91. The Bund Summit Transcript: The Dawn of the Agent Economy
Summary
- Ant Digital Technologies is betting that AI and blockchain will together provide the trust and transaction infrastructure for the agent economy. 边卓群 says the small, 7×24-hour transactions between Agents require a traceable, auditable ledger; 余滨 stresses that large models operate in “a world of probabilities” and need more trustworthy underlying data and infrastructure. The host sums up the company’s long-running strategy: “This isn’t a change-of-track story; it’s an acceleration story” (“不是一个换轨的故事,而是一个加速的故事”).
- Industrial AI’s key inflection point is the shift from asking whether the technology has value to asking how that value gets into the business. 余滨 says DeepSeek’s arrival during the 2025 Spring Festival accelerated that shift. In 2024 and 2025, Ant Digital chose to move directly into complex sectors including banking, insurance, energy and healthcare, taking on more upfront investment. “Nothing ventured, nothing gained” captures the logic: use real business needs to shape models and engineering methods.
- The “Agent superfactory” is a complete system for building and operating digital workers. Its delivery package includes industry and domain-specific models, hundreds of skills, Agent templates, hundreds of digital experts, plus assessment, security and operating mechanisms. Two production methods serve either SOP-driven workflows or task-oriented execution; clients can build Agents or put digital workers straight to work.
- In response to the claim that Agent-to-Business is hard without IM, 余滨 shifts the focus to business and production data. AI could change the organization itself, he argues, so an organization’s data advantage does not automatically translate into production value. 边卓群 adds that the company’s move into finance, energy, healthcare and retail reflects its existing industry experience; “using data to understand a business” is her basis for competing.
- 边卓群 defines the long-term edge in vertical Agents as a cycle of reinvestment in industry knowledge, data and delivery capabilities. In response to the challenge that large models will eventually absorb everything, she acknowledges that model companies can also “grow hands and feet from the brain of a large model” to serve an industry; she does not rule out their entry. Her point is that the moat is dynamic, built as customer work continues to generate capabilities and data.
- The two guests use different yardsticks for when the agent economy reaches its inflection point. 余滨 looks to changing access points: Agents such as Alipay’s “A Bao,” which connect with merchants, are already making the economy visible. 边卓群 looks for people to be freed up: “When we no longer feel anxious talking about AI,” and have time to discuss creativity and a better life, the economy will have truly arrived.
- Both guests expect digital workers to take on real business responsibilities in the next phase. 余滨 believes more digital workers will handle production and business operations “next year,” and says AI could begin to play a key role in operating reports. 边卓群’s timeline for collaboration across software and hardware, embodied and non-embodied systems, and silicon- and carbon-based workers is “possibly in the next 3 or 6 months.”
Deep dive
1. The dual-engine strategy began with internal testing and industry experience
卫诗婕 opens by noting that Ant Digital Technologies is now in its 11th year, has ranked first in China’s blockchain BaaS market by share for 6 consecutive years, and has recently been the subject of pre-IPO financing reports. The Bund Summit roundtable focuses on how the company is combining its blockchain capabilities with AI for business, and how the two lines of business will serve the agent economy.
边卓群 recalls that in 2021 Ant brought together intelligent technology, represented by blockchain; cloud technology, represented by financial cloud and mobile technology; and security technology, represented by privacy computing and security. Internally, they called it “the three armies joining forces.” Ant Group was the “training ground”: technology was tested internally before being packaged as products for external customers. That internal-external loop made the team especially focused on application value, stability and security. 余滨 traces Ant Digital Technologies’ roots to 21 years of technical development and market experience at Ant Group. Working with enterprise and industry clients brought commercial understanding and data that could be used to solve industry problems.
边卓群 says Ant Group has been discussing blockchain for about 10 years. Around 5 years ago, the team began exploring ways to combine AIoT and blockchain, using devices to record data and turn static assets into liquid, intelligent assets. In 2023 and 2024, the discussion expanded to putting real-world assets on-chain and tokenizing them; now it extends to Agent interaction and payments. 卫诗婕 sums it up: “This isn’t a change-of-track story; it’s an acceleration story.”
2. DeepSeek pushed AI toward business value; the hardest use cases became the starting point
余滨 says the team was already working on industrial AI in 2022. After studying ChatGPT in 2023, he saw “ask-anything, answer-anything” products as mainly knowledge tools and interaction layers. Around 2024, rebuilding services around an internal financial AI product made him more certain that AI should be more than a large language model: it should become a “business model” that actually solves business problems.
For him, DeepSeek’s arrival during the 2025 Spring Festival was the second decisive moment. Companies stopped asking first whether AI could be deployed and moved straight to how to create value in their own businesses. The market went quickly from questions about value to implementation. The continued expansion of international products into healthcare and finance has further strengthened his conviction in domain-specific Agents and industrial AI.
The strategy prompted a major internal debate: start with customer service, tools and documents—smaller, more stable investments—or move from the outset into hard use cases in banking, insurance, energy and healthcare. In 2024 and 2025, the team chose the latter, on the logic that “nothing ventured, nothing gained”: real business needs would show them which models and engineering methods were needed, and which problems actually mattered.
边卓群 says the two sides “battle all the time,” with disagreements centered on tactical choices. Building an Agent that understands a company’s business takes substantial time, money and staffing; a general-purpose Agent could reach a standardized commercial model sooner. Which sector to invest in first and whose problems to solve must also change with demand, but the dual-engine strategy of AI plus blockchain has never wavered.
3. Trust is a prerequisite for closing the Agent transaction loop
边卓群 frames the questions this way: when Agents join people as participants in the world, “Where does the interaction run? Who records it? How does payment happen?” She argues that blockchain can support small-ticket, 7×24-hour, multi-point transactions that are traceable and auditable, enabling seamless transactions between people, people and machines, and machines.
余滨 points to another need, rooted in the uncertainty of large models: AI is “a world of probabilities,” so the data and underlying infrastructure it relies on must be more trustworthy. Combining blockchain and security capabilities is how he says the team works to strengthen AI infrastructure. The first priority is an environment that participants can trust.
卫诗婕 asks whether the “last mile” of trust is still open when an Agent can handle tasks and spend money on someone’s behalf. 余滨 compares this with the trust problem in the internet era and offers a stark assessment: “If there’s no way to solve the trust problem, the agent economy may simply cease to exist.” That is why blockchain occupies a key place in the company’s Agent product architecture.
4. The Agent factory brings models, workflows and digital workers into one system
余滨 stresses that enterprise AI is very different from an individual looking for an assistant: models must be combined with internal knowledge and data to become the company’s brain, and traditional software engineering methods must also change. The “Agent factory” captures lessons from internal and client deployments. Its foundation includes industry and domain-specific model platforms, with components, knowledge, development and production methods above them; it is not just a development tool.
The factory comes with hundreds of skills, including MCP, various collectors and knowledge sets, as well as two production methods. One is like an assembly line: AI follows SOPs to complete a defined process, suited to rigorous settings such as finance and manufacturing. The other is task-oriented “intelligent manufacturing,” in which AI is given a task and carries it out. The two methods address different process constraints and execution requirements.
Above that are Agent templates, like production blueprints, and hundreds of digital experts and digital workers across industries, who can also be assembled into expert teams. Clients can use the factory to build their own Agents or select ready-made digital workers to take on roles in business operations. The offering therefore includes both production capabilities and workers ready to use.
Once digital workers are on the job, they also need management, assessment and ongoing operations, so the system includes security, operations and a workbench. 余滨 describes the workbench as the interface through which domain-specific Agents and digital workers interact with the outside world, and mentions versions for finance, healthcare and retail. “Materials, production processes, quality control and risk control” is why he says the factory metaphor fits.
5. Agent-to-Business competition comes down to production data and industry knowledge
卫诗婕 raises the competitive question through Tencent Work Buddy, Alibaba’s Qwen office offering, and the integration of Feishu and Doubao. She also cites the view that “companies without IM shouldn’t rush into Agent-to-Business.” 余滨 responds that IM reveals organizational data, but AI could make organizations flatter. “A company’s most important data should be its business and production data”; value has to be delivered in the production process.
边卓群 argues that the market is large and that demand for personalization is even stronger in the AI era: “No single product can meet everyone’s needs.” She grounds Ant Digital’s market entry in years of data and industry experience. Model capability matters, but so does understanding clients and businesses well enough to turn technology into better user experience and business value.
Finance came first because of Ant’s accumulated experience serving financial institutions; energy was one of the earliest sectors to adopt blockchain. 边卓群 points to charging piles, energy-storage cabinets and solar panels: add AI capabilities to the equipment, then put the data on-chain through blockchain. The principle behind choosing these markets, she says, is “do what we’re good at, do what we understand.”
The healthcare case rests on Ant Group’s investment in “Afu” and related businesses, and its understanding of the industry through hospitals, users and doctors. Retail draws on its experience serving large user bases. 边卓群 cites user populations of 500M and 1B to make the point: retail needs to move from finding customers to finding them more precisely—and offering what they actually want.
6. The challenge that large models will absorb vertical AI gets an industry-focused answer
卫诗婕 challenges the vertical-AI strategy: general-purpose intelligence is absorbing capabilities that were once proprietary, so could large-model companies eventually cover all vertical applications? 边卓群 does not rule out that path: “A large-model company could also grow hands and feet from the brain of a large model to serve an industry.” Her answer is to do what each company believes in and build on its own strengths.
She qualifies the claim to technical leadership: “We may not have the most cutting-edge technology, but we know how to put it to work.” She then distinguishes between a barrier and a moat: a barrier is static, while a moat keeps changing. Serving an industry generates value, as well as data, capabilities and understanding; companies can build on those gains and continue to grow. That, she says, is what makes competitive advantage dynamic.
余滨 points to the distinct combination of capabilities at Ant Digital Technologies: a full-stack product suite spanning AI infrastructure through Agents, alongside more than a decade of investment in blockchain-related technology. The former supports Agent development today; the latter is aimed at the trust infrastructure of the future. He says the company can also help build the infrastructure that a trusted society and an agent society will need once the agent economy emerges.
7. The agent economy will mean new entry points and more time freed up for people
余滨 says the agent economy is already taking shape. In the past, thinking about entry points often meant thinking about mini programs; now the question is whether Agents can integrate with entry points on other AI platforms and get used. He cites Alipay’s “A Bao” as a new-generation entry point for 10M merchants, arguing that new entry points are emerging and economic activity will follow.
边卓群 offers another yardstick: “The era will arrive when we no longer feel anxious talking about AI.” Standardized, high-frequency work will go to AI workers, and companies can become “super Agents.” Once people get their time back, they can turn to creative work—and have room to discuss what is beautiful and what they want the world to become.
Asked to imagine switching roles, 边卓群 says she would give her Agent to 余滨, and 余滨 could rent his Agent to her: “Just leave your knowledge, your emotion and your skills.” Even if someone is not physically present, she says, calling on that person’s Agent could provide support. This is her vision of exchanging energy, value and know-how, rather than requiring each person to do all the work themselves.
8. The next test is whether digital workers take on business responsibilities and show up in results
At the start, 卫诗婕 cites the first time machine-generated internet traffic surpassed human traffic, in June this year, and Ant Group CEO 韩新义’s view that the agent economy could take off in 6 to 12 months. That sets an urgent backdrop for the discussion. The guests’ closing forecasts focus on changes to roles and collaboration inside companies.
余滨 expects more digital experts and digital workers to enter companies “next year” and take on substantive business and production work. He is more cautious about the change to reporting: “Starting next year,” AI could become a key part of company operating reports, making it easier for companies to say exactly what gains AI has delivered.
边卓群 expects silicon- and carbon-based workers to collaborate more closely, with embodied and non-embodied systems, software and hardware moving toward integration—“possibly in the next 3 or 6 months.” Her ask of people is to bring their most intelligent, emotional and aesthetic qualities fully to that collaboration.