Alipay Completes 300 Million AI-Powered Transactions, Signaling the Rise of Agentic Commerce

Over the past decade, the most significant changes in the payments industry have largely occurred at the level of “how users initiate payments.” During the PC e-commerce era, users completed payments on web checkout pages; in the mobile internet era, users scanned QR codes offline or clicked payment buttons in apps online; in the life services era, payments became further embedded into high-frequency scenarios such as ride-hailing, food delivery, travel and accommodation, healthcare, and government services. The core challenge of the payments industry has always seemed to be enabling users to complete payments faster, more securely, and at lower cost.

After entering the AI Agent era, the nature of questions has changed. Future business interactions may no longer require users to personally open an app, search for products, compare prices, and click to pay; instead, users will simply say to an AI Agent: “Order me an iced Americano near me that has good ratings, costs under 30 yuan, and can be delivered within half an hour.” The Agent will then automatically understand the preferences, access mapping, food delivery, coupon, merchant inventory, and payment interfaces to complete the order and payment within authorized limits.

At this point, the payment system must verify not just whether the account is valid, the password is correct, or the merchant exists—but a more complex set of questions: Did the user truly authorize this Agent? Does the Agent’s behavior align with the user’s original intent? Are the transaction amount, merchant, category, and time within the authorized boundaries? If the Agent is compromised by prompt injection, malicious plugins, or fraudulent merchants, can the payment system detect and block it?

Alipay disclosed on May 26, 2026, that AI payments via Alipay have reached 300 million AI agent transactions and support 95% of general-purpose agent frameworks. During the same period, Alipay launched AI Pay, AI Receive, Token Pay, and an AI Wallet, building an open technical protocol framework for agent-based commerce around ACT, the Agentic Commerce Trust Protocol (Agentic Commerce Trust Protocol),. Additionally, public information shows that product development accelerated further afterward: on June 16, the AI version of Alipay, “Abao,” was officially launched and began invitation-only testing; on July 2, “Abao” completed its next product iteration and opened to public testing, initially introducing 72 intelligent service capabilities covering home management, transportation, discount shopping, government documents, wallet and bill management, and companion interaction scenarios.

Individually, 300 million transactions is a product scale metric; but when viewed together with the disclosure of AI payment scaling in May, the AI version of Alipay’s invitation-only testing in June, and the public beta launch in July, the changes are no longer limited to underlying payment capabilities—they are beginning to extend into user-facing product interfaces and service entry points. Around the same time, OpenAI and Stripe launched Agentic Commerce Protocol (ACP) (in-chat instant checkout), Mastercard introduced Mastercard Agent Pay (Mastercard Agent Payment), Visa unveiled Trusted Agent Protocol (TAP) (Trusted Agent Protocol), and Intelligent Commerce Connect (Intelligent Commerce Solution), while China UnionPay released APOP: Open Protocol Framework for Agent Payments.

Therefore, the core judgment of this report is that the Agentic Economy will not eliminate payments; rather, it will elevate the importance of payments. Under the new paradigm of “humans delegating tasks to agents,” payment systems will no longer serve merely as the final step for deductions, but will become the central hub for verifying identity, reconstructing intent, expressing authorization, controlling budgets, allocating responsibility, and building trust.

Three shifts in payment interaction paradigms, source: Created by Clare Yang, Researcher at Web3Caff Research

Author: Clare Yang, Researcher at Web3Caff Research

Cover: Photo by Illia Kholin on Unsplash, Typography by Web3Caff Research

Word count: The entire text contains over 17,100 words.

Note: Due to length constraints, this report is published in two parts. This is Part One (including sections: Key Takeaways: AI-Powered Payments as Infrastructure—Insights from Alipay; From Assistant to Executor: Why Is the Agent Economy Emerging? And Why Is Payment the Final Piece of the Agent Economy Puzzle? ACT, A2A, and A2M: What Are the Underlying Protocols Solving? Alipay’s Full-Stack AI-Native Payment System: From QR Code Scanning to Trusted Payment; Evolution of Authorization Models: What Exactly Are Users Entrusting to Agents?). The remaining sections (Real-World Applications: Which Use Cases Are Most Likely to Emerge First? Global Competitive Landscape: Who Is Competing for the Next Default Payment Option? Regulation and Risk Management: AI Autonomous Payments Cannot Rely Solely on “User Consent”; Business Models: How Will AI Payments Transform Revenue, Cost, and Settlement Structures? Where Will Agent-Based Payments Take the Industry? Eight Critical Questions: What Has AI Payment Actually Changed? Restructuring of Participant Capabilities: Who Is Best Positioned for Key Roles in Agent-Based Payments? Research Boundaries: Which Conclusions Still Require More Data Validation? Final Thoughts: The Real Issue with AI Payments Is How the Payment Industry Understands Trust) will be completed in Part Two.

Table of Contents

• Key takeaway: AI payments as infrastructure, as seen from Alipay samples
• From Assistant to Executor: Why Is the Agent Economy Emerging?
• Why is payment the final piece of the agent economy puzzle?
• ACT, A2A, and A2M: What exactly are the underlying protocols solving?
• Alipay's Full-Stack AI-Native Payment System: From QR Code Scanning to Trusted Payment
• Evolution of Authorization Models: What Exactly Are Users Giving to the Agent?
• Scenario implementation: Which applications are most likely to emerge first?
• Global competitive landscape: Who is vying for the next default payment option?
• Regulation and Risk Control: AI Autonomous Payments Cannot Rely Solely on “User Consent”
• Business model: How will AI payments transform revenue, cost, and settlement structures?
• Where will autonomous agent payments take the industry?
• Eight key questions: What has AI payment actually changed?
• Will AI payments simply be an upgraded version of password-free payments?
• Will the entry point for the payment app be weakened?
• Will merchant relationships be restructured by AI platforms?
• Should a payment platform be a closed ecosystem or an open protocol?
• Will micropayments become the main battlefield for AI payments?
• Will AI payments increase the risk of fraud?
• How will regulation affect the product boundaries of AI-powered payments?
• What is the core barrier for payment institutions?
• Reconstructing participant capabilities: Who is closer to the key position in agent-based payments?
• Research boundaries: Which conclusions still require more data for validation?
• Final thought: The real issue with AI payments is how the payments industry understands trust.
• Key structure diagram
• References

Let’s start with the specific number: 300 million transactions. If viewed merely as the transaction volume of a new feature on Alipay, the conclusion remains at the product level; but when observed within the broader产业链 of the agent economy, it resembles a sample that can be deconstructed: AI payments are evolving from a “pay button” toward “agent-based commercial infrastructure.”

This report is not merely about Alipay’s product launch, nor is it a simple comparison of which payment providers act faster. The fact that Alipay processed 300 million AI payments is worth examining as a starting point because it presents AI payment, AI collection, Token Pay, AI wallet, and the ACT protocol within a unified system, offering a relatively comprehensive view of the potential layers of evolution in agent-based payments: how users authorize transactions, how merchants integrate, how AI applications are billed, how payment platforms manage risk control, and how the protocol layer enables cross-scenario collaboration.

From this sample, the following key judgments can be drawn.

First, 300 million transactions are a signal of scalability, but not full commercial validation.

As of May 26, 2026, public records show that Alipay has completed 300 million AI agent transactions and supports 95% of general-purpose agent frameworks. This indicates that AI-native payments are no longer just conceptual demos but are beginning to integrate into real transaction workflows. However, public data has not yet disclosed total transaction volume, average transaction amount, refund rate, complaint rate, fraud rate, scenario distribution, or user retention rate; therefore, the 300 million transactions cannot be directly equated to a fully mature business model. A more objective assessment is that Alipay has validated the scalability and feasibility of its AI payment infrastructure, but its long-term value still requires additional quality metrics to be confirmed.

Second, the payment entity is expanding from "user clicks" to "user-authorized Agent executes."

Traditional payments assume the user is physically present at the screen to search, place an order, confirm, and pay; the default premise for agent-based payments, however, becomes: the user sets goals, preferences, budget, and constraints, and the AI agent executes the task within authorized boundaries. The entities that the payment system must verify expand from accounts, passwords, devices, and merchants to include agent identity, user intent, authorization scope, budget constraints, and execution process. In other words, the payment industry is no longer just asking “Can this payment be deducted?”, but “Should this agent deduct this payment in this context?”

Third, the Alipay AI payment matrix reflects a shift in product model, not just a change in payment entry points.

AI Pay solves the problem of "in-conversation payments" and "in-task payments" on the user side; AI Receive addresses how merchants and developers can enable their services to be invoked by Agents and receive payments; Token Pay corresponds to token top-ups, subscriptions, and pay-as-you-go billing within AI applications (where "token" refers to the billing unit of AI applications); and the AI Wallet handles user authorization, budgeting, billing, and revocation management. Viewed together, Alipay is not merely embedding a payment button into AI applications—it is attempting to decompose payment capabilities into foundational modules: Pay, Receive, Billing, Management, Risk Control, and Protocol.

Fourth, the significance of ACT lies in elevating Alipay's product capabilities to a protocol level.

If only AI payment and AI receipt are involved, Alipay is simply offering a product feature; but with the addition of ACT, or the Agentic Commerce Trust Protocol, Alipay is addressing a more fundamental question: how can different agents, merchants, devices, wallets, and service platforms express authorization, confirm intent, transmit orders, complete payments, and maintain auditable records? Global reference points include OpenAI/Stripe’s ACP, China UnionPay’s APOP, Visa’s Trusted Agent Protocol, Mastercard’s Agent Pay, and Google’s AP2. Together, they illustrate that competition in AI payments is shifting from the entry layer to the protocol layer—though each protocol approaches this shift from a different angle.

Fifth, the role of payment platforms is evolving from a funds channel to a Trust Layer.

In the agent economy, payment platforms do more than simply process deductions and settlements—they must prove “who authorized, what was authorized, who was paid, whether boundaries were exceeded, and how to trace errors.” This Trust Layer capability is built upon account systems, identity verification, risk control models, authorization management, merchant networks, dispute resolution, and compliance expertise. The more automated the transactions, the more critical it is to have a trustworthy payment infrastructure to underpin them.

Sixth, the Alipay sample is transitioning from "infrastructure release" to "user-side adoption validation."

The 300 million AI agent payments disclosed in May demonstrate that the relevant payment infrastructure is already capable of scalable operation. The launch of Alipay’s AI version, “Abao,” into invite-only testing on June 16, followed by public testing on July 2, shifts the validation to users: Are users willing to access lifestyle services through conversational interfaces? Which tasks will drive high-frequency usage? Can the journey from “finding a service” to “completing the transaction” truly be shortened?

As of July 2, the beta version has launched 72 intelligent service capabilities, meaning Alipay is transforming abstract Agent functionalities into direct service access points for everyday users. What the industry needs to observe next is no longer just whether a platform has released an AI payment feature, but whether users are willing to delegate certain tasks and payment permissions to Agents, whether merchants are willing to open access to their products, inventory, pricing, and after-sales policies to Agents, and whether regulators recognize this “trusted execution” chain of responsibility.

From an external market perspective, the agent economy is no longer just a topic of discussion within tech circles, but is gradually being integrated into long-term planning for enterprise applications, consumer commerce, and payment networks. Institutions and companies such as Gartner, Fortune Business Insights, McKinsey, and Visa have provided various forecasts from different angles, including enterprise adoption rates, the size of the agentic AI market, AI agent participation in consumer commerce opportunities, and AI-driven access growth.

Key projections for the agent economy market size, chart by Clare Yang, researcher at Web3Caff Research

However, these predictions are better suited as trend indicators rather than direct evidence that AI payments have matured. For the payments industry, what truly needs validation is not “how large the agent market is,” but whether users are willing to authorize agents to spend money, whether merchants are willing to open access to their products, inventory, pricing, and after-sales policies to agents, and whether payment institutions can successfully perform identity verification, intent reconstruction, risk control, and liability attribution within automated transactions.

External predictions offer ample room for imagination, but whether the market can truly be activated ultimately depends on whether the underlying logic holds. To understand why AI payments matter, we need to first refocus on the fundamental changes AI itself is undergoing.

Over the past few years, most users have encountered Generative AI. It excels at answering questions, generating text and images, assisting with writing, and summarizing information. Users ask questions, and AI responds—the decision-making power remains primarily with humans. At this stage, AI functions more like an efficiency tool, helping users obtain information faster but not actively executing trades on their behalf.

Starting in 2025, Autonomous Agents are gradually becoming the central focus of the technology industry. Unlike conventional chatbots, Agents possess capabilities for planning, tool invocation, execution, observation, and reflection. Given a goal, an Agent can break down tasks, invoke external tools, observe outcomes, and adjust its approach upon failure. OpenAI’s Agents SDK primarily addresses how Agents invoke tools, hand off tasks, establish safety boundaries, and track execution processes; Anthropic’s MCP protocol, on the other hand, focuses on how models can connect to external tools and data sources in a unified manner.

In other words, the role of AI is evolving: previously, it primarily answered “What should I do?”, but in the future, it may go further to execute “Help me get this done.” This shift will directly impact business relationships. Users will no longer perform actions step by step; instead, they will set goals, budgets, preferences, and constraints, allowing Agents to operate within those boundaries. The essence of business remains unchanged—it is still about matching needs, exchanging value, and fulfilling commitments—but the way transactions are initiated, authorization is expressed, and responsibilities are allocated is changing.

To better understand this change, the role of AI in the business chain can be divided into four stages: information retrieval, generative AI, tool invocation, and agent economy.

The shift in business logic from generative AI to the agent economy, source: created by Clare Yang, researcher at Web3Caff Research

As shown in the table above, the responsibilities of the payment system do not change abruptly but gradually increase alongside the evolving role of AI. During the information retrieval phase, the payment system only needs to support manual user payments; during the generative AI phase, payments primarily serve subscriptions, top-ups, and content purchases; during the tool invocation phase, payments begin to handle pay-per-use billing for APIs, SaaS plugins, and enterprise services; upon entering the agent economy phase, the payment system must simultaneously verify identity, intent, authorization, budget, and audit trails.

In addition, transactions will become more fragmented, higher frequency, and cross-platform. AI tool calls, model inference, API queries, file parsing, image generation, and IoT automated purchases may all generate a large volume of microtransactions. Users cannot repeatedly navigate and confirm each payment of a few cents, dimes, or yuan, and merchants cannot develop separate payment pathways for each agent.

Therefore, the real question agent economics brings to the payments industry is not “Can AI help users find products,” but rather “When AI begins to perform tasks on behalf of users, who verifies that the transaction is trustworthy, controllable, settleable, and traceable?” This is why payments are gradually shifting from being a peripheral tool in the commerce chain to becoming the trust infrastructure of the agent economy.

Reimagining payment locations in the Agentic Economy, source: Created by Clare Yang, researcher at Web3Caff Research

AI recommending products isn’t enough—why must it be able to complete payments? Because the endpoint of a commercial loop isn’t “knowing what’s worth buying,” but “executing transactions in an authorized, traceable, and settleable manner.” An agent that only recommends is still fundamentally an information tool; only an agent that can securely complete payments can enter the layer of commercial infrastructure.

In traditional transactions, users open a page, select an item, click a button, enter a password, or complete biometric authentication, allowing the payment system to clearly determine that the user themselves participated in the transaction. In transactions executed by an agent, the user may not be present for every payment.

The question the payment system needs to answer becomes: Does this deduction still comply with the user’s prior authorization?

The challenge of agent-based payments isn't the deduction itself, but understanding the user's true intent. “Buy me a coffee” may seem simple, but it carries many implicit conditions: proximity, price, flavor, delivery time, merchant rating, whether to use a coupon, whether alternative locations are acceptable, and whether to auto-confirm. If the agent interprets “buy me a coffee” as purchasing an expensive coffee machine, even if the payment technology successfully processes the deduction, the transaction cannot be considered validly authorized.

Therefore, agent payments must simultaneously handle five key concepts: intent recognition, authorization boundaries, revocable authorization, budget control, and accountability tracking. If any one of these elements is missing, users will find it difficult to entrust the “power to spend” to an agent.

Five Trust Issues in Agent Payments, courtesy of Clare Yang, researcher at Web3Caff Research

From the table above, it is clear that agent payments are not merely an upgraded version of traditional password-free payments. Password-free payments typically occur after the user has already entered a specific scenario, such as automatic deduction after a ride-hailing trip or automatic renewal upon subscription expiration; in contrast, agent payments occur when an agent performs the entire workflow on behalf of the user—including searching, comparing, selecting, placing an order, and making payment. The payment system must not only process the deduction but also understand why the agent made this choice and whether it remains within the user’s authorized scope.

Here, we can draw on the metaphor of the “container moment.” Containers revolutionized global trade not because the boxes themselves were complex, but because they turned previously fragmented, chaotic, and hard-to-measure loading and shipping processes into standardized interfaces. Similarly, AI payments for agent-based commerce carry the same significance: it’s not merely about providing a payment button, but about transforming agents’ commercial actions into standardized, authorized, billable, settleable, and traceable processes.

Precisely for this reason, the next challenge for the payments industry won't stop at the product level. A single payment button can complete a transaction, but it cannot support complex collaboration across agents, merchants, devices, and platforms. What will truly enable agent-based commerce to scale is a protocol language that all participating parties can understand and adhere to.

This leads to the question to be discussed in the next chapter: What exactly are ACT, A2A, and A2M solving?

As previously explained, the challenge with AI payments is not whether a payment can be made, but whether the payment aligns with the user’s intent, authorization boundaries, and risk rules. If an Agent only occasionally makes a single payment within one app, the existing product functionality may suffice; however, if the Agent needs to execute tasks across platforms, merchants, and devices, a standardized language understandable by all participating parties is required.

This is the value of the protocol layer.

In the Alipay sample, ACT, or the Agentic Commerce Trust Protocol, can be understood as Alipay abstracting the common challenges behind AI Pay, AI Receive, Token Pay, and AI Wallet into protocol capabilities. It’s not about “how to add another payment button,” but rather: How are Agents authorized? How do they interact with merchants? How do they initiate payments? And how does the payment system determine whether a transaction is trustworthy?

To understand ACT, you first need to understand two foundational concepts: A2A and A2M.

A2A, or Agent-to-Agent, refers to interaction between agents. A user agent may need to collaborate with merchant agents, logistics agents, customer service agents, invoice agents, or enterprise financial agents. For example, if a user says, “Help me arrange a business trip,” the main agent might need to check flight options, book a hotel, verify the budget, request an invoice, and complete the payment. A2A addresses how multiple agents communicate, collaborate, hand off tasks, and maintain audit trails.

A2M, or Agent-to-Machine, refers to agents interacting with machines, services, or APIs. Many agent tasks do not involve communication with another agent but instead call model APIs, cloud services, vehicle systems, IoT devices, enterprise SaaS platforms, or MCP tools. In such cases, the agent may consume tokens, computing power, storage, file parsing, or a single API request. A2M addresses how machine resources are accessed, billed, paid for, and audited by agents.

This also explains why Alipay is not only launching AI Pay, but also introducing AI Receive, Token Pay, and AI Wallet simultaneously: AI Pay addresses user-side payments, AI Receive addresses merchant and developer-side collections, Token Pay enables token top-ups and pay-per-use billing for AI applications, and the AI Wallet manages user authorization, budgeting, and billing. ACT integrates all these capabilities into a unified agent-based commercial trust framework.

Note that the term "token" in this article has two meanings: one refers to usage units in AI applications, such as tokens consumed by model calls, content generation, or tool invocations; the other refers to tokenized payment credentials in payment networks, which replace full account information with restricted, traceable payment credentials. The former is closer to a billing unit, while the latter is closer to a payment security mechanism.

Next, this chapter will introduce OpenAI / Stripe, UnionPay, Visa, Mastercard, Google, and Stripe MPP—not to simply list protocol names, but to illustrate that AI payments are not an issue faced by Alipay alone, but a shared challenge being addressed by global payment networks, AI platforms, and card networks alike.

OpenAI and Stripe's ACP (Agent Commercial Protocol) focuses on product purchases and checkout within AI conversational interfaces; China UnionPay's APOP (Agent Payment Open Protocol) emphasizes clearing networks and cross-institutional collaboration; Visa's Trusted Agent Protocol focuses on helping merchants distinguish trustworthy agents from malicious bots; Mastercard's Agent Pay centers on tokenizing payment credentials within card networks, enabling agents to obtain restricted, controllable, and traceable payment capabilities without direct access to full account information; Google's AP2 (Agent Payment Protocol) emphasizes authorization, authenticity, and accountability chains; Stripe's MPP (Machine Payment Protocol) focuses on machine-to-machine payments, micropayments, and developer billing.

Comparison of agent-related commercial protocols, source: Created by Clare Yang, Researcher at Web3Caff Research

Global Agent Payment Protocol Competitive Positioning, sourced from: Created by Clare Yang, Researcher at Web3Caff Research

From this positioning map, it is clear that Alipay ACT is not equivalent to OpenAI/Stripe’s ACP, Visa TAP, or Mastercard Agent Pay. OpenAI/Stripe focus more on AI conversation interfaces and merchant checkout, Visa and Mastercard emphasize card networks, tokenized credentials, and secure merchant identification, UnionPay APOP leans toward clearing networks and inter-institutional collaboration, while Google AP2 and Stripe MPP prioritize open protocols and machine payments.

The uniqueness of Alipay ACT lies in that it does not enter through a single AI gateway, card network, or clearing network, but instead leverages Alipay’s existing user accounts, merchant network, wallet capabilities, risk control system, and clearing and settlement infrastructure to integrate AI Pay, AI Receive, Token Pay, and AI Wallet into a relatively complete AI-native payment system. In other words, ACT is more like Alipay’s expression of elevating its product capabilities to the protocol layer: addressing trust and authorization at the bottom, connecting merchants and developers in the middle, and supporting users’ payment tasks across various Agent scenarios at the top.

Therefore, after understanding ACT, the next step is to return to Alipay itself and examine how these protocol capabilities are implemented in specific products. AI Pay, AI Receive, Token Pay, and AI Wallet together form the four core modules of Alipay’s natively AI-powered payment system, determining whether Alipay’s AI payments remain as isolated feature updates or evolve into a more comprehensive intelligent agent payment infrastructure.

ACT is a protocol expression; the product matrix is its commercial application. Returning to Alipay itself, AI Pay, AI Receive, Token Pay, and AI Wallet together form the four core modules of an AI-native payment system.

AI-powered solutions for C-end users and Agent scenarios are fundamentally shifting from "scan-to-pay" to "payment within tasks." The payment action may no longer occur within a payment app, but instead be embedded directly within the Agent's task flow.

AI Pay is designed for merchants and developers, addressing how merchants can enable their services to be understood, invoked, and paid for by agents. Payments are no longer just components within websites or apps—they become part of the agent’s toolchain.

Token Pay's billing structure for AI applications converts cost units such as tokens, number of calls, computing power, tasks, or results into payable, authorized, and settleable commercial units.

The AI wallet is designed for user authorization management. It is not an additional payment gateway, but rather a console that helps users manage Agent authorizations, budgets, bills, scenarios, merchants, and risk alerts.

Looking at these four products, Alipay AI Payments focuses not on isolated feature upgrades, but on combining “pay, receive, billing, and authorization management” into a reusable payment infrastructure.

On July 2, the AI version of Alipay, “Abao,” launched its public beta, offering a fresh window into how these capabilities are reaching users. Unlike traditional apps that organize services through pages and functional menus, “Abao” attempts to match users’ intentions—expressed in natural language—to specific services, completing tasks such as inquiries, transactions, coupon claims, shopping, and planning. The initial beta release introduces 72 intelligent service skills.

It’s important to distinguish that “Abao” is not equivalent to AI payment itself. More accurately, it provides a new agent interaction entry point; while AI Pay, AI Receive, Token Pay, AI Wallet, and ACT handle more foundational capabilities such as payment processing, receipt collection, billing, authorization management, and protocol trust. Together, they more closely resemble Alipay’s full journey from an “AI payment product matrix” to an “agent service entry point + payment infrastructure.”

Alipay's AI-native payment "quad" matrix, source: Clare Yang, researcher at Web3Caff Research, excerpted from public reports on the Alipay AI Payment Ecosystem Conference

Alipay's Native AI Payment Product Matrix Breakdown, sourced from: Clare Yang, Researcher at Web3Caff Research

Alipay AI Payment Key Milestone, image source: Clare Yang, researcher at Web3Caff Research

Thus, Alipay AI Payment focuses not just on adding a payment entry point, but on breaking down payment, receipt, billing, and authorization management into reusable foundational capabilities.

But the product matrix can only answer “whether it can be done”; what truly determines whether users are willing to use it is the authorization boundary. Next, we need to discuss: how much payment authority are users willing to grant to the Agent?

The real sensitivity around agent payments isn't whether AI can make payments, but the extent to which users have authorized them. There's a fundamental difference between users allowing an agent to recommend products and users permitting an agent to automatically place orders and make payments.

Therefore, the implementation of AI payments will not happen overnight but is more likely to follow a gradual path: from per-transaction confirmation, to rule-based automatic deductions, and finally to budget or fund pool authorization.

Batch confirmation is the earliest model most easily accepted by users. Users see exactly where each payment is going, the amount, and the merchant before deciding whether to pay. This approach provides strong security and is well-suited for high-value, low-frequency, or unfamiliar scenarios. OpenAI’s Instant Checkout also currently emphasizes user confirmation before execution, with payment authorization tied to specific amounts and merchants. However, the drawback of batch confirmation is clear: it is inefficient and unsuitable for high-frequency, low-value, repetitive Agent tasks.

Rule-based automatic payments are the second stage. Instead of approving each transaction individually, users set rules in advance—such as a maximum of 30 yuan per transaction, no more than two transactions per day, only allowed during weekday mornings, restricted to specific categories or merchants, and requiring secondary confirmation if dynamic price increases exceed 30%. Scenarios like coffee, food delivery, ride-hailing, parking, and AI tool usage are better suited to this model. Its core is not “seamless,” but “bounded automation.”

Budget pools or funding pools are better suited for enterprises and complex tasks. Enterprises can allocate monthly budgets to specific departments, projects, or Agents, enabling Agents to procure cloud APIs, office SaaS, travel services, or data services within the budget. At this point, the payment system’s role extends beyond mere deduction—it must also handle budget aggregation, approval rules, invoice matching, cost allocation, and audit trails.

These three modes correspond to the gradual building of user trust: initially, users require confirmation for every transaction; as scenarios become stable and risks sufficiently controlled, users may delegate certain rules in advance to an Agent; in enterprise settings, authorization further becomes part of budgeting, approval, and auditing systems.

Evolution of Agent Payment Authorization Models, courtesy of Clare Yang, Researcher at Web3Caff Research

Under this framework, the value of an AI wallet will be primarily reflected in authorization management: users need to see which agents are authorized, how much they can spend, in which scenarios they can act, and whether actions can be paused or revoked in case of anomalies.

Therefore, the authorization model determines how far AI payments can go. The product matrix addresses “whether it can be done,” while the authorization design addresses “whether users will dare to use it.” Only when users can see, control, and revoke access can agents move beyond occasional payment assistance into more frequent and complex payment scenarios.

Once the authorization boundaries are clearly defined, the focus shifts further to scenario selection: which applications are best suited to adopt this payment method first?

Alipay ACT Protocol Official Website

Shanghai Securities News / Sina Finance: AI payments exceed 300 million transactions; Alipay launches AI Wallet and Token Pay

[3] IT Home: Alipay completes 300 million AI payments, launches AI Wallet and Token Pay

[4] OpenAI: Buy it in ChatGPT, Instant Checkout and Agentic Commerce Protocol

[5] OpenAI Agents SDK Documentation

[6] Anthropic: Introducing the Model Context Protocol

[7] Google A2A Protocol Specification

[8] Google Cloud: Announcing Agent Payments Protocol AP2

[9] Visa: Official Introduction to the Trusted Agent Protocol

[10] Visa: Intelligent Commerce Connect

[11] Mastercard: Mastercard unveils Agent Pay

[12] Stripe: Introducing the Machine Payments Protocol

[13] UnionPay / PR Newswire: UnionPay Launches Agentic Payment Open Protocol Framework

[14] Gartner: 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026

[15] Fortune Business Insights: Agentic AI Market Size

[16] McKinsey: The Agentic Commerce Opportunity

[17] FATF: Recommendation 16 on Payment Transparency

European Commission: AI Act enters into force

[19] Bank of England: Artificial Intelligence Consortium

Disclaimer: This report or content (collectively referred to as the report) is prepared by Web3Caff Research. The information contained herein is for reference purposes only and does not constitute any prediction, investment advice, recommendation, or offer. Investors should not rely on such information to buy, sell any securities, cryptocurrencies, or adopt any investment strategy. The terms used and opinions expressed in this report are intended to aid understanding of industry trends and promote responsible development within the Web3 new economy, including the blockchain industry, and should not be interpreted as explicit legal opinions or positions of Web3Caff Research. The views expressed in this report reflect only the personal opinions of the authors as of the stated date and are unrelated to the official stance of Web3Caff Research; these views may change with subsequent developments. The information and opinions contained in this report are derived from proprietary and non-proprietary sources deemed reliable by Web3Caff Research, but do not necessarily encompass all data and are not guaranteed for accuracy. Accordingly, Web3Caff Research makes no representations or warranties regarding its accuracy or reliability and assumes no liability for any errors or omissions arising in any other manner (including liability to any person due to negligence). This report may contain “forward-looking” information, which may include predictions and forecasts; however, this document does not constitute a guarantee of any such predictions. The decision to rely on the information contained in this report is entirely at the reader’s own discretion. This report is provided for informational purposes only and does not constitute any investment advice, recommendation, or offer to buy or sell any securities, cryptocurrencies, or adopt any investment strategy. Please strictly comply with relevant laws and regulations in your country or jurisdiction.