CHKLC 23rd Anniversary cum HK CG & ESG Excellence Awards2025-Keynote Address-Dr. Kai-Fu Lee

Thank you very much. Thanks especially to Professor Chan for inviting me to speak to such an esteemed group of people. What I want to talk to you about is the most important technology revolution not of the decade, not of the century, not of our lifetimes, but of the entire human history of the human race. This is something really to look after because if you go back and see how much each technological revolution increased the GDP, we see massive increase going to industrial, to internet, and now we go to the AI agent revolution, and that is going to propel everything to an unbelievable height. In my talk, I will explain how.

I'm not going to ask for a show of hands, but I believe if I asked you how many of the companies here use AI agents daily, I don't expect to see any hands. Oh, one hand. Okay, maybe one hand. That's good. But that is not enough because imagine if Thomas Edison were in a similar talk and asked how many people are using electricity, or Marc Andreessen how many people are using the internet, and one hand goes up out of several hundred, that is not enough. Next slide please.

Before I go into AI agents, I just want to tell you, you're probably thinking what about DeepSeek, what about OpenAI? Well, this is all built on OpenAI and DeepSeek. OpenAI and DeepSeek are types of LLM, large language models, that serve as the brain of the agent. But most of us think of ChatGPT or DeepSeek or DBA as the standard application. But that is all wrong. You've all been misled. The chatbot is merely the first example of an interesting app. It is not the ultimate app. It's like if any of you used the Macintosh once upon a time, chatbot is like MacPaint. Does anyone use MacPaint anymore? It's long been eradicated by Photoshop, much more powerful. So it's a sample app, an early test. It is not the answer. It is not even close to the beginning of the answer.

We've now gone through a very fast revolution of applications built on large language models. We've seen AI applications that are tools that can generate videos and images. Microsoft launched a copilot which watches what you do. But today we're about to witness the age of AI agents. An AI agent is when you tell the AI what you want to accomplish and it figures out the steps and gets it executed. That is completely different from a chatbot. If I told a chatbot I want to go to Hong Kong and go to these three restaurants, visit with these three people, and speak at this conference, it can probably give me a hypothetical itinerary. But what good is that? What I want is what my assistant does: she knows what airline I like, what hotel I like, who I meet with, their priorities, and how to arrange things so the transportation is convenient. She would book the flight, book the hotel, arrange all the meetings, and all I have to do is get on the plane knowing everything has been done. That is what an agent or an assistant does. They do, they don't just write, they don't just answer. So that's AI agent.

To show you, on the very top we see that an agent takes a task, breaks it down to steps, and accesses tools and knowledge base and delivers the final result. That means on the left side they have a brain which might be OpenAI or DeepSeek, and on the right side they have an infinite memory of you, your corporate experience, your preferences, just like your executive assistant does. Then on the bottom, they actually execute tasks. When I talked about booking my travel to Hong Kong, what did that include? Finding a flight, so they use Ctrip, Booking.com. What else? Alipay, WeChat Pay, or PayPal to pay for my trip. It knows how to access Professor Chan's WeChat, so it bombards Professor Chan until he answers when we can meet, then it fills in the schedule, location, car, and it has access to my payment account. Everything is done. They do exactly what my assistant does. That's what makes agents really special because they offload tasks from us and they execute them better and faster. Next slide please.

There are three types of agents. One is kind of rote learning: it memorizes a trip, a hotel, an airline, meetings, restaurants, people, dinners around this time or that time, how you pay. But now because LLMs have gotten smarter, they can not only do tasks that are fixed, which are workflow agents, but they can do tasks that have never been done before. So if I want to do something crazy, like I tell my agent I want to do a work trip to Hong Kong with these two meetings, but I want to try this new restaurant, and I need to be in Jiuai and eat with my daughter, then I want to go to Disney with my granddaughter, then I might visit Macau if there's time, and do I need a visa for the Philippines? This is a totally crazy request that a workflow agent wouldn't be able to handle because I'm mixing business and pleasure and family. But a smart secretary can easily handle that. So a reasoning agent will break down each task, and even on tasks it hasn't done, it will figure out how to do it. If I told the agent, 'When can I go travel to Mars?' it can actually answer that. It may not be able to book a trip, but it can tell you the projected time when Blue Origin and SpaceX might have a Mars expedition. That is the beauty of a reasoning agent. It thinks, it reasons, it plans, and then it executes. After that is something even more amazing. Reasoning agent is kind of here today, and I'll show you a demo. But multi-agent is the even more exciting thing that will transform all of your organizations.

Why is that? Because multi-agent is when agents talk to each other. Think about the original PC. Some of you may be old enough to remember the first PC. It wasn't connected to the internet, was it? That was the human-machine interface, a human and software talking to each other, just like chatbots today more or less. But in the future, as chatbots become agents, why shouldn't the agents talk to each other? Why should Professor Chan have to tolerate Dr. Lee's agent bombarding his WeChat, knowing this is not Kai-fu talking? So my agent should just bombard Dr. Chan's agent. Then they sort things out. Then guess what happens? My trip to Hong Kong can be planned in 10 minutes or maybe one minute because all the agents know everything. They don't have to go back and forth: 'Are you free Friday at two? No. Can we do three? No. How about four?' They know the entire schedule. Agents talking to each other is magic, just like PCs talking to each other created what today is known as the mobile internet. If you unplug the internet, remove the Wi-Fi, what can you do with your PC? Almost nothing. Connecting agents to each other is the next big revolution. It's called multi-agent. It makes them networked. Next slide.

What does it mean when it's all networked? Imagine your company will have HR that can take a job description and screen resumes, go to LinkedIn, send emails to recruiting agencies, headhunters. Bunch of resumes come in. It sorts them top to bottom, and AI will conduct the first interview. By the way, if your company doesn't use AI for the first interview, you're really missing out. It not only saves your HR manager time, it allows you to have a pipeline of 100 times more candidates because you don't need human time interviewing. It's not for the hiring decision, it's for screening. It allows you to take a much bigger pool and therefore hire much better people. Coming back to this, then arrange for the human interview, negotiate offers—those can all be done by AI agents. You might want the human in the loop to provide tender loving care for the candidate, but AI can pretty much prepare the script and your HR manager just needs to read it with emotion.

So what happens then to the org chart? You've got all the HR agents working there. When they work together, magic happens again. Think about the performance evaluation agent. Surely you have a performance review every year, and you have some superstars who earn big raises and some lower performers who get let go. You do it as a singular exercise in your company. That's wrong. Why? Because when you do this by agent, your performance review agent will tell your recruiting agent, 'I just gave these guys a huge raise and I fired these people. So in the future, don't hire me any of these people I would currently fire. Find what it is about them that's intrinsic to them not working in the company—whether it's capability, training, or value fit—and these are the superstars, get more of them.' When you do this for five years, your company will accumulate a virtuous cycle which will get you more and more people who fit your organization and your skill set requirements.

At the top level, this is your executive management team. They're all agents—or actually initially they'll be humans managing agents. Eventually they'll be agents that the CEO can summon. This is great news for you CEOs. You always wanted to be in charge, always want to give an order to a team that listens, executes without question, and gets it done before the next day. That's going to happen. But imagine when the CMO in your company has an idea: 'I have this great idea for a product, let's go build it.' Then the CTO would say, 'This is really stupid, it doesn't possibly work, don't you know technology?' I'm sure your team has had discussions like that. That will never happen again because all of your CXOs will be built on the entire knowledge base of every imaginable discipline, every imaginable part of your company history, and it will just be different in the way they optimize their KPI. Your CMO will be a very capable technology person and will not propose a product that's not technically feasible. Think about how efficient that is. Next slide please.

About this revolution going forward, we can really think about an organization that's super efficient. These agents are super fast, hardworking, work 24/7, don't get tired, don't complain, don't have labor unions, and they're infinitely scalable. Imagine each of you must have a superstar on your team. Don't you wish sometimes you could have 10 of that superstar? Not possible, human cloning is not legal. But if you have a superstar AI agent, you can make a thousand of them. Every one of your agents will be equally stellar. When you replicate them, they'll multiply the value to your company. You can do international expansion because they know all the languages in the world. Your human resource cost goes up, the AI agent cost goes down three to five times per year going forward. So it's going to be cost zero in about five years. That's what's going to happen. Your organization will become very efficient because the more agents you have, the less politics, the less information asymmetry, the less boundaries, the less bureaucracy. The agents are just prioritized to get you, the CEO, your job done.

To be fair, it's going to be a difficult process to get agents in the company because the humans will be in resistance. I'll get to that in a moment. But it's important to realize that we need to position the message to the employees that initially it's you, the individual contributor, managing agents for efficiency. Later it's you working with the agents. Later there will be more agents and fewer people. But if you do a good job, all the people keep their jobs, but we get 10 times more agents to give you 10x the power. That's the message we have to deliver. But in reality, some companies will be like that. AI agent will be a force for growth. Other companies, AI will be a force for human replacement. That is simply inevitable. But the growth is what we want to hear. So I'm going to focus more on the growth, recognizing there will also be replacement.

How does the growth look like? This is not a Kai-fu Lee dream. This is written by Sequoia in the US. This is a trillion dollar growth, a trillion dollar race. The way it's going to happen is AI should not be paid by API, by workplace replacement. Yes, you can replace people, but that's not important. What you want is a multiplier to your revenue, multiplier to your profit. That's what's going to make AI super powerful. If you imagine, we're working with an insurance company. They have a superstar insurance sales team. What we want to do is take that superstar insurance sales team, offload their work, take part of their discussion with their customers. This works with every salesforce that you have. AI can talk to customers in chat, but when it's human to human, it's handed back to the person. Then AI as a chat vehicle—imagine if your super salesperson gets the AI to do their job on WeChat, WhatsApp, and maybe even on the phone, but the salesperson does the job person to person. Then surely the amount of time saved is on the order of two-thirds. That means each salesperson can generate three times the revenue. That is the power of AI. It's not replacing two-thirds of your salesforce; it's making the best of your salespeople become three times more productive. That's what this paper is about. Next slide please.

How to build the agent? What model is good? You guys are in Hong Kong. Some of you are in China. The answer is obvious: using the Chinese open source models. Which one to pick? Let me pick for you. If you use our platform, we'll always give you the best open source model, which changes all the time. But if you're interested in open source versus closed source, I don't have enough time to go into details. You can look for a Financial Times column that I wrote yesterday which explicates all the details and differences. Next slide please.

What we do at 01.AI, Lingiwan Wu, is we build the tool to let you build the agents. This tool is called World Wise. Next slide please. Let me just show you how this works. This is a demonstration of World Wise. This product is being used by AIIB, which is a bank, the Asian Infrastructure Investment Bank in China. We built them an agent, and you can pretty much ask the agent to do the work of a research analyst. So the banks in the room, the financial funds in the room, this is something your research analyst may do. Let's click and see the demo. The request I ask is: develop a well, first develop industry analysis. We develop it very simply. Then we ask the question, we run it. This is running in real, almost real time. Make a proposal for AIIB to invest in below countries in AI. Then it splits it up into seven tasks all by itself. There's no human involvement. Then it's now executing. It's going off to web search, to maps, to understand where is Kazakhstan, what's the GDP, what's the current AI usage rate. Then it develops an answer. If you had a human do it, it would be a 100-page report. But AI is done in 10 minutes, and it's not a report but a full website including dynamic content as well as code. This is the outcome. It's done. Now you click on this. It shows you the countries, the investment areas, and it develops a map. I didn't ask for a map. It developed a map. Here's Jordan, Saudi Arabia, Turkmenistan, and how they perform and rate in terms of AI. These are the things you can go into: the top countries for investment, the top areas, infrastructure, data center, models, applications over the five years, how to invest, what return to expect, what's the benefit for the economy, what's the multiplier. It's all done by AI, zero human involvement. If you don't believe it, we can arrange a time. I'll show this to you in real time, and I can ask any question about investment. Now, it's not perfect. Your best research analyst can do better than this. But can this save 80% of the job? Absolutely. Next slide.

Now I come back to the fundamental problem: most of you are in what I call traditional industries. Traditional industries know a lot about their industry, but they don't understand how to use AI, which model to pick, how to build agents, whether their IT team or CIO is suitable. AI companies don't want to work with you because to them it's much easier: take my API, take my chatbot, it's good enough. I'll sell you an agent, use my cloud. That's the message you will hear. But then you're on your own. Or you can call McKinsey, and they'll build you a beautiful PowerPoint, but they can't implement. So this is where we're stuck: the technology is here, and we are not ready to find a way to work together. Next slide.

What I'm doing with 01.AI is we decided to be the good Samaritan. We're going to do what no one else does: we want to bridge the gap between advanced AI technologies and traditional technology companies. We will go to your company, understand your strategy, know your pain points, see how you make money, what are your KPIs, and then we will help you craft an AI strategy the same way McKinsey does, except ours can be implemented. Hope McKinsey is not in the room. How do we know ours can be implemented? Because we will do it for you. You're going to say, 'I don't want that. I don't want to pay you forever.' I'll say, 'We'll do it with you.' When we're done in a year or two years, we'll hand everything to you. Your team will be trained. If any of you remember the beginning of the PC revolution, you didn't know how to do it. You hired someone like Accenture or whatever they were called, Anderson Consulting, to come and help you build it. Then later you got your IT department. We're all over in the same place again. So we will be that Anderson Consulting and McKinsey to you, from strategy to implementation.

One other thing that's important: it must be driven top down because your people are going to fight tooth and nail. They don't want this to happen. It will affect their job, their political status, and their sustainability and value to your company. But that is life. It's time to move on. Some groups will become great, some people will become greater, and others will be eliminated. So it is not time to feel for your people and slow down the change because it is a ruthless future, and only the paranoid, only the visionary will survive. We're here to partner with you. Next slide please.

What kind of implementation have we done? We'll work with you on anything, but we ideally want to work with you on core business because what's the point of saving you 20 people in customer support? You can do that with any company's readymade product. We want to customize and do for you that which is most important to you: your corporate responsibilities, corporate values, and what delivers your quarterly results. Let me give you four quick examples of what we have done or are doing. This is a real customer, a $30 billion mining company. We optimized their entire supply chain one step at a time with the simple goal of maximizing iron ore throughput delivery to the customer base. I don't have time to go into a lot of details, but you can imagine we tweak their autonomous vehicles, their shipments, their train schedule, their ship schedule. In the end, we hope to save them a hundred million dollars a year going forward. This is a two-year project, and we hope to achieve that, but we're already showing excellent early results. The models we tweak for them are doing much better than the human work that's being done now. Next slide.

Another example is a financial institution. We're working with one of the mega funds in the world that is currently interested in doing three things. One is recommending to customers, which basically means robo-advisor but in a way that's responsible and customer focused. Secondly, make smart decisions that include credit risk, transaction risk, anti-fraud. The third is to make strong compliance and clearance and regulatory compliance, all within tracking all the portfolio companies for erratic or new things that you need to know about. I can't really say a lot more about the details, but the obvious answer is it's about sales, products, compliance, and risk. Next slide please.

The third example: I think there are a lot of real estate companies here. You probably think this has nothing to do with real estate. No, it has a huge amount to do with real estate. One is the design. If you're building a large mall, you'll want to know what merchants to bring in that will maximize the revenue. You need to understand your shoppers much better. If you're a property management firm, you're the greatest portal, the most trusted portal. You're much more trusted than Douyin and Xiaohongshu. You know this because some of you sell very expensive eggs. I know that because I live in a property in Beijing. If you can sell very expensive eggs to me, you can sell a lot of other things. You're the trusted connection for the residents of your home, and you're not maximizing that property. Do you know the profile of your residents within each of your buildings? Do you know what elevator ads to place? Do you know what message to give? Do you know how your security guards, when they become friends with the residents, you can upsell them? Do you know your demographics? It's all a big black box, and it needs to be customized, tailored, exactly targeted to make you the most money. And of course, with all the home recommendation, sales and optimization, and contract planning. Next slide please.

Lastly, I know there are many of you who plan press conferences, events like this event. I'm sure a dozen people worked hard for six months for this event. I'm sorry to tell you next year maybe two people can do it for two months, and elevate the rest to do more interesting work. For example, you can imagine Professor Chan can say, 'I need to design the brochure, the website, the PowerPoint, the logo, the seating charts, send out invitations to all these CEOs and see who's coming to decide who's sitting where.' This can all be changed dynamically. You can have people supervising it to make sure the protocols are followed, but it can probably learn the protocols from the past, and then the result comes out. Professor Chan says, 'Oh no, we can't sit Secretary Qiu here, that's not respectful.' Then you can tweak everything around. Or you can say, 'I want the logo to be orange.' Then it's done, everything and everywhere. This is multi-agent at work. We're working with a company in China that is using this to accelerate event planning, but it can be done much more than that. Next slide please.

So basically, this is what we do. We build a platform, an agent platform. You don't have to use ours, but you've got to get those agents developed. It's got to be done driven top down. Your people are going to resist, but if you don't push, then you're not going to run ahead and you might fall behind. Next slide please. This is basically how we see the ecosystem, and we're landing in China at least Hong Kong. Next slide please.

Lastly, I just want to have an invitation for any of you to contact me by WeChat or by email. I think Professor Chan wanted me to say a few words about Hong Kong. I'm full of hope for Hong Kong because you have some of the best industries suitable for AI transformation. Financial and real estate are certainly ready for that. Much of your data is already digitized. The three keys to successful AI delivery are a committed CEO, completed transformation, and a welcoming IT team. If you feel your team has that, then you can be that first example of successful implementation. You can be that lighthouse account for Hong Kong. I think Hong Kong also has the advantage that universities have a lot of great people, and there are a lot of people coming from the mainland who can be engineers, participants, and entrepreneurs. So all the things are in the right place. But I have to say the Hong Kong companies are a little bit behind American and mainland Chinese companies, and I hope this talk will change that and enable you and encourage you to embrace AI as soon as possible. Thank you.