On September 20, Baidu AI Cloud, FluxA and Visa showcased an Agent Spend solution. After a user provides a coding task, a budget cap and payment permissions, the agent can compare prices, performance and availability across different models, data services and computing resources, and complete the procurement accordingly.
This exploration, aimed at overseas markets, offers a new scenario for agent payments. The agent's tasks have expanded from consumer services such as ordering food and group buying to procuring digital resources for work tasks.
Just one day later, Ant Group announced an adjustment to Alipay's organizational structure, merging the former Alipay business group, the digital payment business group and the Sesame Credit business unit into a new Alipay business group. It also proposed bringing together the user entry point, payment execution and trust capabilities to accelerate its push into agentic commerce.
Ant Group CEO Han Xinyi said at Alipay's AI payment ecosystem conference in May this year: "AI payment is not the last step; it becomes the trusted commitment engine of agentic commerce."
This reflects a shared vision among major tech companies for agent payments: payment is no longer just the deduction step after an order is generated, but must enter the entire process of user delegation, service selection and transaction fulfillment.
As a result, competition in agent payments comes down to three questions: What can agents buy, who can be selected by agents, and once an agent makes a choice, who is responsible? The things being pried open are therefore not just the payment step, but the entire division of labor involving user entry points, merchant distribution and responsibility boundaries.
An expanding shopping list
Hub research found that agent payments are breaking through existing transaction boundaries along three paths.
The first exploration shifts service boundaries from within a platform to across institutions and service providers.
In January this year, Qwen received a task in a public demonstration to order 40 drinks. After the user gave quantity, flavor and delivery requirements, the agent filtered products on Taobao Flash Sale, matched discounts, generated an order and completed payment after authorization.
At that time, goods, fulfillment, payment and after-sales were all within the Alibaba ecosystem. Qwen's task was to break a natural-language sentence into search, comparison, product selection and ordering, then call each already-connected service one by one.
In April, China UnionPay released the Agent Payment Open Protocol framework, APOP, and completed five production-system verification transactions covering air tickets, hotels, in-car consumption and utility bill payments.
In a publicly displayed air ticket order, the Umetrip AI assistant compared flights, linked passenger information and used a Bank of Communications credit card to complete the purchase.
Compared with the milk tea transaction within Alibaba's system, APOP already involved multiple parties: Umetrip provided flight and ticketing services, Bank of Communications managed the account and credit card, and UnionPay connected the payment network. The user's request for "cheap" was translated into routes, dates, prices and refund/change rules, which traveled with the order across different institutions.
At a similar time, overseas institutions were also exploring ways to let independent service providers quote, collect payment and deliver directly to agents.
For example, Stripe, a U.S. payment infrastructure company, and Tempo, a payment settlement network project, released the Machine Payments Protocol (MPP), turning independent service providers' interfaces into transaction windows for agents.
Take a market analysis task as an example. A user can ask an agent to collect public information from competitors' websites and give it a budget and payment permissions. The agent sends a query to a web data service provider that supports MPP, and the interface returns the price and payment requirements for this data call. After payment is completed, the service provider delivers the required web content for the agent to complete subsequent analysis.
In this process, Stripe handles payment and settlement, allowing this kind of per-use, usage-based purchase to be embedded in the agent workflow.
The second exploration writes fund permissions into transaction rules in advance.
Tencent's public relations director Zhang Jun once used a simple analogy when the AI-exclusive card was released: when asking someone to do something for you, you need to give them the money to buy what they need; the same applies when asking an agent to do it.
In June, WeChat Pay launched an AI-exclusive card, with WorkBuddy among the first products to integrate it. Users first transfer funds into the card and set a limit, then ask the Meituan life assistant to generate a group-buying order and confirm the deduction on their phone.
Under this model, before tasks such as group buying and food ordering are handed to the agent, the amount and purpose have already been written into the rules of the exclusive card, while users still retain the right to confirm each transaction.
Mastercard's Agent Pay for Machines demonstrated a continuous procurement scenario: after a user asks an agent to build a flower shop website, the agent can purchase a domain name, web hosting, images and a checkout page in sequence within budget.
Continuous procurement requires the same budget to cover multiple payments. Domain names, hosting, images and checkout pages each generate expenses, and the payment system needs to verify permissions and balances for each transaction.
The third exploration extends the object of purchase from C-end life services to the digital resources needed for B-end tasks.
In previous cases, drinks, group buying and air tickets mainly served personal consumption.
In September, Baidu AI Cloud, FluxA and Visa showcased the Agent Spend solution, putting tokens, data services and computing power on the agent's procurement list. After the user provides a coding task, a budget cap and payment permissions, the agent compares prices, performance and availability across different models, data services and computing resources, and then decides what to buy.
Baidu vice president Yuan Foyu pointed out that as AI agents move from executing tasks to truly participating in commercial activities, tokens are no longer just the cost of model calls, but will become an important resource in the operation of the agent economy.
Enterprises can therefore record resource procurement, actual expenditure and task results in the same project ledger, and payment also enters resource selection and budget management.
Who can be selected by agents
No matter how much the shopping list grows, however, it ultimately still comes back to user choice.
The original reason users hand choices to agents is to reduce the attention they spend on product selection and price comparison.
When buying a winter quilt in the past, a consumer might first be drawn in by homepage recommendations, advertising atmosphere or even influencer promotion, then compare fillings, warmth ratings, sizes, prices and return/exchange terms across product pages, and finally claim discounts and complete payment.
After an agent takes over, the user only needs to specify size, warmth level, budget and delivery time. The agent can then compare similar products in a comprehensive mall based on filling, weight, price, inventory, delivery and discounts, and give a purchase recommendation.
For merchants, agent participation in product selection may challenge traditional marketing logic.
Traditional e-commerce sells to people. Homepage exposure, visual creativity and brand tone serve consumers' interests, aesthetics and even emotions. But once practical orders are led by agents, product selling points need to become structured, functional and machine-readable standard parameters.
Whoever has more complete information, a more suitable price and more reliable inventory and delivery is more likely to enter the candidate list.
Hub noted that explorations to write product capabilities into standard agent interfaces have already appeared.
In March 2026, OpenAI launched a product discovery feature in ChatGPT based on an agentic commerce protocol. Retailers such as Target, Sephora and Nordstrom can synchronize product catalogs and promotional information, and ChatGPT compares specifications, prices, inventory and promotions accordingly to offer candidate products.
This represents a change at the front end of payment: product names, parameters, prices, inventory, delivery and promotions need to be provided in a readable, comparable and continuously updated way in order to enter the agent's candidate list.
Domestic exploration is also underway. In May 2026, Alibaba connected Qwen to the full product catalog of Taobao and Tmall, so information such as products, prices, inventory, delivery and after-sales can be called by the agent along with the transaction chain within the platform.
Han Xinyi pointed out that agentic commerce is shifting toward supply matching driven by user intent, and merchants need to become the most trusted, best-matched and capable service providers.
In the future, e-commerce product information will inevitably be continuously standardized and incorporated into the agent decision chain. For a winter quilt, that may mean filling, weight, warmth rating, size, price and delivery time; for home appliances, it means energy efficiency, functions, size, installation and after-sales conditions.
As merchants begin exploring new operating logic, payers are also facing path challenges brought by agents recalculating payment plans.
The actual payment cost of online shopping is often not the listed price of the product. Store discounts, platform coupons, national subsidies, bank benefits and credit card cashback affect the actual amount paid, while account limits, installment eligibility and the acceptance scope of payment methods also affect whether a transaction can be completed.
This content is scattered across subsidy rules, store descriptions and payment interfaces on major e-commerce platforms.
Take the sale of the iPhone 18 Pro as an example. Hub research found that after stacking platform A's subsidies and store discounts, the price could be 900 yuan lower than the official price. Platform B offered fewer direct subsidies, but its directly operated online store provided a large phone bill gift. In terms of payment methods, Huabei supports 12 interest-free installments, while WeChat Fenfu supports 24 interest-free installments. Several state-owned major banks, joint-stock banks and city commercial banks, including ICBC and CCB, all had 24 interest-free installment promotions for this model. Among them, Ping An Bank, Shanghai Pudong Development Bank and China Merchants Bank added benefits such as repayment coupon packages, double points and temporary credit limit increases for existing customers on top of interest-free installments, while Bank of Nanjing and Bank of Jiangsu offered benefits such as phone accessory coupons.
Faced with complicated discount rules, many users choose to give up comparing and stick to a fixed payment path: go to the most familiar e-commerce platform, find the flagship store, and pay with Alipay, WeChat or a debit card.
The complex calculation logic creates a threshold for comparison, which instead allows payment channels to hold onto their closed loop through user inertia. In the future, if discounts and subsidies are redesigned around agents and the price comparison threshold disappears, whether a payment channel can remain in the final solution will no longer be a matter of habit, but a matter of calculation results.
Whether it can enter the optimal solution given by the agent is a new issue for payment channels.
At the end of 2024, Ant Group divided the Alipay platform and digital payment into different business groups. In September 2026, it merged the Alipay platform, digital payment and Sesame Credit again, so that user demand, agent authorization, payment execution and service fulfillment are connected into the same chain.
These moves mean Alipay is rearranging the foundation for participating in conversational and AI-assisted transactions. What needs to be verified next is whether it can connect user authorization, merchant discounts and payment benefits into the agent's calculations and continue to remain in the final payment solution.
How to be responsible for agents
After a payment channel enters the candidate solution, authorization, verification and fulfillment determine whether the transaction can be executed.
Preset authorization and per-transaction confirmation can coexist. The closer a transaction is to automatic execution, the more responsibility needs to be assigned to every party involved in the transaction.
For example, a delegation to "book a cheap flight" must first specify the destination, date, cabin class, refund/change scope and price cap. The available limit of the payment account, the scope of merchants that can be used and the method for returning ticketing information also form part of the authorization.
Users need to clarify the goal, budget, account and prohibited zones of the delegation. The agent platform needs to record what information it based its recommendation or execution on and whether it exceeded the scope given by the user. The payment service provider needs to verify the account, authorization, limit and risk control conditions.
For example, the travel service provider is responsible for truthfully displaying flights, prices and refund/change rules, and for completing ticketing, refunds and after-sales service. The payment network and rule-making party take on another task: enabling identity, authorization and dispute information to be transmitted among travel, banks, merchants and payment institutions.
UnionPay's Agent Payment Open Protocol framework, APOP, incorporates agent identity, user identity, transaction intent and payment authorization into the same verification process. The agentic commerce trust protocol turns user delegation, merchant confirmation of transaction conditions, payment service provider verification and record retention into transaction information that can be transmitted among different participants.
Fund boundaries need to be implemented through account design.
For example, WeChat's AI-exclusive card limits the agent's payment scope through separate funds, limits and per-transaction confirmation. The payment method binding rules of the agentic commerce trust protocol associate payment tools with real accounts, entrusted agents and applicable conditions, and support setting amount caps, validity periods and merchant scope.
At the same time, merchants also need to confirm who exactly is making the request.
When automatically accessing retail websites, user-authorized agents, ordinary crawlers and malicious traffic may exhibit similar behavior. Visa's Trusted Agent Protocol attempts to let agents prove their identity and authorization source to merchants through signatures bound to time, merchant and purpose.
The Self-Discipline Convention on the Application of Agent Payments released in August has incorporated authorization, limits, risk control, human services and fallback mechanisms into industry rules, and requires institutions to leave verifiable and traceable records for model decisions, payment instructions and risk control.
In-platform consumption can continue to use existing account, fulfillment and after-sales systems. Cross-institutional services and cross-provider procurement split authorization, accounts, payment and delivery among more parties, and the responsibility chain will first be tested in these scenarios.
When a ticket is booked incorrectly, the budget is exceeded or a service is not fulfilled, the system needs to be able to reconstruct the user's delegation, the agent's choice, the payment service provider's verification and the merchant's delivery.
After all, only when responsibility can be traced to a specific link can agent payments be expanded to transactions with a higher degree of freedom.
Comments