cannot fully measure.

What large businesses need to prepare for

Large businesses should treat agentic commerce as a distribution and risk issue, not simply as a new checkout feature.

They may need a dedicated policy layer that governs agent identity, spending permission, product eligibility, discounts, substitutions, authentication, refunds, and escalation.

They will need to distinguish a trusted agent from a malicious automated script. They will need to recognize when an agent is acting within its authority and when an instruction may have been manipulated.

Fraud systems will also change. Traditional fraud detection relies on signals such as device behavior, browser fingerprints, mouse movement, and unusual navigation. An agent may not generate those signals. Large businesses will need new signals based on identity, authorization, payment scope, agent reputation, transaction history, and the relationship between the person and the agent.

Large businesses should also prepare for protocol uncertainty.

ACP, UCP, AP2, MPP, x402, network tokens, wallets, and payment provider integrations may continue to evolve. A company that builds a deeply customized system around one standard may find itself locked into an ecosystem that changes direction.

The safer strategy is to maintain a protocol independent business layer that can expose catalogs, receive orders, process payments, manage fulfillment, and handle disputes across several agent channels.

Large companies should also model what happens if the agent becomes the first point of contact.

How much margin can be spent on referral or ranking fees? What happens if an agent prefers a competitor? Can the company preserve brand presentation inside an AI interface? Can it continue to sell premium products if an agent is optimized for price?

The challenge is not just accepting machine orders. It is remaining commercially visible when someone else controls the explanation of the product.

What the next adoption curve may look like

Historical comparisons suggest that AI commerce will probably develop in stages.

The first stage is controlled experimentation. This is the stage now visible in payment pilots, approved merchant programs, software services, and structured product feeds. Human approval remains common. The participants are companies that can define the transaction and measure failure.

The second stage is repeatable routine use. This may include software subscriptions, business payments, travel bookings, event tickets, replenishment, and high volume online categories. The agent becomes useful because the task occurs often enough that saving time outweighs the desire to perform every step manually.

The third stage is broader delegation. This would require consumers to accept more open ended authority, merchants to support reliable substitutions and returns, payment systems to assign responsibility clearly, and regulators to define protections when an agent acts beyond the user’s intention.

A reasonable historical extrapolation is that the first two stages could develop over several years if the infrastructure works and trusted platforms distribute the experience. Broader autonomy could take longer because it depends less on whether the software can complete a checkout and more on whether consumers believe the consequences are reversible.

This is also why machine commerce may move faster in software than in physical retail.

A failed API request can be measured in milliseconds and refunded through a defined system. A damaged sofa, an incorrect prescription, a missed hotel cancellation, or an unwanted subscription creates a more complicated dispute.

The adoption curve will therefore not be one curve.

It may be fast for machine to machine software payments, moderate for enterprise procurement, selective for high volume retail, and slow for purchases that depend on taste, trust, emotion, or physical inspection.

The potential future

The first possible future is that agents become a new layer of search and comparison. The person remains the decision maker, but merchants increasingly compete for inclusion in the agent’s shortlist.

The second is that agents become marketplaces. The agent or platform controls discovery, ranking, payment permission, and post purchase support. Merchants remain responsible for fulfillment but become dependent on a small number of channels.

The third is that machine commerce separates from human retail. Software agents buy data, compute, browser sessions, content, API access, and other digital resources while people continue to make more personal or ambiguous purchases themselves.

The fourth is a hybrid model in which people use different levels of delegation for different categories. They may allow automatic purchases for routine software or household supplies while retaining active control over travel, gifts, fashion, health, and local services.

The fourth future may be the most socially plausible because it respects how people already behave with money. The same person who wants contactless payment at a train station may still prefer cash for a private purchase. The same consumer who uses AI to find a product may not want an AI system to choose it without review.

The future of AI shopping may therefore be less about replacing the customer than about splitting the customer role into separate permissions.

Search may be delegated.

Comparison may be delegated.

Payment may be conditionally delegated.

Judgment may remain human.

Responsibility may remain contested.

The machine customer is a role, not a population

The phrase machine customer is useful because it describes a change in the transaction record. It becomes misleading if it suggests that the machine has become the consumer.

A person may still live with the outcome. A business may still carry the cost of fulfillment. A bank may still handle the dispute. A merchant may still be judged by a purchase it did not initiate and a recommendation it did not control.

The central question is not whether machines will shop.

It is which parts of shopping people will delegate, which institutions will profit from that delegation, and whether the market will remain visible enough for people to understand what they have given away.

The next customer may appear in a merchant’s system as an AI agent.

But the future of commerce will be decided by the permissions, limits, and liabilities that remain attached to the human being behind it.

Sources