Summary

The core of this software-stock re-rating is that enterprises are shifting what they pay for from seats to auditable business outcomes. The episode says global software stocks lost nearly $1T in a week after Anthropic launched 11 enterprise plugins on Claude Cowork; 张韶峰 argues that the technology did not suddenly leap forward, and that OpenClaw and these plugins were merely “the last straw.” Licenses will not disappear immediately, but “the curtain has definitively risen on a new paradigm”; the old model’s decline and eventual extinction are only a matter of time. He also recalls that in 2021 investors asked 百融 to adopt American-style SaaS, while he argued that China could skip SaaS and go straight to TaaS (Transaction); investors later suggested BaaS (Business, business-level services).
Private data and complex workflows are the main factors that can extend SaaS’s lifespan, while industry know-how confined to human minds is not a durable moat. General-purpose models struggle to replicate applications without access to private-domain knowledge; a workflow may contain 20 steps, and 张韶峰 uses repeated multiplication by 0.99 or 0.98 over 25 steps to show how error probabilities accumulate, potentially falling below an enterprise’s tolerance threshold and giving traditional vendors perhaps 3 years of “redemption.” 刘一鸣 immediately argues that it may not even be 3 years, asking whether that estimate is too optimistic; 张韶峰 does not directly confirm this, instead turning to a third category of barriers. He still believes that almost no Chinese SaaS companies meet these conditions, and that only a small fraction of American companies might.
RaaS expands the software budget to cover salaries, business outsourcing and end-to-end outcomes, and 张韶峰 estimates that its market is at least dozens of times larger than SaaS. 百融 cut internal contract review from 56 minutes of human labor per contract to 4 minutes, with a silicon-based employee handling the remaining 52 minutes; pricing can therefore be based on roles, hours or units, with costs potentially just one-third to one-fifth those of carbon-based employees. AI BPO can go further and price on end-to-end outcomes; 张韶峰 compares this with traditional sales channels taking 20%, 30% or 40% of sales, without saying that 百融 itself uses those fixed rates.
For traditional software giants, their customers and installed-base revenue are both a moat and the heaviest obstacle to self-disruption. Moving from one-time On-premise revenue plus 15%–20% maintenance fees to SaaS, where upfront revenue is lower and renewals uncertain, creates an inherent conflict of interest; AI transformation also faces legacy customers with much lower tolerance for errors. Startups, by contrast, are “barefoot”: they have no legacy revenue to cannibalize, and early users are more willing to accept mistakes. By the time the new paradigm is broadly accepted, the “new king” may already have a mature product and organization.
AI Agents are being treated as governable, measurable and retireable workers rather than merely software features. 张韶峰 says 百融’s more than 1,000 carbon-based employees correspond to 200K–300K silicon-based employees, putting the silicon-to-carbon ratio at roughly 130–150:1; silicon-based employees have organizational structures, employee IDs, email addresses, tenure, JDs, KPIs and designated trainers. When performance is poor, they may first improve themselves, then be retrained by carbon-based partners or “retired from service”; when they perform well, “all a silicon-based employee needs is electricity and chips,” while ultimate responsibility and economic incentives remain with humans. He also once built a silicon-based assistant to handle internal approvals, and the need was later standardized into a company-wide silicon-based employee.
Automation does not necessarily mean layoffs in equal numbers, but whether workers can migrate into new roles will determine how the gains are distributed. 百融 originally used 50 carbon-based employees to serve 2,500 small-business customers, each contributing RMB100K–500K in annual revenue; it later shifted to 18 silicon-based employees and 5 carbon-based employees. The remaining 45 people were not laid off: they learned to produce agents, moving from a cost center to a profit center, and revenue increased. 张韶峰 retained an important caveat: “I’m afraid not every enterprise can do this.”
The value stack for enterprise Agents may sit across dedicated models, AI PaaS, IAVs and Agent Stores rather than being captured by a single general-purpose model vendor. 百融 says its dedicated models handle more than 95% of problems, with the remaining roughly 2%–5% of corner cases routed to general-purpose models; AI PaaS fills in the common capabilities traditional ISVs need to become Independent Agent Vendors. If an ecosystem partner creates ¥100 of revenue, it could notionally take ¥70 and the platform ¥30; the principle is, “If you don’t make money, we don’t charge you either.”
Recruiting, legal, business and tax services show that Agents will first absorb search, communication, unstructured-content production and multi-step workflows, while humans retain review, judgment and accountability. Recruiting Agents can understand the capabilities behind “microservices” rather than merely search the keyword, and candidates may be required to use Cursor, Claude Code, Codex or Trae on-site; 百鉴 cut internal contract review from 56 minutes to 4 minutes and generated cross-border consulting reports in tens of minutes. 张韶峰 expects that in 3 years, professionals may “be left with only one respectable job: review and sign-off,” but that signature will still represent credentials, judgment and final responsibility.

Deep dive

1. The software empire built on selling seats by the head is giving way to outcomes

The episode opens with Anthropic’s launch of 11 legal, finance, sales and data-analysis plugins on Claude Cowork, and says the global software sector lost nearly $1T in market value in a week. 张韶峰’s view is not that technology suddenly jumped: OpenClaw had already shown that Agents could work like humans; the enterprise plugins simply made the threat visible to a sufficiently large audience.
张韶峰 puts the Chinese market even more bluntly: “Software as a product industry has never existed.” Its output may be only 4% of America’s. American SaaS enjoyed at least 15 years in the sun, but what enterprises really want to buy is a partner—silicon-based or carbon-based—that can solve their problems one by one.
He recalls that in the winter of 2021, investors asked 百融 to become an American-style SaaS company. He argued that China could skip SaaS and move directly to TaaS, or Transaction as a Service: complete transactions for customers and share the value based on results. Investors later suggested BaaS, with B standing for Business and emphasizing business-level services.
Licenses and subscriptions will not die immediately; the old and new paradigms will coexist for a long time. The direction of the industry, however, has already changed. 百融 has charged by outcomes since its founding because “the curtain has definitively risen on a new paradigm”; the old model will gradually move downmarket, and its eventual extinction is only a matter of time.

2. Private data and complex workflows can buy only a temporary “redemption period”

The first hard moat is private-domain data. If critical industry knowledge is embedded there, OpenAI, DeepSeek, Qwen or Google Gemini cannot replicate the application using only a general-purpose model. “Even a major vendor can’t build it, because it doesn’t have your private-domain data.”
The second is complex workflows that sit below an enterprise’s error tolerance. A workflow may contain roughly 20 steps; 张韶峰 again uses 0.99 or 0.98 multiplied continuously over 25 steps to show how errors accumulate until the result becomes unacceptable. He therefore believes traditional ERP or SaaS may have a 3-year buffer. 刘一鸣 immediately challenges him that it may not even be 3 years, asking whether the estimate is too optimistic; 张韶峰 does not directly confirm this and instead turns to the third barrier.
Industry know-how that exists only in the minds of carbon-based employees is a softer moat because competitors can poach the people who hold it. It must be converted into private knowledge, data, complex workflows or a combination of the three. 张韶峰 estimates that “almost no” Chinese SaaS companies meet the criteria, while only a small portion of American companies may qualify.
Traditional software will not necessarily disappear. If it is cheap enough, silicon-based employees may still use it as a screwdriver. “Software is a tool that carbon-based life uses to complete an end-to-end task”; the user is simply becoming silicon-based. If the tool is not optimized for the new user, an Agent company may build a more suitable wrench of its own. AI-native companies and SaaS vendors will therefore both cooperate and compete.

3. For the “old kings,” the hardest barriers are installed-base economics and legacy customers

张韶峰 uses Kodak and Nokia to explain the predicament of the “old king”: new technology is often already in hand, but adopting it would damage existing revenue. On-premise software generates a one-time payment followed by maintenance fees of roughly 15%–20%; switching to SaaS lowers upfront revenue while making renewals uncertain, so self-cannibalization is difficult to embrace.
The second lock comes from customers. The enterprise customers loyal to mature vendors are often conservative themselves, and can use a single error in any new solution to prove that it is “not as good as before,” further locking in the supplier. This is the same logic behind the market’s earlier belief that Google would not dare disrupt its search revenue.
Startups have no installed-base burden: “I’m barefoot; all I have is change.” Their internal champions are also more likely to be AI-native newcomers who grew up with ChatGPT, Gemini or Qwen. By the time the new paradigm is broadly accepted, these people may have become the core of the organization, and the market may already be saying, “This simply isn’t your business anymore.”

4. “Silicon-carbon co-governance” brings Agents into the formal organization

刘一鸣 cites 黄仁勋’s vision of Nvidia growing from 32K employees to 50K humans plus 100M AI assistants, with Agents autonomously recruiting other Agents and maintaining their own Slack channels. 张韶峰 says he agrees 100%, noting that 百融 presented an organizational chart for “silicon-carbon co-governance” at its annual meeting before the 2025 Spring Festival.
百融 currently has more than 1,000 carbon-based employees and 200K–300K silicon-based employees, with 张韶峰 giving the silicon-to-carbon ratio as roughly 130–150:1. The company has around 200 types of silicon-based roles, spanning front office, middle office, back office and functional departments; a single role can also contain multiple silicon-based employees.
“Silicon Employee Home” displays departments, teams, names, email addresses, employee IDs, tenure, start dates, JDs and performance, much like an HR system. Metrics include working hours and successful responses. Each Agent generally corresponds to a carbon-based partner responsible for training, deployment and ultimate accountability; underperformers may first improve themselves, then be retrained or “retired from service.”
张韶峰 says he once built a silicon-based approval assistant during Chinese New Year’s Eve and the first day of the lunar new year, initially to judge whether colleagues’ requests were urgent. If a request was genuinely urgent, it could notify him, call him or send an email. The need was later standardized into a company-wide silicon-based employee.
The organization may not retain a strict “carbon above, silicon below” hierarchy. 张韶峰 believes humans reporting to silicon-based superiors will certainly be technically possible, with a future structure perhaps forming an interwoven silicon-carbon network, although the very top may remain carbon-based. Online learning also means Agents need not wait for the underlying model to be upgraded; they can evolve based on work results and colleague feedback.

5. The threshold for a “human feel” came before ChatGPT; production validation arrived in 2020

百融 launched its voice-Agent project in October 2017, roughly one quarter after the Transformer paper was published, and released its first version in 2018. Initially, third-party research found that around 15% of customers thought the voice sounded mechanical and the logic rigid. By around 2019, perhaps only a few per thousand customers could detect that it was AI, or merely felt that something was odd; by the time ChatGPT was released, 张韶峰 says the figure was down to a few per ten thousand.
百融 says its AI handles roughly 100M phone calls a day, with adjustable speaking speed, emotion and dialect. When 张韶峰 tested it in Sichuanese, he felt the Agent “already speaks much more standardly than I do.” While serving 8K–9K enterprise customers, the Agent could also answer complex questions over WeCom about bills, business consumption, hang-ups and network jitter, and return spreadsheets directly.
The real “aha moment” came at the start of the COVID-19 outbreak in 2020. When large numbers of China Postal Savings Bank call-center employees could not return to work, the bank demanded an emergency AI deployment. The feedback was that users “basically couldn’t tell it was AI.” For the first time, 张韶峰 became convinced that “we may have actually pulled this off.”

6. Employees resist externalizing knowledge; the transition turns on whether they can put on the “Iron Man suit”

刘一鸣 discusses port dispatchers whose veteran workers regard experience as their livelihood. Even when an Agent project required them to convert their know-how into text, they refused to contribute, and management pressure was ultimately needed. 张韶峰 adds the example of veteran traditional Chinese doctors who both “look down on AI” and refuse to hand over knowledge that becomes more valuable with age; only an incentive system can resolve the problem.
张韶峰 describes the replacement dynamic not as “AI versus humans,” but as a productivity race among humans: “The people who replace you in the future won’t be AI; they’ll be carbon-based people who are better at using AI and put on the Iron Man suit earlier.” In this metaphor, AI is not the opponent but the exoskeleton that turns the human body into “JARVIS.”
百融 once used 50 carbon-based employees to serve 2,500 small-business customers, each contributing RMB100K–500K in annual revenue, making offline service uneconomical. After shifting to collaboration between 18 silicon-based employees and 5 carbon-based employees, the other 45 people were not fired; they moved into producing Agents for other enterprises, turning a cost center into a profit center while increasing revenue.
This is not a guarantee for everyone. 张韶峰 explicitly acknowledges that some people will get stuck on changing their mindset or lack the willingness to learn. 刘一鸣 closes with the example of ATMs, emphasizing that technology may change roles—from counting cash to financial planning—rather than simply deleting jobs in proportion to the number of machines.

7. RaaS expands the business model from software gross margin to the entire operating cost pool

张韶峰 expands RaaS as Results as a Service and estimates that its total market is “far larger than SaaS—not just 2x or 3x, but at least dozens of times larger.” SaaS is merely the wrench and screws, whose value is difficult to measure; RaaS delivers end-to-end outcomes or clearly defined tiers that map directly to a customer’s labor costs, workload or revenue.
百融’s silicon-based contract-review specialists cut carbon-based time per contract from 56 minutes to 4 minutes, allowing the remaining 52 minutes to be converted into labor costs. If quality stays constant while output rises and costs are only one-third, one-quarter or even one-fifth those of carbon-based employees, the service can be priced by monthly salary, hours or units.
The first model is AI employee dispatch, settled by fixed compensation for a role, time or workload. The second is AI BPO, which outsources entire processes such as call centers or sales and prices on end-to-end outcomes. 张韶峰 uses traditional sales channels taking 20%, 30% or 40% of sales as an analogy for the pricing logic of outcome sharing.
The third model is to open 百融’s AI PaaS and let ecosystem partners become IAVs, or Independent Agent Vendors. A partner could take ¥70 of every ¥100 in revenue while 百融 takes ¥30, and could also receive technology, know-how or even capital support. The constraint is: “If you don’t make money, we don’t charge you either; we charge only when you make money.”

8. Enterprise Agents need dedicated models and PaaS; general-purpose models are only the “brain”

张韶峰 corrects the claim that 百融 does not build models. The company does not target AGI or a general-purpose model, but it does train dedicated models for specific industries and domains. Voice applications may use pre-training from scratch to control response time and human-likeness; other applications use post-training on open- or closed-source models.
By his account, dedicated models can handle more than 95% of problems, while the remaining roughly 2%–5% of long-tail corner cases are routed to general-purpose models. This architecture addresses the trade-offs that make general-purpose models slow, prone to hallucinations, inaccurate or expensive, while the Agent Builder orchestrates Agents—that is, silicon-based employees.
He uses cloud computing to explain why AI PaaS is necessary. PaaS emerged above IaaS because SaaS vendors did not want to repeatedly build common capabilities; likewise, traditional ISVs that move directly from a major vendor’s foundation model to becoming IAVs will lack the intermediate infrastructure required for enterprise services. Chinese tech giants have relatively little ToB exposure, making it harder for them to understand these delivery “pains” firsthand.

9. Agent Stores do not naturally belong to model vendors; enterprise applications can remain multi-platform

百融 envisions an Agent Store that accommodates both internally developed and ecosystem Agents, with end customers choosing freely and the platform and developers sharing revenue. Silicon-based employees are also divided into EX and CX: EX serves employee experience, such as contract processing and recruiting; CX faces customers directly, handling complaints, marketing and product introductions.
刘一鸣 asks why OpenAI or Anthropic cannot simply take this layer. 张韶峰 says they certainly can: OpenAI’s GPTs were essentially an Agent Store, but “it hasn’t even mentioned them this year; it never got off the ground.” Coze is closer to a ToC Agent Store, while the enterprise-grade store has at least not yet been fully figured out.
Anthropic is focused on productivity and ToB and may eventually enter enterprise-Agent distribution, but that does not imply a monopoly. 张韶峰 compares the structure with IaaS: Amazon, Microsoft, Google and Oracle have coexisted for years, so the model, PaaS and application-ecosystem layers are more likely to form a dynamic structure of both cooperation and competition.

10. Recruiting, legal, business and tax services reveal the clearest substitution paths

刘一鸣 asks whether recruiting cycles could shrink from 28 days to 2 days. 张韶峰 does not independently verify that figure, instead breaking down the mechanism: Agents can conduct hiring interviews, organize JDs, search resumes, and schedule appointments by email and phone, eliminating the operational burden of preparing dozens of SIM cards and preventing high-frequency calls from being blocked.
The machine’s advantage is not just speed. A human recruiter may mechanically search for the three words “microservices,” while an Agent can expand the concept into specific work and determine from a candidate’s lengthy description whether they possess the relevant capabilities, reducing the repeated back-and-forth with hiring managers to widen the search.
百融 requires some candidates to use AI on-site to complete assignments: “I don’t want you to write code manually.” Applicants can open Cursor, Claude Code, Codex or ByteDance’s Trae; the same applies to copywriting and video roles. The company reviews each quarter whether departments have added silicon-based employees that are actually deployed. In the future, people who cannot use AI “will probably not be hired.”
The legal, business and tax platform 百鉴 lets senior directors or junior partners aged 35–40 call on silicon-based professionals to generate deliverables such as Word documents and PowerPoint decks, then revise them through dictation and provide the final sign-off. The organizational logic is to siliconize junior and mid-level execution while leaving carbon-based professionals with the expertise, credentials and accountability for review and final responsibility.

11. The end state for professional services and traditional enterprises depends on end-to-end delivery

百鉴 grew out of 百融’s experience cutting internal contract review from 56 minutes to 4 minutes. In one case, an Agent completed a manufacturing company’s overseas-expansion report in tens of minutes, automatically searching competitors and other General Motors suppliers for cross-validation. The partner said this kind of fifth-chapter validation was normally performed only by senior mid-level or above partners.
The company had previously hired a global top-five consulting firm that deployed 5–6 professionals for 5 months and charged roughly RMB5M. The client invested nearly RMB100M to build a plant but failed, then succeeded at another location; the Agent’s report directly recommended the site that later worked. The partner rated the original report at more than 200 pages and perhaps 90 out of 100 in form, but wrong in conclusion. The AI report was around 80 pages, and professionals could expand its presentation in 2–3 days.
百融 subsequently equipped more than 100 partners at the institution with silicon-based employees, saying each professional’s income increased by at least 3x. Cross-border consulting projects below RMB1.5M had previously been difficult to take on because of legal, labor, tariff, logistics and cross-border coordination costs. 张韶峰 expects professionals in 3 years may mainly be responsible for “review and sign-off,” but judgment, credentials and accountability remain indispensable. He also relays the McKinsey CEO’s claim that 25K of its 60K employees are silicon-based.
张韶峰 identifies 3 types of work best suited to Agents: human-like interaction; generating and processing unstructured data such as PowerPoint, Word, PDF, images, video and audio; and orchestrating multi-step workflows with some degree of complexity and flexibility.
For traditional enterprises, he recommends first distinguishing between foundation models and Agents: “The large model is only the brain.” A complete Agent also needs hands, feet and the nerves connecting them. Cement, spicy-snack, chili-powder or financial companies with annual revenue of RMB1B–2B do not need to obsess over the underlying model; they only need to ask about quantity, quality and cost. Most enterprises are better off buying or customizing Agents that can start working immediately.
On “one-person companies,” 张韶峰 believes ToC will arrive before ToB. A Prosumer could lead a large workforce of silicon-based employees, and if valuation systems reach 100x sales, “a one-person $1B unicorn is entirely possible.” ToB still requires validation by large enterprises, trials with personal assistants and trust built around critical decisions, but it is ultimately “entirely possible” there as well.

The Ruthless Sales Culture Behind ElevenLabs Growth | Carles Reina Debt Spiral or NEW Golden Age? Super Bowl Insider Trading, Booming Token Budgets, Ferrari's New EV