What can AI really bring to the economy when it starts working?
On September 10th, at the Huangpu River in Shanghai, the 2026 Inclusion · The Bund Conference officially opened. This year's theme is "Creating a New AI Economy".
Compared to previous years, this is a meaningful change.
Since the emergence of ChatGPT, discussions about artificial intelligence have long focused on parameters, computing power, Scaling Law, and model capabilities. People were concerned about whose model was stronger, who had more GPUs, and which benchmark would next surpass humans in the next generation of models.
But after three years, as large models move from chatting to handling tasks, agents begin to enter real transactions, and robots start to take part in production processes, the question has quietly changed: we are no longer only discussing what AI can do, but rather what AI can truly create — new transactions, new organizations, new productivity, and new economic growth.
This also forms one of the most noteworthy underlying themes of this year's Bund Conference.
The economists, scientists, and entrepreneurs on stage are still talking about models, agents, quantum computing, and scientific discovery, but more and more discussions have shifted to some of the oldest questions in economics: where does productivity come from? How does innovation become growth? How do new technologies create new companies and jobs? And how does technological progress ultimately translate into social prosperity?
Artificial intelligence is gradually becoming an economic variable from a technical one.
From Obedience to Action
Over the past few years, the deepest impression that large models have left on ordinary people is that they have become increasingly "good at speaking".
Ask a question, and it gives an answer; ask it to write an email, and it generates one; give it some material, and it completes a summary. The underlying logic remains that people issue commands, and AI responds to them.
But in 2026, this logic is changing.
Zhang Hongjiang, investment partner at Yuen Capital and academician of the U.S. National Academy of Engineering, summarized an evolutionary path on the main forum: Agents are evolving from "conversations to assistants, from passive to active, from individuals to groups."
The real importance of this statement is that AI is beginning to have "action ability".
An AI model that can only answer questions is essentially still an information tool; but when an agent begins to understand tasks, call tools, plan steps, and complete tasks, it first truly enters the economic process originally composed of human labor. Further, when different agents begin to collaborate, exchange information, and even complete transactions, it changes not just human-computer interaction, but the way the economic system itself is connected.
This change has already appeared in very specific scenarios.
In a panel discussion titled "When Agents Become Transaction Entities," Ant Group CEO Han Xinyi shared a small example: last week, he used the AI health application "Afu" to buy some healthy snacks like nuts and seaweed, spending over 40 yuan.
40 yuan is certainly not a huge transaction. What is worth noting is that an agent has taken its first step from "telling you what to buy" to "buying it for you."
Liu Zuohu, senior vice president and chief product officer of OPPO, also mentioned that related consumption by "Xiaobu" has nearly doubled in several months. Although the base is still small, users have started to get used to having agents help them shop.
Once AI can act, it inevitably begins to touch payment, accounts, goods, and services. The basic structure of internet commerce used to be "people find products," "people click buttons," and "people complete payments." In the future, it may instead become people setting goals, agents finding solutions, comparing prices, calling services, and completing transactions.
Therefore, an entire set of commercial infrastructure must also change accordingly.
Does an agent have an identity? Who authorizes it to spend money? Who is responsible if it buys the wrong item? How can a machine-made transaction be traced back?
Han Xinyi summarized these as authorization, identity, capability assessment, financial security, and traceable auditing. In traditional internet, people are familiar with KYC - Know Your Customer; in the agent era, a new issue is emerging: KYA, Know Your Agent.
Jorn Lambert, chief product officer at Mastercard, also emphasized the importance of commercial trust. Zhou Jingren, chief scientist at Alibaba Group, believes that as long as constraints, secure execution environments, and model boundaries are clearly defined, the trust issue of agents is not unsolvable.
As more machines are able to complete tasks, conduct transactions, and form collaborative networks, a truly meaningful agent economy may emerge.
This "ability to act" is also starting to move from the digital world into the physical world. On the same day's embodiment intelligence panel, several robot entrepreneurs formed a common judgment: embodiment intelligence will not simply replicate the development path of large models. What really determines whether robots can move beyond demos is not just model capabilities, but the full capacity composed of data, hardware, system stability, and ROI in real scenarios.
Ultimately, robots must answer not "can they move," but "can they stably enter factories, stores, and service scenarios, and truly become part of the productive force?"
From Foundation to Growth
One of the most important keywords in the AI industry over the past few years has been "foundation": larger models, more advanced chips, more data centers, and larger capital expenditures.
But looking back at past technological revolutions, it was not the rails themselves that changed the world, but the reflow of goods and people; the value of the Internet was not in how many fibers were laid, but in the emergence of new economic forms such as e-commerce, search, social networks, and mobile payments.
Is AI also undergoing the same process?
Philipp Agion, recipient of the Nobel Prize in Economics, gave a direct answer using economic growth theory.
According to his and his collaborators' calculations, considering only the tasks of AI automating the production of goods and services, the productivity growth rate could increase by approximately 0.68 percentage points per year over the next decade. If further considering AI's promotion of the "idea production" process, i.e., innovation, it could add another 0.4 percentage points.
In other words, 0.68 percentage points might even be the lower limit.
This may also be the most important layer of understanding "AI New Economy": what is truly worth paying attention to is not how much revenue the AI industry itself creates, but whether the production function of the entire economy changes because of it.
Previously, when a company increased output, it usually meant adding employees, capital, and equipment. In the future, when more knowledge work can be handed over to agents, and when a company's internal experience, processes, and knowledge can be modeled, the same scale of people and capital may produce outputs of an entirely different magnitude.
Zhang Hongjiang believes that in the agent economy, the definition of corporate assets may also be rewritten. Previously, the core assets of a company were talent, customers, brand, and patents; in the future, computing power, proprietary models, high-quality data, agent collaboration systems, and continuously iterated workflows will also become new production factors.
If the Internet reduced the cost of information dissemination and transactions, then agents are further reducing the costs of knowledge labor, coordination, and execution.
A startup can now call a group of agents to complete programming, design, marketing, customer service, and operations. A "one person + a group of agents" enterprise is no longer just an idea. Enterprises may become smaller, but their capabilities may actually become greater.
It is here that AI first truly moves from a cost item of a company to a growth item.
Of course, not all cutting-edge technologies immediately turn into productivity. At the main forum, Lu Chaoyang, executive director of the Shanghai Institute of the University of Science and Technology of China, gave a cooling down to the hot quantum computing: quantum computing is not a super GPU that can "accelerate everything," and there is currently no recognized algorithm that can accelerate large model training.
This actually provides another dimension to understand "AI New Economy": what is truly important is never how cutting-edge the technology sounds. The greatest value of a technological revolution has never been to do existing things faster, but to constantly create previously non-existent supply and demand. The steam engine created railways, electricity created modern manufacturing, the Internet created platform economies, and AI will eventually create what, of course, we don't yet have a complete answer to.
But the outline is already beginning to take shape.
From Machine to Human
At the end of every technological revolution, it always returns to a fundamental question: what happens to people?
In a conversation on the main forum that day, writer Liu Zhenyun and Professor Ma Yi, dean of the School of Computing and Data Science at the University of Hong Kong, discussed the same issue — when machines become increasingly smart, what should people truly hold onto?
Liu Zhenyun shared an interesting experience. There are now countless videos online that feature his face and imitate his voice to express opinions, “95% of which are fake.” AI can even continue to “speak on his behalf” based on works like "One Heap of Chickens" and "One Sentence Worth Ten Thousand."
But in his view, being able to imitate doesn’t mean creating.
"It can imitate 'One Heap of Chickens' to write 'One Heap of Geese,' but I haven't thought of or written a work that AI can't imitate."
Ma Yi summarized this difference as "commonality" and "uniqueness." Today's large models have already mastered literary, historical, and scientific knowledge well, but these are more of the accumulated common knowledge of humanity. However, everyone's joys, sorrows, values, and life experiences differ.
As machines become better at mastering commonalities, people need to find their own uniqueness more than ever.
This also provides another perspective on today's most common AI anxiety.
When facing students and parents asking “what should I learn so I won't be replaced by AI?” Ma Yi believes that seeking a skill that AI can never learn is perhaps an old way of thinking. Instead of avoiding AI, it's better to learn to use AI to amplify our own abilities.
Liu Zhenyun used agricultural mechanization to explain employment anxiety. In the past, agriculture relied on a lot of manpower, but later tractors and harvesters entered rural areas, allowing one person to cultivate hundreds of acres of land, but people didn't stop working. Ultimately, new productivity forms new production relations.
"AI is definitely not a flood or a monster."
What is truly worth worrying about may not be machines becoming more like humans, but humans becoming less like themselves after using machines.
Liu Zhenyun said that when he goes to Hong Kong, he can certainly ask AI which restaurants are good, but he prefers to ask Ma Yi: "When will you treat me to dinner?" AI can recommend restaurants, but it cannot truly invite you to dinner.
This joke highlights a boundary that technology can never bypass.
AI can provide knowledge, but cannot completely replace relationships; it can simulate communication, but cannot replace genuine connections between people. Ma Yi also reminds us that if all knowledge acquisition turns to AI, students may indeed learn more freely, but they may also become more isolated. In the end, people must enter society and interact with real people.
And even further away, AI may even help people create new knowledge.
Professor Wang Mengdi from Princeton University raised another question that day: how far is AI from true autonomous discovery?
Today's large models are good at learning the most common and mainstream knowledge in probability distributions, but important scientific discoveries often occur outside the long tail. She and her team are trying to let AI truly access the experimental world: proposing hypotheses, calling experimental equipment, observing results, and then re-propose hypotheses.
If the main way AI creates value today is still improving the efficiency of existing work, then the future huge increment may come from the second level of ability — creating new knowledge, new materials, new drugs, new production processes, and industries that don't exist today.
