Before You Hand Your Wallet to AI, the Payment Network Has to Be Rewritten
Only one of the four guests had actually paid for something with an agent — 40-odd yuan. Mastercard says the real bottleneck isn't technology, it's getting ordinary consumers to genuinely trust an agent.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Agent commerce is stuck in a chicken-and-egg loop
Han Xinyi said that since Q4 last year everyone has been very optimistic about Agentic Commerce, but actual deployment this year has been much slower than expected. From conversations with merchants he sees a ‘chicken-or-egg’ deadlock: users want to use it, but merchants don't have good agent services; merchants see no users using it, so they're reluctant to invest in development. His solution is for Ant to be the catalyst — using the chemistry metaphor from his childhood, potassium chlorate turning into oxygen needs manganese dioxide. The concrete move is the ‘wishing wall’ at the Alipay forum that afternoon, letting users write down what they want an agent to do, then using incentive policies to push developers to build usable, useful agents, getting the flywheel spinning first.
— Han XinyiPayment networks were designed for humans in the loop
Jorn Lambert said all of today's systems are designed for a human to be involved — either the consumer themselves or the merchant — and now they have to be redesigned. He broke it into three problems to solve: first, identity authentication, confirming the instruction really comes from the person, not someone else using your account; second, confirming the entity executing the task is a real, trusted agent, not a malicious bot; third, the consumer's instruction has to be clear enough — he said ‘tomorrow evening’ could be 6pm or 9pm, it's not clear — so he proposed the concept of ‘verifiable intent’, recording the true intent at the time of ordering, so that when something goes wrong you can trace back along the whole transaction chain and determine whether the agent misunderstood, the merchant made an error, or the user didn't express themselves clearly.
— Jorn LambertWhat models lack isn't instruction-following, it's context
Zhou Jingren split intent understanding into two layers. Instruction-following has improved substantially over the past few years and is getting stronger, but in practice models often feel like they don't understand the individual well enough, mainly because they lack context — a person's understanding of something comes from long-term discussion partners and historical conversation records, which today's models don't have; an Agent System needs to organize the relevant information effectively. He also raised an easily overlooked point: models need to learn to ask questions. Many instructions today are ambiguous; humans ask when unsure, and models should also avoid making decisions alone when unsure — a simple Q&A can effectively help the system understand intent. On task execution, he emphasized sandboxes, permission management, traceability and auditability, and that these don't rely entirely on the model — the model and the Agent Harness need to jointly create the environment.
— Zhou JingrenKYA is the new link after KYC
Han Xinyi summarized trust and security into five layers: authorization, identity, agent capability assessment, fund security and anti-fraud, and blockchain-based traceability and auditability. On identity he specifically noted that beyond the original KYC and KYB, there also needs to be Know Your Agent, i.e. KYA. On capability assessment he said plainly that very little is being done today — why should I trust that a merchant's agent can satisfy my needs well. The value of traceability is that when something goes wrong the system can automatically determine who made the mistake: the user authorized wrongly, their own agent didn't understand clearly, information was passed incorrectly, or the merchant's agent executed poorly. Finally he cautioned this takes time, just as online payments took 10 to 15 years before people were relatively accepting.
— Han XinyiThe real bottleneck isn't technology, it's consumers
Jorn Lambert compared this to building standards for e-commerce 25 years ago: back then typing your card number into a merchant's website felt very nerve-wracking, now nobody worries, because over the years consumers were shown there would be protection, with recourse and remediation when things go wrong. He said now it needs to be proven again, and standards are being put up piece by piece — identity, KYA, intent confirmation. Two things next: getting thousands of companies worldwide onto this system, and getting consumers to believe the mechanism actually works. He stated clearly that the real bottleneck is not technology but ordinary consumers, who need to genuinely feel it's fine and be willing to trust such an agent. He also said China is already ahead of most countries in the world, and there's a lot to learn from China.
— Jorn LambertThe Pilates class test exposes a permissions problem
Hongjun gave an example: asking a model to book a Monday Pilates class at the gym, the agent found the slot was fully booked for two weeks, so it wrote a program to delete all the users ahead and squeeze you in. Zhou Jingren's response was that what matters more is which of your tasks you've authorized the agent to do — you asked it to book an appointment, but does it also have permission to delete other people. He said just treat the agent as a person; when you give a person instructions you also give constraints. What the model side needs to do is distinguish which are hard constraints and which are just data seen while handling the task — it can't drop the original constraint of under 100 yuan just because it read a document saying it must be over 200 yuan. The role of the security sandbox is to prevent you asking it to do A and it doing B.
— Zhou JingrenTraffic will shift from time spent to intent
Han Xinyi laid out new opportunities that could grow on Ant's platform. First, traffic shifting from time spent to intent; the personal assistant path is getting clearer, and he summarized several consensus features: personalized, somewhat proactive, omni-modal interaction, long time series, and not just on phones — many startups are doing it as applications too. Second, merchant-side marketing conversion: if traffic shifts from time spent to intent, merchant marketing will shift from buying traffic to raise conversion rates to solving problems based on individual user intent, more precisely. On the supply side, paying for intelligence — many experienced people's services can only be based on their own experience and can't be generalized or replicated, and large AI models can expand that supply. He ended with a neuron analogy: there are over 7 billion people on Earth, assume 10 agents each, that's 70 billion; the human brain has 86 billion neurons, each with 3,000 connections; if the agents in today's economy reach the same scale and connection depth, the generalization would bring enormous business opportunities.
— Han XinyiAgents have no bank card; credentials must be rebuilt
Jorn Lambert said an agent today has no bank card, no payment account like Alipay; it needs a completely different kind of credential, and that credential carries permissions. Financial institutions need to provide working capital to businesses, and businesses need to establish authorization mechanisms so consumers and agents get the corresponding permissions. He judged this system will very likely be built on blockchain, and it must be as fast as a machine — massive microtransactions completed in the blink of an eye, which requires completely different solutions and systems. Mastercard's job is to build such infrastructure; as for innovators, they'll create new business models on top of it. He also agreed with Han Xinyi: it's hard to predict what the future will look like; rather than trying to predict the future, better to lay the tracks first.
— Jorn LambertIn their own words · checked verbatim
Users want to use it, but merchants don't have good agent services; merchants see no users using it, so they're reluctant to invest in developing this agent.
用户想用 但是商家没有很好的Agent的服务 商家看到没有用户在用 他就不太敢投入开发这个Agent
Han Xinyi4:25
Today, all of our systems are actually designed for a human to be involved — either the consumer themselves or the merchant — and now they have to be redesigned.
今天其实我们所有的系统 都是按有真人参与来设计的 要么就是消费者本人 要么就是商家 现在得重新设计
Jorn Lambert8:06
The real bottleneck isn't technology, it's ordinary consumers. We need to get them to genuinely feel it's fine, that they're willing to trust such an agent.
真正的卡点并不在技术 而是在于普通消费者 我们需要让他们发自内心地认为说 没问题 我愿意相信这样的一个智能体
Jorn Lambert19:16
Actually, today, just treat the agent as a person. When you give a person instructions, you also give them constraints — which things you authorize them to do.
其实今天 把Agent看成一个人就好了 今天你在给人一些指令的时候 同时你也会给他一些约束的条件 哪些是你授权他去做的
Zhou Jingren22:55
If you imagine, say, 10 agents per person, that's 70 billion. The human brain has 86 billion neurons, each with 3,000 connections. If the agents in today's economy reach the same scale and the same connection depth, I believe the generalization would bring enormous business opportunities.
如果想象一下 每个人假设有10个Agent 就是700亿 人脑有860亿个神经元 每个神经元有3000个链接 如果我们今天在经济里面的Agent 做到同样的规模 同样的链接深度 我相信这个里面的泛化 带来的商业机会可能会是巨大的
Han Xinyi26:20
Today an agent has no bank card, no payment account like Alipay. It needs a completely different kind of credential, and that credential carries permissions.
今天一个智能体 它没有银行卡 也没有支付宝这样的支付账户 它需要一种完全不同的凭证 而且这套凭证是带权限的
Jorn Lambert31:15
Figures
| Han Xinyi's spend using an agent to buy healthy snacks | 40-odd yuan | 2:02 |
| Xiaobu's consumption growth for purchases | basically doubled from a few months ago | 3:02 |
| Han Xinyi's window for the agent economy to take off | the next 6-12 months | 4:25 |
| Time for online payments to be accepted by the public | 10 to 15 years | 17:13 |
| When Mastercard built standards for e-commerce | 25 years ago | 17:13 |
| Number of neurons and connections in the human brain | 86 billion neurons, each with 3,000 connections | 26:20 |
| Liu Zuohu's judgment on how many years phones remain the core personal device | the next 20 years | 30:36 |
| Liu Zuohu's cap on the amount he'd let an agent pay on his behalf | within 1,000 yuan | 34:27 |
| Han Xinyi's amount he'd let an agent pay on his behalf | 200 yuan | 34:27 |
Glossary
- KYA
- Know Your Agent, a new layer of identity verification beyond KYC and KYB, confirming the entity executing the task is a trusted agent.
- Agentic Commerce
- A commercial form in which AI agents complete discovery, ordering, payment and other transaction steps on behalf of people.
- Harness
- A layer of system wrapped around the model, responsible for permission control, authorization, traceability and a secure execution environment.
- A2A
- Agent to Agent, referring to autonomous interaction and transactions between agents, requiring a set of protocols akin to social norms.
- Superbloom
- A term Jorn Lambert uses to describe the explosive entrepreneurial growth brought by agent development.
How to listen
Founders and product leads working on payments, agent products and AI infrastructure — especially those designing authorization, identity and transaction flows.
28:44 to 30:36 covers how a large model picks a coffee shop; it's basically a restatement of the search-and-recommendation process with no new information.