On August 24, the Payment & Clearing Association of China rolled out the "Self-Regulatory Convention on Agent Payment Applications" (referred to as the "Convention"), which took effect immediately upon its release. The document lays out a series of requirements covering agent management and consumer rights protection, among other areas.

Liu Gang, founder of the payment industry website PayCircle, told reporters that the Convention draws a clear "safety red line" for AI agents stepping into the core payment arena, pushing the sector from unregulated expansion toward high-quality, compliance-driven growth while safeguarding public funds and legal rights.

Wang Pengbo, chief analyst at Botong Consulting, noted that the Convention stresses that all agent-based payment activities must operate within the existing payment framework and adhere to baseline compliance standards, such as current payment, anti-money laundering, and consumer protection regulations. It also explicitly requires that core payment functions—like account management, transaction processing, and fund clearing—be handled exclusively by licensed institutions, a stipulation that will directly shape how many businesses take shape going forward.

Defining the Boundaries of Payment Authorization

As artificial intelligence technology accelerates and reshapes how people live, work, and how industries are structured, the Convention has emerged to steer agent payment applications toward safe, compliant, and sustainable growth.

On the agent management front, the Convention asks member institutions to build on the "Know Your Customer" (KYC) framework by exploring a "Know Your Agent" (KYA) mechanism. This involves identifying and verifying agents connected to the payment chain, enforcing real-name registration for users of agent payment services, and clarifying the agent-user relationship through formal agreements.

In Liu Gang's view, establishing the KYA mechanism tackles the trust dilemma head-on. It means agents will carry independent identity markers with tiered management, blocking malicious agents from tapping into the payment pipeline and resolving the "who is actually paying" identity question at its source.

When it comes to authorization, member institutions must define clear boundaries for agent payment applications and sign authorization agreements with users. These agreements should spell out key details such as transaction limits, designated deduction accounts and their priority, and validity periods.

On risk management, the Convention requires member institutions to set up a graded, category-based management system aligned with the functional attributes and risk levels of agent payment applications. For scenarios involving cross-agent platform calls or large-value transactions, a cautious approach is mandated. Institutions are also expected to establish emergency response protocols for unexpected incidents.

Additionally, following the principle of "whoever applies AI technology bears the responsibility," agent algorithms and model risks must be folded into the broader risk management framework, with ongoing assessments of robustness, transparency, traceability, and output reliability to mitigate issues like model black-box effects or hallucinations, ensuring that agent-involved payment transactions stay explainable and auditable.

Liu Gang pointed out that, in response to potential large-model pitfalls such as hallucinations, algorithmic bias, or injection attacks, the Convention calls for a full-lifecycle risk management system and a credible evidence-storage mechanism. This guarantees that agent payments are not just fast, but also stable and accurate, preventing user funds from being jeopardized by technological failures.

The Payment & Clearing Association of China has stated that any agent payment applications touching core payment functions—like account management, transaction processing, or fund clearing—must be carried out by licensed entities, including banking financial institutions, non-bank payment firms, and clearing houses. Member institutions should operate under the principle of "whoever provides the payment service bears the responsibility," taking on primary accountability for users' account and transaction security, fund safety, and information protection.

In line with the ethos of "payments for the people," the Convention requires member institutions to account for varying levels of digital literacy across different groups, preserving necessary human-operated services and alternative payment channels to avoid worsening the digital divide. More prudent risk alerts, permission controls, and transaction safeguards should be applied to vulnerable groups such as the elderly and minors.

Liu Gang emphasized that strengthening consumer agency and keeping "human options" available reflects the human touch of technology for good, curbing algorithmic price discrimination while reserving a "slow lane" for seniors in the digital age.

Wang Pengbo observed that the Convention first clarifies accountability for core payment functions, reinforcing the obligations of service providers, and second, stresses the need for a dedicated identification and management mechanism for agents themselves. This direction aligns with how major institutions are developing agent payments—requiring pre-checks, risk testing, and emergency handling for independently initiated agent transactions, while refining arrangements around transaction authorization, risk grading, and user rights protection.

Fastening a "Safety Belt" for Agent Payment Applications

Looking back at the evolution of the payment industry—from cash to card swiping, the surge of QR code payments, and the rise of facial recognition and frictionless payments—China's payment sector has never stopped innovating. Now, agent payments are moving from concept validation to real-world commercial deployment, with major players diving in.

In May, Alipay announced that its "AI Payment" had completed 300 million AI agent transactions and now supports 95% of common agent frameworks. That same month, Alipay introduced the world's first Token Pay service and an AI wallet product, joining its earlier AI Pay and AI Collect offerings to build a full-stack AI-native payment ecosystem tailored for the AI era.

In June, JD.com unveiled an agent autonomous payment protocol, categorizing agent payments into six levels from L0 to L5. The protocol zeroes in on the intermediate L3 and L4 tiers: L3 allows agents to initiate payment requests within a single task, with the system deciding whether to approve based on user-defined boundaries, while L4 grants agents broader authority to complete payments autonomously as long as amount, scenario, user, and other factors fall within preset parameters.

Around the same period, WeChat Pay launched an AI-exclusive card. Once an agent is authorized and connected, users can simply state their consumption needs in a conversation and experience an automated flow from smart recommendations to order placement and payment. Currently, the AI-exclusive card supports Meituan services within WorkBuddy.

Also in June, UnionPay Merchant Services rolled out a full-scenario AI payment solution, connecting the chain of "inquiry, selection, ordering, and payment." For example, in closed environments like tech parks or industrial zones, it offers conversational meal reservation services, allowing users to handle everything from queries and membership verification to ordering and payment through chatting with an agent.

Wang Pengbo highlighted that market attempts to link various agents to payment chains are multiplying, with internet giants and platforms driving real-world implementations and supporting protocols and industry ecosystems rapidly taking shape. However, he cautioned that this emerging niche still harbors significant gray areas—such as unclear division of responsibilities and risk handling that lack unified consensus, as well as data management and consumer protection without standardized benchmarks.

The Payment & Clearing Association of China stated that, based on trends in agent technology development and user authorization depth, the evolution of agent payments can be mapped across three stages: agent-assisted payment, autonomous agent payment under preset conditions, and autonomous agent payment under broad conditions. Member institutions should remain application-oriented and scenario-driven, prudently exploring agent-initiated payment transactions.

From an industry standpoint, Wang Pengbo believes the Convention can effectively steer market innovation, confining technological advances within safe and manageable bounds, reducing potential risks like money laundering and fund security that stem from chaotic experimentation, and offering market participants clear operational guidance.

"It must be emphasized that the industry reality is that agent payments are still in their infancy—fully autonomous payment scenarios have yet to scale widely, and most applications remain confined to assisting transaction processes," Wang Pengbo added.