Author: Liu Honglin
Good evening, teachers and partners.
Thank you very much to Professor Zhu Haikun and Dr. Yang Guang for sharing. The topic I shared today can just put the key words shared by the two together: currency, AI. Let’s talk about how robots spend money.
In the past two years, I have often shared compliance topics such as Web3 and digital assets with everyone on the CEIBS campus. But talking about Web3 from the perspective of AI is a new attempt for me today. I want to start with a recent experience I had.
At first, I just wanted to ask a question
Some time ago, I wanted to buy some vitamin tablets, but I couldn’t remember whether I should supplement B1 or B2, so I opened the Ant Afu brand owned by Ali and wanted to have a chat with it.
What surprised me was that during the conversation, it recognized my purchase intention and proactively recommended pharmacies near me and vitamin products available for purchase, listed by price and distance. The subsequent purchase process was connected to the Taobao flash sale. I followed the prompts to place the order and pay. I did not open Alipay separately during the whole process. The items were delivered in about 20 minutes.
When I first opened this app, I really just wanted to ask a question, but while chatting, I bought something. What strikes me here is that the distance between consultation and consumption has suddenly shortened. Looking along this line of thinking, AI applications such as Doubao and Qianwen are also accessing daily services such as ordering meals, shopping, and hotel reservations.
In the past, I always felt that AI payment and robot payment were still far away from us. But it seems to have entered our lives silently, but we are not particularly aware of it.
From a business perspective, AI originally mainly answered questions and provided information, but now it has begun to provide us with decision-making reference and even participate in execution. Especially after the Spring Festival this year, intelligent tools such as OpenClaw have become popular, and various office assistants have also appeared one after another. It feels like a "hundred regiments of war".
I feel like I'm back to being a programmer recently. I'm doing vibe coding every day, talking to AI and letting it help me write code and build tools. The PPT, pictures and many creative ideas shared today were also created together with AI in this process. So I made a joke at the scene: If the speech is good today, it is because the AI is doing a good job; if the speech is inappropriate, it may be that humans are not cooperating well.
Look at this AI-generated "The Last Supper". Everyone is sitting together and eating. Who will pay the bill?
In AI collaboration, humans put forward a demand, and AI selects models, calls tools, and performs tasks. Who decides and pays for the service consumption that occurs behind the scenes? This is no longer a pure joke.
I now use Codex, as well as the APIs of Kimi and DeepSeek. After using it a lot, there is a very specific confusion: Where are these Tokens consumed every day - the unit of measurement used by the model when processing text and other information - spent? Was it all used on my mission? Are the model services I purchased and the ones actually called the same thing?
Every time OpenAI’s model is updated, everyone discusses whether their ability has become stronger. As someone who uses it every day, I have a very direct feeling: Why does this token burn so fast? Is it completing my task seriously, or is it trying and consuming the same problem over and over again? There are so many usages recorded in the background. As the person making the payment, it is actually difficult for me to check them one by one.
With so many tokens consumed every day, the robot has to use various skills and tools, but the consumption process is still like a black box to me. I want to know who the money was paid to, what I bought, and how the task was accomplished, rather than just seeing a decrease in my account balance.
Furthermore, is the service I purchased “watered down”? For example, if I pay for a high-priced model, but the actual work is done by another lower-cost model, can I know the difference? This is not a joke, but a real thing. Some time ago, some services of OpenAI and Claude failed, but some third-party APIs are still available. So someone made a joke, if you thought it was Claude doing the work, could it be that "Mr. Dou Bao" was actually doing the work for you behind the scenes?
What I really want to ask is, is there any way for paying users to verify: which model was called, how many tokens were processed, and whether the money was actually used for my task?
Someone once said that with the arrival of AI, will Web3 become obsolete? My personal experience is that I have never hoped that the verifiable records of the blockchain can play a role in AI services as much as I do now.
Whoever affects decisions affects transactions
In addition to unclear visibility of the execution process, we will also encounter another change: AI is guiding consumption decisions. Here are two examples, one is Beanbao and the other is Meta.
The fees charged by the Doubao hotel channel have caused a lot of discussion. According to public reports at the time, the comprehensive rate for the software service fee plus payment handling fee for related hotel orders was approximately 12%. What we are talking about here is the transaction fee for specific hotel channels, not a unified fee for all beanbao services; the platform has also responded that hotel recommendations do not charge advertising fees, and merchants cannot pay to purchase recommended rankings.
But from a business perspective, this is still a change worth watching. We originally hoped to use AI to retrieve information faster and find services more accurately. But when it comes to closing the transaction, it is linked to the merchant's customer acquisition cost and channel fees. Businesses that revolve around traffic and transactions on the traditional Internet are beginning to enter AI dialogue.
Going back to the example of me buying vitamins, I may not have decided to consume them initially. AI constantly recognizes my questions and intentions, gives me options, and may push me to make purchasing decisions. This is a very subtle change.
Recently, I watched a long interview with Professor Zeng Ming. The discussion about the need for intelligent business to accurately identify user intentions moved me very much. Where does the intention come from? A large part of it is hidden in the conversation between users and AI. For entrepreneurs and operators, what they need to think about in the future is how their products and services enter these conversations, how they are understood, and how they enter the user's choice range.
Overseas Meta provides another example. In the past, we were familiar with its advertising business: advertising based on user data, and merchants buying exposure and clicks. Now, it has taken a step further to provide merchants with AI reception services such as Business AI.
After the user clicks on the ad, he first communicates with the merchant’s AI staff. Based on the products and business information provided by the merchant, this assistant answers questions, recommends colors and styles, assists in making appointments, completes purchases, and only transfers to manual intervention when needed. It turns out that the process between "seeing the advertisement" and "being received" has also begun to be taken over by AI.
So in my opinion, Meta's role is extending to selling services. It not only affects what customers see, but also begins to participate in how customers understand the product and make decisions after seeing it. The boundaries between Internet advertising and customer operations may change as a result.
Whoever can influence users' decisions can influence transactions. The payment happens later, but the forces that determine whether and to whom the transaction occurs may have already been at work in the previous conversation.
This force is also coming from the other direction: payment companies are also starting to move towards the front end of transactions. A typical example is Stripe's announcement to acquire the model routing platform OpenRouter, with media reports that the amount was US$7 billion. What interests me more than the amount is why a payment company would be so interested in model calling?
My understanding is that in the AI era, companies will spend a considerable part of their budgets on AI services. In the past, we paid employees salaries and bought office software. In the future, we will also have to pay for models and tools for the AI work process. How much money AI employees spend will also become an account that operators need to manage.
How a company collects money, payment companies are already providing services; how a company spends money, especially the ongoing expenditure on AI calls, it also hopes to provide solutions. This can explain why in the process of serving cross-border payment customers, we will see some companies begin to pay attention to and layout API-related businesses.
Going back from the traffic company is to identify intentions, influencing decisions, and facilitating transactions; going forward from the payment company is to enter the process of enterprises using services and generating expenditures. This is a set of very interesting business actions in the AI payment track that I have observed so far.
Don’t worry about letting go, even if you ask every time
When AI starts to do things and spend money for us, can we really rest assured?
There is a story circulating on the Internet that sounds like a joke: a user asked AI to help make a restaurant reservation, and the AI generated a decent notification of successful reservation. The user went to the store with the notification, but the merchant said there was no such order at all. "Then you go to AI to eat it." Whether the story has happened or not is not yet determined, but everyone can understand the joke: AI says it has been done, but it does not mean that it has been done in the real world.
If it was just a joke, we would just laugh and pass it by. But when similar issues go to court, things get serious.
In a case in Hangzhou, a user asked AI for college application information and received an incorrect answer. In the subsequent conversation, a statement was made that he was willing to pay 100,000 yuan in compensation. The user sued the development company based on this, and the claim was eventually rejected by the court. The promises generated by AI in the chat cannot be directly regarded as effective compensation promises made by the platform.
We need to be precise here: this does not mean that the AI service provider does not have to bear any responsibility as long as it prompts "errors may occur". The details also depend on service obligations, fault, loss and causation. AI itself does not have the status of an independent subject that bears civil liability, nor does it mean that those who use it or operate it are exempt from liability.
There is another case in Tongzhou, Beijing. The lawyer asked AI to assist in drafting litigation materials, and the reference case submitted looked very similar to the case to be tried. However, after verification, it was found that there were fictitious contents that were inconsistent with the real judgment. The court criticized this in its judgment.
These things are not far away from us. Even information and judgment may be wrong, so how can we safely hand over assets and money to AI?
I now use API services such as DeepSeek. Before going to bed every night, I often assign a task to the AI: What are you going to accomplish tonight? After arranging it, I still have to check the account limit, and I don’t dare to top up too much. If it tries and calls again and again on the same issue, the bills may become unbearable.
You see, just by calling the model, I already have to control the limit and risk, not to mention other consumption. So at this stage, it is really difficult for us to let AI completely let go and do its thing.
But on the other hand, if you ask me every time, it will be very frustrating. Call a tool once, spend a little money, and ask me if I agree; change a service, and ask again. This feeling is similar to when managers open OA every day and face a lot of pending approval items: you may not have time to look at them carefully, but they are just waiting for you to click "Agree".
It's not possible to let go 100%, and it's not possible to take it back 100%. Where is the balance in the middle? I don’t think there is a unified answer at this stage, because everyone’s level of trust in AI, the risk of the task, and the loss they can bear are different.
On the surface, we are discussing the development stage of Internet payment and AI. At this point, I feel more and more that the core is still trust.
In the past, cash was used, then bank cards, online banking, and then Alipay and WeChat payment were used. The forms have been changing. But the question is always: Whose money is it? Who has the right to spend? Has the money arrived? What should I do if I spend the wrong money? Who will bear the responsibility? In the AI era, we still cannot escape these problems.
Think one step further: If AI becomes more and more involved in economic activities in the future, will we need to re-discuss its legal position?
A very important institutional invention in human business society is the legal person such as a company. It is not a natural person, but the law gives it specific subject qualifications. Is it possible that in the future, AI will also obtain some kind of legal identity, have an account, accept credit, and be subject to corresponding restrictions? Will the company be included in the list of dishonest people? Will there be similar arrangements for AI in the future?
If something happens to AI, will it be locked up in a dark room for a few years? This is of course discussing future possibilities, not that today’s law has recognized such a subject. The question it really raises is: if AI acts more and more independently, how will responsibilities and constraints keep up?
Returning to today’s reality, the existing financial system has not yet regarded AI as an independent trading entity. Can you ask an AI to open a bank account on its own behalf? It is neither a natural person nor a company established according to law. At this stage, it still needs to return to the personal or corporate account and authorization.
If I just recharge it with money and let it be used in a certain system, this is still similar in terms of control relationship to recharging an application account. Being able to perform tasks does not mean it has independent economic and legal identity. How intelligence will develop in the future is worthy of continued discussion.
The same goes for permissions. We may give a child a hundred yuan pocket money, but we will not directly give a million yuan and let him spend it as he pleases. How much money can the AI put in its hands, how can it be used, and under what circumstances must it come back to ask, and arrangements still need to be made. Existing laws are not completely without rules, but there are still many specific issues to be resolved on how these rules adapt to the autonomous execution of AI.
There is also attribution of the results of human-machine collaboration. Today's PPT has a lot of AI participation. I also joked that I was like a "human microphone" to bring the content to the scene. So who is the author of these contents, what parts can be protected, and how should people's creative contributions be treated? It cannot simply be reduced to "Whoever presses the generate button owns the ownership."
So I think a very important proposition in future AI competition is how to build trust in AI. It involves technology, but also trading rules and legal responsibilities.
Now there have been some different explorations. For example, x402 promoted by Coinbase uses the HTTP 402 payment status code to allow programs to know how to pay as required when accessing paid resources and then obtain services. Everyone is familiar with 404, which means that the resource cannot be found; 402 was originally reserved for payment needs, and x402 designed a specific payment process on this basis.
It can be compared to the Internet TV at home: after watching a section, you will be prompted to pay for subsequent content. But what the AI service needs is for the program to process payment requirements and vouchers, rather than asking people to scan a QR code every time.
The AP2 - Agent Payment Protocol - proposed by Google solves another problem: Is this money really what you allowed the AI to spend? It converts the purchase content, amount and payment authorization approved by the user into a certificate that can be verified with cryptographic signatures. For example, if you allow AI to buy a ticket on a certain date and within a certain price range, merchants and payment institutions cannot just listen to the AI saying "the owner agrees", but they must also check the corresponding authorization records. If a dispute does arise, there are materials for comparison: What did you agree to at the time and what did you buy in the end? This agreement does not limit the use of stablecoins, and payment methods such as bank cards can also participate. It provides a basis for checking authorization, and specific responsibilities must still be judged in conjunction with transaction rules.
Amazon AWS's AgentCore Payments is more like a "manageable wallet" for AI. Developers connect to wallet services such as Coinbase or Stripe Privy, and then set the expenditure budget and validity period of the task. AI purchases charging interfaces, tools or content within this range; the infrastructure checks the amount during payment, and records of calls and expenditures are also left. In this way, you don't need to come back to find someone to agree every time you spend a small amount of money, and you don't need to give all the permissions of the wallet to the model. This capability has been officially launched, but which wallets can be connected and in which regions it can be used depend on specific service support.
In my opinion, although these solutions have different focuses, they still revolve around people authorizing AI, making execution traceable, and controlling permissions. They are still in the exploratory stage, and they cannot be understood as having formed a set of general industry rules applicable to all scenarios.
Why I’m bullish on the combination of crypto payments and AI
Following the key word trust, we will continue to deduce, who is the best solution to 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 translated into Chinese as "trust machine" or "machine that creates trust". Today, I think it has a very specific use scenario: when there is a transaction between several AIs and service providers who are unfamiliar with each other, they cannot just rely on each of them to say "I authorized it" and "I paid it." It is best to have a record that everyone can check together and that cannot be easily changed by one party afterwards. Of course, writing it down does not mean that the content is naturally true, but it can at least help us connect the authorization, payment and execution processes. Will the technology discussed more than ten years ago have the opportunity to play a greater role when AI starts to do things and spend money on its own?
I strongly agree with one judgment: when crypto assets were born, they seemed to be for human use, but what is really possible for them to be used on a large scale may be machines.
For domestic consumers, Alipay and WeChat payment are already very convenient. For a person to relearn wallets, private keys, and network options for daily payments, migration and learning costs are high. This is why, just because there is a new payment method in technology, it does not mean that everyone will switch to it.
The situation is different for machines. AI does not need to learn an interface like humans do. It can directly call the interface and execute according to the rules. Many steps that are troublesome for humans can be completed automatically between programs.
So I borrowed Trump’s joke: “No one knows better than me.” Here, what we want to ask is, do encrypted assets better understand the needs of AI payment?
In the past few years, we have been exposed to many businesses involving cross-border payments. A large part of my interest in crypto payments comes from these specific scenarios.
For example, cross-border micro-payments may occur in live streaming rewards, game recharges, or a very detailed digital service. The payment itself is small, but the fixed fees and access costs on the traditional route may make the deal uneconomical. Certain on-chain payment solutions have advantages in this type of scenario, provided that the chosen network fee is appropriate, both the payer and the payee can use it, and the business itself meets local requirements.
Imagine a scenario where an overseas user wants to reward a game anchor with one dollar. How much will his cross-border handling fee be? If the cost of payment exceeds the value of the service, it will be difficult for the transaction to occur. AI’s continuous invocation of small-amount services will make this contradiction more prominent. This is why I am particularly optimistic about small-amount, high-frequency scenarios. Specific fees must be calculated based on the network and payment path, and a low rate cannot be regarded as a uniform level for all crypto payments.
Another advantage is programmability. Applications such as on-chain lending and fund distribution are all taking advantage of this capability. The digital renminbi application discussed earlier today also involves the idea of writing conditions into payment arrangements. Currencies can be executed according to preset conditions, giving trading more flexibility. Of course, whether the conditions are met in reality still requires reliable judgment.
There is also a larger background, which is the openness of the Internet.
I talked with a colleague who came back from the United States before. He said that in the United States, many small stores and entrepreneurial brands have their own websites. If you want to find nearby services, you can often find them directly online. But when he returned to China, he would initially wonder why some merchant information could not be found on public web pages. Only later did I discover that a lot of content and services were in official accounts, applications, and their respective platform ecosystems.
In the past decade or so, a large amount of good content has been produced on the Internet, but different platforms are not fully interoperable. I call this state "electronic fence." In the AI era, if you ask your assistant to find information and complete tasks, it will also encounter these access and authorization restrictions.
Therefore, I judge that AI will bring a force that promotes the reopening of the Internet.
For the assistant to call services across platforms, information and interfaces need to be accessed in an appropriate way. It does not happen automatically. It also requires platforms, content parties and users to find new arrangements for cooperation and charging.
Similarly, if AI is looking for services around the world, its payment system will be difficult to be limited to the ecosystem of a certain application forever. From this perspective, I am bullish on the role of cryptocurrencies and blockchain in transferring value across networks. AI promotes service collaboration, and blockchain provides some verifiable transaction and execution capabilities. In my opinion, there is strong room for integration between these two things.
This is an industry judgment and does not mean that AI can only use cryptocurrency, nor does it mean that domestic business can directly copy overseas solutions. Technical potential, specific costs and applicable rules still need to be viewed in their respective business scenarios.
To explore AI payment, do three things first
Many of the things discussed today are still very cutting-edge and very early. If you are an AI entrepreneur or an entrepreneur in the payment industry and want to explore this direction, I would like to make three suggestions from a practical perspective.
One is control permissions. From a user's perspective, my approach is very simple: charge less first. From the perspective of an entrepreneur, you can’t just rely on users to be careful. What is allowed to be purchased, to whom to pay, how much can be spent in a single transaction and cumulatively, and when the authorization expires, should all fall into the product settings and authorization terms; if the limit is exceeded or the recipient of payment is changed, the user must reconfirm, and the user must be able to withdraw the authorization at any time. Especially for corporate accounts, whoever has the authority to authorize AI must also correspond to the internal approval authority. An ordinary operating account cannot be allowed to obtain unlimited spending capabilities.
For some of the crypto investment and trading clients we have served in the past, losses may not come from business judgment, but may also come from hacker attacks. After AI participates in the operation, the authority is expanded and the interface is exposed, which may also put assets at risk. Therefore, safety must be considered during product design.
Another is to arrange the responsibilities after mistakes are made. The AI buys incorrectly, the platform deducts incorrectly, and the merchant fails to deliver. These are three different problems. The agreement should clearly state who will handle complaints, how refunds and compensation will be made, and how the platform and merchants will share responsibilities. At the same time, authorization, order, transfer and deduction records should be kept so that there is a way to restore the process if something goes wrong. For format clauses involving exemptions and limitations of liability, it is necessary to provide prompts and explanations. Do not use the sentence "AI generated, for reference only" to push all the responsibilities that should be borne to the user.
Another is protecting data, keys and account information. For the sake of convenience, it is very risky to stuff private keys, payment passwords or all customer information into the conversation. A more appropriate approach is to isolate the key from the model and let the AI work through a restricted interface; whatever data is needed, only give it the parts necessary to complete the task, and control the retention period and access rights. If you call an external model, you must also check to whom the data will be handed over, whether it will be retained or used for training, and whether personal information will be exported abroad. Asset protection and privacy protection must be arranged when accessing services.
The above are some observations I have shared with you from the perspective of Internet development and legal practice.
If these thirty minutes of content are sorted into a core context, what I want to express is:
In the AI era, whoever can win trust can influence business decisions;
Whoever can influence business decisions can influence business 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. OK, thank you all.
Live Q&A: What currency will AI eventually use?
Question: AI payment will eventually face acquiring and supervision. What currency will AI use in the future? Is it a stable currency, digital renminbi, or other forms? What do you think about how stablecoins can enter the actual payment system?
Liu Honglin: My personal point of view is that AI itself does not have the same preference as humans for the technical form of currency. AI doesn't care about which currency it uses conceptually, it cares about whether it can complete the payment. Whether it is the digital form of traditional legal currency, central bank digital currency, or stable currency, the key is whether it can complete the task within the scope of authorization. It does not naturally like or reject a currency just because it comes from a centralized or decentralized system.
The second point of view is that I am optimistic about the gradual connection between encrypted payment and compliant financial account systems. As long as companies are doing cross-border trade, there will be a need for global fund settlement. Even without considering AI, the demand for more efficient settlement methods for cross-border trade and digital services will drive continued development in this direction. But the growth trend does not mean that a specific model has been licensed. Discussing this connection does not mean that mainland individuals or corporate accounts must directly receive and pay cryptocurrency.
We can pay attention to how compliance agencies in Hong Kong and other overseas markets can undertake payment, exchange and settlement services within the scope allowed by the local area, and then connect them with the traditional legal currency system. What is discussed here is the possible business structure. Whether it can be done depends on the entity, transaction, funding source and local regulatory requirements. It does not mean that setting up another company overseas or transferring more funds will automatically solve the compliance problem.