Vol.58 Discussion of Several Questions on 2B-AI Agent—Transcript of the 2025 Generative AI Business Summit
Summary
Enterprise demand for AI Agents exploded in 2025, driven by simultaneous improvements in model capabilities, invocation protocols, and market budgets. Domestic models such as Qwen 1.5/2.5 have brought most enterprise tasks into the usable range, while DeepSeek R1 further improved performance on complex tasks; MCP and A2A have lowered the cost of tool integration. Future Time Intelligence’s Q1 demand has already exceeded the whole of last year, while 众数信科 has been running at full capacity since the Spring Festival, with 吴炳坤 describing the market as “selling masks during the pandemic.”
Consumer-facing general-purpose Agents may ultimately converge on entry points such as phones and glasses, leaving startups more defensible in To B’s proprietary data and industry know-how. 众数信科 chose higher education, shipping and maritime, and Zen Buddhism, based on a common profile: “high knowledge density, weak AI capabilities, and relatively close to money.” Future Time Intelligence plans to serve State Grid, the “three oil majors,” and other leading customers first, then move down to mid-market enterprises and professional users as the technology matures.
The core upgrade from SaaS to Agents is a shift from delivering tools to replicating experts and delivering outcomes directly. 吴炳坤 describes the previous SaaS cycle as “equality of information,” while this cycle is about “equality of skills.” 杨劲松 believes vendors can charge for service outcomes and share in the value created, or even enter traditional industries directly with 3-to-5x efficiency and participate in their profit pools.
Enterprise AI adoption is moving from departmental procurement to a top-executive project, expanding system coverage while lengthening sales and delivery cycles. 高佳辉 observes that projects are no longer about a CIO or one business unit buying a tool, but about connecting every system across the group. Vendors therefore need to run large bespoke projects alongside standardized digital employees paid a monthly “salary.”
Vertical Agents will not follow a one-size-fits-all product model; the real moat is encoding industry differences into workflows. Luxury hotels still need people to serve guests, making AI better suited to digital store managers and process optimization; tens of thousands of economy and business hotels pursue standardization and need the group to centralize guest interactions. “Don’t deify it”—adding the letters AI does not automatically reduce costs or raise efficiency.
Personal super-Agents and enterprise Agents may eventually interact directly. 高佳辉 envisions personal Agents communicating and even “bargaining” directly with hotel-group Agents instead of continuing to simulate human browsing on OTAs. That would reshape traffic entry points and platform value, but for now it remains a directional judgment; enterprise deployment may come first.
The biggest constraint on scaling is not demo quality but the accuracy, stability, and end-to-end traceability required by To B. Enterprise customers want projects launched in January, deployed in February, and put straight into production, prompting 高佳辉 to “pour a bucket of cold water” on the expectation. 吴炳坤 wants models to develop enterprise memory, while 杨劲松 expects vertical reinforcement learning to codify expert capabilities.
Deep dive
1. Agents Evolved from Language Models into Action Systems in 2025
庄明浩 used OpenAI’s L1–L5 framework to set the terms: L1 is the Chatbot; L2 covers reasoning models such as DeepSeek, OpenAI o1, and o3; only L3 is the Agent. The first two levels process language, while L3 begins turning capability into “behavior.”
His core judgment is that much of the industry’s discussion from 2022 to 2024 remained focused on describing Agent characteristics, whereas 2025 is the year people began to “see” real products and business deployments. On the day of the forum alone, Shanghai hosted three AI conferences of comparable scale, making the market’s heat a shared industry consensus.
That shift also explains the change in enterprise demand: customers are now asking directly, “What can AI do?” and demanding that it enter production workflows, call tools, and complete tasks.
2. Three Companies Are Embedding Agents in High-Knowledge-Density Industries
吴炳坤 said 众数信科 was founded four years ago. It spent its first 2 years on digital-city data operations and its next 2 years participating in research on domestic general-purpose large models, eventually positioning itself around private enterprise models and the business Agents built on top of them.
众数信科 selected three seemingly unrelated sectors—higher education, shipping and maritime, and Zen Buddhism. Their common traits are high densities of proprietary knowledge, weak existing AI capabilities, and being “relatively close to money.”
Future Time Intelligence, where 杨劲松 works, was founded in June 2023 by a team from the DAMO Academy large-model group. It calls itself an Agent Native company, offering an enterprise-grade general-purpose AI Agent platform designed to put Agents into production workflows through products.
高佳辉 followed a path from technology to smart hotels, digitization, and finally AI transformation, providing hotel groups with customer-facing digital concierges, internal digital store managers, and board-level digital secretaries: “You just pay me a monthly salary.”
Hotel customers had been especially concerned about the privacy risks of uploading data to general-purpose models. Advances in private deployment, local training, RAG, and knowledge retrieval have made “large models plus vertical industries” much more practical.
3. Model Maturity, Protocol Standardization, and DeepSeek Jointly Ignited Demand
吴炳坤 recalled that when the company launched an early Agent in September 2023, the underlying models had weak instruction-following capabilities and “failed very easily.” Model capabilities improved this year, while DeepSeek sharply lowered the cost of trial and error, rapidly heating up market discussion and experimentation.
杨劲松 believes Qwen 1.5/2.5 lifted domestic model capabilities to a new level, while DeepSeek R1 further improved the quality of complex tasks. Planning, memory, and tool calling—the capabilities Agents require—have also become more viable as protocols such as MCP and A2A have taken shape.
Commercial demand followed. Future Time Intelligence’s Q1 demand exceeded the whole of last year, while 众数信科 was running at full capacity across the team after the Spring Festival. At most half of its 10 customers could be retained, and the company said it had served about 60% of Xiamen’s state-owned enterprises and roughly one-fifth of Fujian’s in a single month, with the market feeling like “selling masks during the pandemic.”
4. Vertical Focus Is Fundamentally a Search for Moats Foundation Models Cannot Absorb
吴炳坤 initially weighed To C against To B. His conclusion was that To C products would likely be “blown up” by rapid model iteration, with many hit apps already disappearing as a result. As compute costs continue to fall and algorithms are increasingly open-sourced, the key To B assets are customer data and industry know-how.
众数信科 screens for sectors with knowledge barriers, insufficient competition, weak AI capabilities, and willingness to pay. Despite the team’s digital-city background, government revenue accounted for less than 10% of last year’s revenue because To G “is not a particularly good commercial ecosystem.”
杨劲松’s end-state view is that general-purpose To C Agents will eventually concentrate in entry points such as phones and glasses, becoming a battleground for large technology and model companies. Startups will struggle to build durable moats, so the plan is to start with leading enterprises and move down-market as the technology matures.
高佳辉 sees vertical companies and large technology firms as partners more than competitors for now. Large firms provide “atomic capabilities”—models, maps, music, DingTalk, Tmall Genie, and more—while startups combine them and embed them into hotel workflows.
5. When Large Enterprises Buy AI, Protection Against Disruption Comes Before Cost Savings
杨劲松 divides large-enterprise decision-making into 3 levels: first, determining whether new technology could disrupt the existing business; second, looking for new product formats; and only third, pursuing cost reduction and efficiency gains. Strategic investment declines from the first level to the third.
He used Meituan as an example. If an “AI friend” that understands the user were to handle restaurant reviews and food-delivery choices, then call Meituan through an API, Meituan’s value as an entry point could be weakened. That is the first change large enterprises need to consider.
This is why large enterprises are often the most aggressive when new technology is still immature. Future Time Intelligence’s early customers were mainly state-owned enterprises, State Grid, the “three oil majors,” and other large organizations. The company sees this as a phase of technology diffusion, not a permanent boundary around its customer base.
6. The Hotel Industry Shows That Even One Vertical Cannot Use a Single Agent Template
高佳辉 advocates conducting a “health check” before starting a project: assess the company’s infrastructure, data, and digital foundation before deciding what AI can do. “Add the 2 letters AI and your company immediately reduces costs and raises efficiency” is not realistic.
Luxury hotels emphasize human service and do not need AI to replace customer-facing interactions. A digital store manager is better suited to organizing service processes, analyzing operating data, and turning experience-based management into standardized workflows.
Economy and business hotels, which may operate tens of thousands of outlets, are pursuing standardization. They need guest interactions centralized at the group level rather than handled independently by each property, giving a group-level Agent greater value.
7. Agents Rewrite SaaS’s Unit of Value and Open the Door to Outcome-Based Pricing
吴炳坤’s distinction is that internet SaaS delivered “equality of information,” while large models and Agents deliver equality of skills and productivity. The former delivers the same software and is easily understood by customers as “copy one for me”; the latter enters production and reshapes workflows.
众数信科 combines basic capabilities such as Q&A, creation, and review with customer knowledge bases to create different Agents. Beyond private models and an AI middle platform, customers are willing to keep paying by the number of Agents they use.
杨劲松 emphasizes that SaaS still requires an expert to use tools to complete a task, while an Agent can “replicate the capabilities of an industry expert” and deliver the service outcome directly. Pricing can therefore shift from one-time tool sales to outcome-based fees and a share of the value created.
When proving ROI is too difficult, an Agent company can enter traditional industries itself. 杨劲松 compared this with the shift from the Yellow Pages to online retail: instead of persuading every company to adopt the technology, use 3-to-5x efficiency to re-enter the industry’s value distribution.
8. Enterprise AI Transformation Has Become a Top-Executive Project, with Scale and Timelines Both Expanding
高佳辉 observes that over the past year, AI projects at hotel groups have often been pushed directly by the chairperson or CIO. They are no longer about one department buying a specific feature, but about connecting systems across the entire group.
Only the top executive can coordinate this kind of transformation, but broader coverage also lengthens project timelines. 逻辑科技 is therefore taking on group-level AI transformation while turning roughly 80% of common capabilities into standardized products for customers with weaker ability to pay.
吴炳坤 is also seeing Agents gradually combine with hardware and offline scenarios. Enterprises are no longer buying just standard software, but AI-enabled hardware and workflows embedded in daily production that can continuously improve efficiency.
9. Personal Agents and Enterprise Agents Will Form a New Interaction Interface
高佳辉 believes Manus demonstrates a shift in the human–machine interaction paradigm, but it currently still browses the web and simulates a person booking a hotel. In the future, a personal super-Agent may connect directly to a hotel group’s Agent and even bargain on the user’s behalf.
He makes the uncertainty explicit: “When a universal super-entry point will emerge, I don’t know either.” Enterprise Agents may nevertheless appear in large numbers before personal Agents, becoming a new service interface for companies first.
吴炳坤 is more focused on internal enterprise memory. The path from general-purpose models to industry models, enterprise private models, and then Agents is continuous; the next step is for models to internalize memory, creating “expert-level AI doubles” and enabling the “permanent transmission of enterprise intelligence.”
10. The Next Phase Is About Codifying Expert Capability, but Production-Grade Reliability Comes First
杨劲松 expects vertical reinforcement learning to become an important route to Agent deployment. Coding and mathematics have already reached very high levels through reward models and large datasets, and may even “surpass ACM champions.”
His goal is to capture expert capabilities from the way users work with Agents and replicate them in Agents, allowing “1B knowledge workers worldwide to work at 10x efficiency,” potentially reducing the working week for everyone to 4 or even 3 days.
高佳辉 “poured a bucket of cold water” on the idea. Errors in To C may not matter much if the system merely retrieves a few facts or offers suggestions; To B requires accuracy, stability, and traceability at every step. Companies therefore cannot expect to launch a project in January, go live in February, and put it immediately into production. He hopes all 20M hotel rooms nationwide will eventually use the company’s digital employees.
庄明浩 closed by reducing the discussion to 3 phrases: “from language to behavior,” “from describing features to seeing,” and “the industry is still very early and far from mature.” Agents are entering work scenarios, but infrastructure, business models, and delivery standards still need to be built out together.