Good evening, professors and partners.
Thank you very much to Professor Zhu Haikun and Dr. Yang Guang for their sharing. The topic I will share today happens to bring together the key words from both of their presentations: currency and AI. Let us talk about how robots spend money.
Over the past two years, I have often shared with everyone on the CEIBS campus on compliance topics such as Web3 and digital assets. But discussing Web3 from the perspective of AI is a new attempt for me today as well. I would like to start with a recent experience of my own.
At first, I just wanted to ask a question
Some time ago, I wanted to buy some vitamin tablets, but I could not remember whether I should take B1 or B2, so I opened Ant Afu, an app under Alibaba, and wanted to have a chat with it.
What surprised me was that during the conversation, it identified my purchase intent, proactively recommended nearby pharmacies and vitamin products I could buy, and listed them by price and distance. The subsequent purchase process connected to Taobao Flash Sale; I followed the prompts to place the order and pay, and throughout the whole process I did not open Alipay separately. In about twenty minutes, the items were delivered.
When I first opened the app, I really just wanted to ask a question, but as we chatted, the items ended up being purchased. What struck me here was that the distance between consultation and consumption had suddenly shrunk. Looking further along these lines, AI applications such as Doubao and Qwen are also connecting to everyday services such as food ordering, shopping and hotel booking.
In the past, I always felt that AI payments and robot payments were still far away from us. But they seem to have already entered our lives quietly and unobtrusively; we just have not particularly noticed it.
From a business perspective, AI used to mainly answer questions and provide information; now it has begun to give us decision-making references and even participate in execution. Especially after this year's Spring Festival, agent tools such as OpenClaw became popular, and office assistants from various companies appeared one after another, giving a bit of the feel of a "battle of a hundred regiments."
Recently I have felt as if I were back in the state of a programmer, doing vibe coding every day, talking with AI while having it write code and build tools for me. The PPT, illustrations and many of the ideas used in today's sharing were also produced together with AI in this process. So let me make a joke here: if today's talk goes well, it is because AI did a good job; if it is not quite appropriate, it may be that we humans did not cooperate well.
Look at this AI-generated "The Last Supper." Everyone sits together and finishes the meal; who pays the bill?
In AI collaboration, a human raises a need, and AI selects models, calls tools and executes tasks. Who decides and who pays for the service consumption happening behind this? This is no longer a pure joke.
I now use Codex, and I also use the APIs of Kimi and DeepSeek. Having used them a lot, I have a very specific confusion: where exactly are all the tokens consumed every day—the unit of measurement used when a model processes text and other information—actually spent? Are they all spent on my tasks? Is the model service I purchased the same thing as what is actually being called?
Every time OpenAI releases a major model update, people discuss whether its capabilities have improved. As someone who uses it every day, I also have a very direct impression: why do these tokens burn so fast? Is it really working earnestly on my task, or is it repeatedly trying and repeatedly consuming tokens on the same problem? The backend records so much usage, but as the person paying, it is actually very hard for me to verify it item by item.
So many tokens are consumed every day, and the bot also has to call various skills and tools, but the consumption process still feels like a black box to me. I want to know who the money was paid to, what I bought, and how the task was actually completed—not just see that the account balance has decreased.
Going further, has the service I purchased been "watered down"? For example, I paid for a high-priced model, but the one actually doing the work is another, lower-cost model. Can I know the difference? This is no joke; it is a real thing. Some time ago, parts of OpenAI's and Claude's services experienced outages, yet some third-party APIs remained available as usual. So some people joked: you think Claude is doing the work—could it actually be "Doubao" working for you behind the scenes?
What I really want to ask is: do paying users have any way to verify which model was actually called, how many tokens were processed, and whether this money was truly spent on my tasks?
People used to say that once AI arrived, would Web3 be finished? My own experience is exactly the opposite: I have never hoped as much as I do now that blockchain's verifiable records could play a role in AI services.
Whoever influences decisions influences transactions
Beyond the execution process being opaque, we also encounter another kind of change: AI is guiding consumption decisions. Here are two examples, one is Doubao and the other is Meta.
Doubao's hotel channel fees have sparked a lot of discussion. According to public reports at the time, the software service fee plus payment processing fee for the relevant hotel orders came to a combined rate of approximately 12%. This refers to transaction fees for a specific hotel channel, not a uniform charge for all Doubao services; the platform also responded that hotel recommendations do not charge advertising fees and that merchants cannot pay to purchase recommendation rankings.
But from a commercial perspective, this is still a change worth paying attention to. We originally hoped to use AI to retrieve information faster and find services more accurately; by the transaction stage, it has become linked to merchants' customer acquisition costs and channel fees. The business that used to revolve around traffic and transactions in the traditional internet is beginning to enter AI conversations.
Going back to my example of buying vitamins, initially I may not have already decided to make a purchase. AI continuously identifies my questions and intentions and presents options, which may also push me to make a purchase decision. This is a very subtle shift.
Recently I watched a long interview with Professor Zeng Ming, and his discussion about intelligent business needing to accurately identify user intentions struck a chord with me. Where do intentions come from? A large part of them is hidden in the conversations between users and AI. For entrepreneurs and operators, what they need to think about going forward is how their products and services can enter these conversations, how they can be understood, and how they can enter the range of users' choices.
Overseas, Meta offers another example. In the past, we were familiar with its advertising business: ads were served based on user data, and merchants bought impressions and clicks. Now it has taken a step further, providing merchants with AI reception services such as Business AI.
After a user clicks an ad, they first communicate with the merchant's AI employee. This assistant answers questions, recommends colors and styles, helps with appointments, and completes purchases based on the product and business materials provided by the merchant, transferring to a human only when necessary. The process that previously lay between "seeing the ad" and "having someone receive the customer" is now beginning to be connected by AI.
So in my view, Meta's role is extending toward sales services. It not only influences what customers see, but also begins to participate in how customers learn about products and make decisions after they see them. The boundary between internet advertising and customer operations may therefore change.
Whoever can influence users' decisions can influence transactions. Payment happens later, but the power that determines whether a transaction will occur, and with whom, may already have been at work in the earlier conversation.
This force is also coming from another direction: payment companies are beginning to move toward the front end of transactions. A typical example is Stripe's announcement of an acquisition of the model routing platform OpenRouter, with media reports putting the amount at USD 7 billion. What interests me more than the amount is why a payment company would take such a strong interest in model invocation.
My understanding is that in the AI era, enterprises will spend a considerable portion of their budgets on AI services. In the past, we paid employee salaries and bought office software; going forward, we will also pay model and tool fees for the process of AI work. How much AI employees actually spend will also become an account that operators need to manage.
Payment companies are already providing services for how a company's money comes in; they also want to provide solutions for how a company's money goes out, especially the ongoing expenses of AI invocation. This explains why, in the course of serving cross-border payment clients, we see some companies beginning to pay attention to and build out API-related business.
Moving backward from traffic companies means identifying intent, influencing decisions, and facilitating transactions; moving forward from payment companies means entering the process by which enterprises use services and incur expenses. This is a set of very interesting commercial moves I have observed so far in the AI payments track.
Letting go is uneasy, but asking every time does not work either
When AI starts doing things and spending money on our behalf, can we really feel at ease?
A story has circulated online that sounds much like a joke: a user asked AI to help book a restaurant, and the AI generated a convincing-looking booking confirmation. The user took the confirmation to the restaurant, but the merchant said there was no such order. "Then go eat with your AI." Whether the story actually happened is not something to conclude for now, but everyone can understand the punchline: AI saying something is done does not mean it has been done in the real world.
If it were just a joke, we could laugh it off. But when similar issues reach the courtroom, the matter becomes serious.
In a case in Hangzhou, a user asked an AI for information about university admissions and received an incorrect answer; in a subsequent conversation, a statement expressing willingness to pay 100,000 yuan also appeared. The user sued the developer on that basis, but the claim was ultimately dismissed by the court. A promise generated by an AI in a chat cannot be directly treated as a valid compensation commitment already made by the platform.
This needs to be stated precisely: it does not mean that an AI service provider bears no liability at all simply because it warns that errors are possible. The specific determination still depends on the service obligations, fault, loss and causation. An AI itself does not have the legal person status to independently bear civil liability, nor does that mean the person who uses or operates it is thereby exempt from liability.
There is another case in Tongzhou, Beijing. A lawyer asked an AI to assist in drafting litigation materials; the reference cases submitted appeared very similar to the case to be heard, but upon verification they were found to contain fabricated content inconsistent with real judgments. The court criticized this in its judgment.
These matters are not far from us. If even information and judgments can be wrong, how can we feel at ease entrusting our assets and money entirely to AI?
I currently use API services such as DeepSeek. Every night before going to sleep, I often assign the AI a task: what you need to complete tonight. After assigning it, I still have to check the account balance, and I do not dare to top it up too much. If it repeatedly tries the same issue and repeatedly makes calls, the bill may become unbearable.
As you can see, merely calling a model already requires me to control the budget and manage risk, let alone other spending. So at this stage, it is indeed difficult for us to let AI run completely freely.
Conversely, if it comes to ask me every time, that is also maddening. Calling a tool once and spending a little money requires asking whether I agree; switching to another service requires asking again. It feels much like a manager opening the OA system every day and facing a pile of pending approvals: you may not have time to look carefully, but it is just waiting for you to click "approve."
Complete delegation is not acceptable, and taking everything back is also not acceptable. Where is the balance in between? I do not think there is already a unified answer at this stage, because each person's degree of trust in AI, the risk of the task and the loss they can bear are all different.
On the surface, we are discussing internet payments and the development stage of AI. Having come this far, I increasingly feel that the core issue is still trust.
In the past we used cash, then bank cards and online banking, and then Alipay and WeChat Pay; the forms have kept changing. But the questions remain the same: whose money is this? Who has the right to spend it? Has the money arrived? What if it is spent wrongly, and who bears liability? In the AI era, we still cannot escape these questions.
Let us think one step further: if AI participates more and more in economic activities in the future, will we need to revisit its legal position?
A very important institutional invention of human commercial society is the legal person, such as a company. It is not a natural person, but the law confers on it a specific legal person status. Could it be possible in the future that AI also obtains some kind of legal status, holds accounts, receives credit and is subject to corresponding constraints? Companies can be placed on the list of dishonest debtors; could similar arrangements for AI emerge in the future?
If AI causes a problem, will it also have to "spend a few years in the penalty box"? Of course, this is a discussion of future possibilities, not a statement that today's law already recognizes such an entity. The real question it raises is this: if AI increasingly acts independently, how should responsibility and constraints keep pace?
Returning to today's reality, the existing financial system does not yet treat AI as an independent transaction entity. Can you have an AI open a bank account in its own name? It is neither a natural person nor a company established in accordance with law; at this stage, it still needs to rely on the accounts and authorizations of an individual or an enterprise.
If I simply top up an amount for it and let it use that amount within a certain system, this still resembles topping up an application account in terms of the control relationship. The fact that it can perform tasks does not mean it already has an independent economic and legal identity. How intelligent agents will develop in the future is a matter worth continued discussion.
The same applies to authority limits. We might give a child one hundred yuan in pocket money, but we would not simply give him one million to spend as he pleases. How much money an AI may hold, how it may use that money, and in what circumstances it must come back and ask still need to be arranged. Existing law is not entirely without rules, but how those rules should adapt to AI's autonomous execution still leaves many specific issues to be resolved.
There is also the question of ownership of outputs from human-AI collaboration. A great deal of AI was involved in this PPT today, and I joked that I was like a "human microphone," bringing the content to the venue. So who is the author of this content, which parts can obtain protection, and how should human creative contribution be viewed? Nor can this simply be reduced to "whoever presses the generate button owns everything."
So I believe that a very important proposition in future AI competition is how to build trust in AI. It involves technology, but also transaction rules and legal liability.
There are already some different explorations. For example, x402, promoted by Coinbase, uses the HTTP 402 payment status code so that a program accessing paid resources knows how to pay as required and then obtain the service. Everyone is familiar with 404, which means the resource cannot be found; 402 was originally reserved for payment requirements, and x402 designs a specific payment process on that basis.
One analogy is the internet television at home: after watching a segment, you are prompted that subsequent content requires payment. But what AI services need is for programs to handle payment requirements and credentials, rather than having a person scan a QR code every time.
Google's AP2—the Agent Payments Protocol—addresses a different question: was this money really spent with your permission for the AI to spend it? It turns the purchase content, amount, and payment authorization approved by the user into credentials that can be verified with cryptographic signatures. For example, if you allow the AI to buy a plane ticket on a certain date within a certain price range, the merchant and payment institution cannot simply listen to the AI say "the owner agreed"; they must also be able to check the corresponding authorization record. If a dispute does arise, there is also material to compare: what you agreed to at the time and what was ultimately purchased. This protocol does not limit payments to stablecoins; payment methods such as bank cards can also participate. It provides a basis for verifying authorization, while specific liability still needs to be determined in light of the transaction rules.
Amazon AWS's AgentCore Payments is more like equipping AI with a "wallet that can be kept under control." Developers connect wallet services such as Coinbase or Stripe Privy, then set a spending budget and validity period for a task; within that scope, the AI purchases paid interfaces, tools, or content. At the time of payment, the infrastructure checks the limit, and calls and expenditures also leave records. In this way, there is no need to come back and ask a person to click approve every time a small amount is spent, nor is there a need to hand all wallet permissions to the model. This capability has been officially launched, but which wallets can be connected and in which regions it can be used still depend on the support of the specific service.
In my view, although these solutions have different emphases, they still revolve around the user authorizing the AI, making execution traceable, and keeping permissions under control. They are still at the exploratory stage and should not be understood as having already formed a set of industry-wide common rules applicable to all scenarios.
Why I am optimistic about the combination of crypto payments and AI
Following the keyword of trust, let us continue the reasoning: who is best positioned to solve AI trust?
What comes to mind is blockchain.
The October 31, 2015 issue of The Economist discussed blockchain under the title "The Trust Machine," which is often rendered in Chinese as "trust machine" or "machine that manufactures trust." Today, I think it has a very concrete use case: where transactions occur among several AI systems and service providers that are not familiar with one another, it is not enough for each to simply say "I authorized this" or "I paid." It is better to have a record that everyone can jointly verify and that cannot easily be altered unilaterally after the fact. Of course, recording something does not mean the content is inherently true, but at least it can help us link authorization, payment and execution. Could a technology discussed more than a decade ago have an opportunity to play a greater role now that AI is beginning to handle matters and spend money on its own?
I very much agree with one assessment:cryptoassetsappeared at birth to be for human use, but what may truly enable their large-scale use is perhaps machines.
For domestic consumers, Alipay and WeChat Pay are already very convenient. Asking a person to relearn wallets, private keys and network selection for everyday payments entails high migration and learning costs. This is also why a new payment method at the technical level does not mean everyone will switch to it.
The situation is different for machines. AI does not need to learn an interface the way a person does; it can directly call APIs and execute according to rules. Many steps that are troublesome for humans can be completed automatically between programs.
So I borrowed Trump's meme: "No one knows more about it than me." Here, what we want to ask is whether crypto assets better understand the needs of AI payments.
Over the past few years, we have handled quite a few businesses involving cross-border payments. A large part of my interest in crypto payments comes from these specific scenarios.
For example, small-value cross-border payments may occur in live-streaming tips, game top-ups, or a very granular digital service. The payment amount itself is small, but fixed fees and onboarding costs along the traditional path may make the transaction uneconomical. Certain on-chain payment solutions have advantages in such scenarios, provided that the selected network fees are appropriate, both payer and payee can use them, and the business itself complies with local requirements.
Imagine a scenario: an overseas user wants to tip a game streamer one US dollar. What would the cross-border transaction fee be? If the payment cost exceeds the value of the service, the transaction is unlikely to occur. As AI continuously calls small-value services, this tension will become even more pronounced. This is also why I am particularly optimistic about small-value, high-frequency scenarios. Specific fees must be calculated based on the network and the payment and collection path; a particular low fee rate cannot be treated as a uniform level for all crypto payments.
Another advantage is programmability. Applications such as on-chain lending and fund allocation are all leveraging this capability. The digital renminbi applications discussed earlier today also involve the idea of writing conditions into payment arrangements. When money can be executed according to pre-set conditions, transactions gain greater flexibility. Of course, whether the conditions have actually been met in reality still requires reliable judgment.
There is also a broader backdrop: the openness of the internet.
I previously spoke with a colleague who had returned from the United States. He said that in the United States, many small shops and startup brands have their own websites, and if you want to find nearby services, you can often find them directly online. But back in China, at first he felt that some merchants' information could not be found on public web pages. Only later did he realize that much content and many services are within official accounts, applications and their respective platform ecosystems.
Over the past decade or so, a great deal of good content has been produced on the internet, but different platforms are not fully interconnected. I call this state "electronic fences." In the AI era, when you ask an assistant to find information or complete tasks, it will also encounter these access and authorization restrictions.
Therefore, I judge thatAI will bring a force that pushes the internet to reopen.
For assistants to call services across platforms, information and interfaces need to be accessible in appropriate ways. This will not happen automatically; platforms, content providers and users still need to find new arrangements for cooperation and charging.
Likewise, if AI seeks services globally, its payment and collection system can hardly remain confined forever within the ecosystem of a single application. From this perspective, I am optimistic about the role of cryptocurrencies and blockchain in cross-network value transfer. AI drives service collaboration, while blockchain provides a degree of verifiable transaction and execution capability; in my view, these two have strong potential for combination.
This is an industry judgment; it does not mean that AI can only use cryptocurrencies, nor does it mean that domestic businesses can directly copy overseas solutions. Technological potential, specific costs and applicable rules still need to be examined within their respective business scenarios.
To explore AI payments, first do three things well
Much of what we discussed today is still very cutting-edge and very early-stage. If you are an AI entrepreneur or a payment industry entrepreneur and want to explore this direction, I would offer three practical suggestions.
One is control authority. From the user's perspective, my own approach is quite simple: start by topping up a smaller amount. From the entrepreneur's perspective, you cannot rely solely on users being careful. What may be purchased, to whom payments may be made, how much may be spent per transaction and in aggregate, and when authorization expires should all be built into product settings and authorization terms; exceeding the limit or changing the payee should require re-confirmation, and users should be able to revoke authorization at any time. This is especially true for enterprise accounts: who has the authority to authorize AI should also correspond to internal approval authority, and an ordinary operational account must not be given the ability to make unlimited expenditures.
Some of the crypto investment and trading clients we have served in the past did not necessarily suffer losses because of business judgment; losses may also come from hacker attacks. Once AI participates in operations, excessive permissions and exposed interfaces can likewise place assets at risk. Security must therefore be considered from the product design stage.
The other is to properly allocate responsibility after errors occur. An AI purchasing the wrong item, a platform deducting the wrong amount, and a merchant failing to deliver are three different problems. The agreement should clearly state who handles complaints, how refunds and compensation are made, and how the platform and merchant share liability, while retaining authorization, order, invocation and deduction records so that the process can be reconstructed if something goes wrong. For standard-form clauses involving disclaimers and limitations of liability, proper notice and explanation must also be provided; one cannot use a single sentence—"AI-generated, for reference only"—to shift all responsibility that should be borne onto users.
Another is protecting data, keys and account information. For convenience, stuffing private keys, payment passwords or all customer materials into a conversation creates significant risk. A more appropriate approach is to isolate keys from the model and let AI handle matters through restricted interfaces; provide only the data necessary to complete the task, and control retention periods and access permissions. If external models are invoked, it is also necessary to clearly understand to whom the data is provided, whether it will be retained or used for training, and whether it involves cross-border transfer of personal information. Asset protection and privacy protection must be arranged at the time the service is integrated.
That is some of the observations I would like to share with you from the perspectives of internet development and legal practice.
If I were to reduce the content of these thirty minutes to one point, what I want to express is this:
- In the AI era, whoever can win trust can influence business decisions;
- whoever can influence business decisions can influence commercial transactions.
What we are discussing today is not only the future of transactions, but also how trust in this society should be established and achieved. All right, thank you everyone.
Live Q&A: What currency will AI ultimately use?
Question: AI payments will ultimately have to face acquiring and regulation. What currency will AI ultimately use in the future—stablecoins, digital renminbi, or some other form? How do you see stablecoins entering the actual payment and collection system?
Liu Honglin: My personal view is that AI has no human-like preference for the technical form of money. AI does not care which conceptual camp the currency it uses belongs to; what it cares about is whether payment can be completed. Whether it is a digital form of traditional fiat currency, a central bank digital currency, or a stablecoin, the key for AI is whether it can complete the task within the scope of authorization. It does not naturally favor or reject a currency merely because that currency comes from a centralized or decentralized system.
My second view is that I am optimistic about the trend of crypto payments gradually connecting with compliant financial account systems. As long as enterprises engage in cross-border trade, there will be demand for global fund settlement. Even without considering AI, the demand from cross-border trade and digital services for more efficient settlement methods will also drive this direction to continue developing. However, a growth trend does not mean that a particular specific model has already obtained permission. Discussing such connection does not mean that mainland individual or enterprise accounts must directly receive and pay in cryptocurrencies.
We can pay attention to how compliant institutions in Hong Kong and other overseas markets undertake payment, exchange and settlement services within the scope permitted locally, and then connect with the traditional fiat currency system. What is discussed here is a possible business structure; whether it can be done still depends on the entity, the transaction, the source of funds and the regulatory requirements of various jurisdictions. Merely setting up one more company overseas, or transferring funds through one more hand, does not automatically resolve compliance issues.
