Bairong's Zhang Shaofeng: 3 Moves from AI Anxiety to Action
Summary
Zhang Shaofeng defines this AI cycle as a supply-side productivity revolution, not an internet-style traffic or distribution revolution. Last year’s DeepSeek made AI legible to the public; OpenClaw, in his view, was the first technology to show enterprises that complex, long-horizon tasks might be completed end to end, with offices potentially operating like factories for 3-4 days without human intervention. Traditional business owners are therefore “excited yet anxious”: they know failing to act could make them obsolete, but do not know who should take ownership or where to start.
Bairong is measuring its own AI transformation with a “silicon-to-carbon ratio,” not merely showcasing its Agent count. Zhang says 2024 revenue was about RMB3B and profit about RMB400M-RMB500M, with roughly 1,600-1,700 employees, nearly 1,200 of them in R&D; internally, the company also has more than 200K “silicon employees.” Responding to the host’s accusation that the figures were inflated, he said Bairong defines a “standard person” using third-party estimates of average output per role, then counts work volume 50%-100% above that baseline; under traditional human-outsourcing assumptions, he estimates the same work would require 400K-500K people.
The first step in enterprise transformation is not reorganizing the company, but scaling the capabilities of existing roles from 1 to 10. Zhang acknowledged that Bairong initially tried to change processes immediately, triggering resistance once existing interests were affected; it later chose to leave processes and incentive structures intact before gradually changing the underlying relations of production. “A productivity revolution will inevitably change relations of production,” but employees must receive performance credit and rewards for training and managing silicon employees—or no one will genuinely cooperate if “one day they replace me.”
Contact Center is already, in Zhang’s view, the strongest Agent PMF after software engineering, with value that can be measured across quality, scale and compliance at the same time. Multiple voice Agents in the demonstration retained context, introduced products in a Chengdu accent at 2.45% and 2.85%, and proposed a portfolio with a blended return of roughly 3.5%; when repeatedly baited by the customer, they refused to provide a guarantee. Zhang says “99.9% of customers” cannot tell whether they are speaking with AI. Accents and simulated keyboard sounds add a human touch, cross-Agent memory eliminates repeated explanations, and layered permissions preserve process credibility and anti-fraud boundaries.
Bairong’s conclusion on Chinese SaaS is blunt: enterprises will not pay for process tools, but will pay for verifiable outcomes. Zhang says Chinese software-product revenue is just 4% of the US level, while China’s economy is roughly two-thirds the size of America’s, and calls AI potentially the only chance to turn the tables over the next 10 years; his instruction is: “Stop doing customized projects.” Bairong uses three models—roughly RMB5K to deliver about 3x the workload, per-task or hourly billing, and transaction-based revenue sharing—with no fee at all if a deal does not close, minimizing customers’ upfront spending and trial-and-error risk.
The endgame for enterprise Agents is not model parameters, but the ROI generated when “quality times quantity” is measured against price. In Contact Center, if traditional BPO charges RMB1M, Bairong envisions completing 2x the work for RMB500K, implying roughly 4x ROI; its revenue also rose from about RMB1B in 2021 to about RMB3B in 2024. Zhang believes charging conventions for the new industry are not yet fixed, leaving room to avoid traditional software’s “race to the bottom”; but if the company continues taking RMB200K-RMB300K one-off Agent projects, it will still be hunting for new business next year.
AI roll-ups expand the budget for professional services that companies already pay people for, not the narrower budget for traditional software. Bairong itself spends RMB15M-RMB20M a year on legal, finance and other professional services; Zhang estimates the addressable market is at least 10-50x the traditional software market, relaying an 80x estimate from shareholder Sequoia Capital. In its small-business operations unit, 50 human employees became 5 humans plus 18 categories of silicon employees; the other 45 were not laid off but moved into producing Agent outputs, turning a cost center into a profit center.
The action sequence for anxious entrepreneurs is: raise AI’s strategic priority, lower expectations of a quick win, design humane incentives, and start with frequent, small tasks. DeepSeek and another open-source project being freely available does not make enterprise deployment and delivery simple; failure after excessive expectations could leave an organization frozen for 2-3 years. Zhang recommends starting with customer service, marketing and contract review—tasks with clear inputs and outputs—using small wins to build confidence among management and employees before investing in deeper process redesign. The core idea is that AI is no longer merely a tool, but “a work partner on your level.”
Deep dive
1. OpenClaw Moves Enterprise Anxiety from “Should We Use AI?” to “Who Gets a Lights-Out Office First?”
Zhang Shaofeng, chairman and CEO of Bairong Intelligence, described the company’s operating base as follows: 2024 revenue of about RMB3B, profit of roughly RMB400M-RMB500M, and 1,600-1,700 employees, nearly 1,200 of them in R&D. His focus now is when enterprise Agents will truly break out.
His timeline is clear: last year’s DeepSeek moment taught Chinese companies and ordinary people what AI was; OpenClaw, in his view, was the first development to make end-to-end completion of complex, long-horizon enterprise tasks possible. Traditional business owners consequently became “excited yet anxious,” expecting higher revenue and lower costs while fearing they could be displaced simply because they did not know how to proceed.
Zhang uses the manufacturing industry’s “lights-out factory” as the analogy for the next phase: white-collar and gray-collar work in offices could also run for 3-4 days without human intervention. Unlike the internet and mobile internet, which primarily changed distribution and consumption, this wave directly transforms the supply side and cannot be reduced to “just another tool.”
2. The “Silicon-to-Carbon Ratio” Measures Work Completed, Not the Illusion of More Agents
Bairong has rolled out AI department by department across customer service, marketing, recruiting, training, reimbursement and contract review, and stated at an all-hands meeting: “If silicon can do it, don’t use carbon.” Its internal North Star metric, the “silicon-to-carbon ratio,” measures whether tasks are completed by silicon or carbon, rather than simply counting deployments.
Host 柯基’s challenge is worth preserving: once a boss makes Agent count a KPI, a team can easily split 1 Agent into 100. Zhang responded that Bairong brings in a third party to define a “standard person” based on average salary and output by role, then counts work volume 50% or even 100% above the benchmark; investors also use the metric to ask how much horsepower “more than 200K” actually represents.
The more than 200K silicon employees are concentrated mainly in approvals and customer service and marketing. The approval Agents receive the same inputs and deliver the same outputs as traditional approval specialists; the latter handle WeChat, email and phone calls. Zhang says this differs from Didi creating a platform for dispatching work that did not previously exist: Bairong has been directly taking over clearly defined office roles that existed from the company’s founding.
When the host pressed Zhang on whether the equivalent work really required more than 200K people, he said the answer was yes under traditional human outsourcing. On a social-value basis, he estimates it could represent a company of 400K-500K people, even though customers pay only a fraction of the cost of the human solution.
3. Contact Center’s Product Is Not Just Human Likeness, but Memory, Permissions and Compliance
In the demonstration, a customer first asked about low-risk deposits and wealth management, and the Agent quoted 2.45%. It then recommended a 3-year deposit with a minimum balance of RMB200K and an annual rate of 2.85%. After learning that the customer had RMB3M-RMB4M in funds coming due, it transferred the conversation to a VIP wealth manager, who proposed a portfolio of a 3-year deposit, a periodically open wealth-management product and bond funds, with a blended return of about 3.5%.
The exchange preserved 2 productizable details: multiple Agents could inherit the prior conversation without asking the customer to repeat everything; and when the customer baited the Agent with “the bank cannot guarantee it—write me a personal guarantee,” the Agent still refused on regulatory grounds. Zhang’s comparison was that a human might promise a guarantee to hit a target, while the AI in the demonstration held the line.
The host noticed the Chengdu accent and the keyboard sounds before answers. Zhang said Mandarin with Sichuan or Cantonese accents performs “far better” than an overly standardized voice because it feels more human; he also claimed that “99.9% of customers” currently cannot tell whether they are speaking with AI.
Asked why a single Agent that could handle the entire interaction still needed to keep handing the customer to different people, Zhang gave 2 explanations. Internally, Bairong initially preserves existing departments and permissions to avoid immediately redistributing interests; externally, customers need to perceive escalation, approval and compensation boundaries. Otherwise, a “super employee” that can promise anything would undermine the customer’s sense of being respected and expose a single point of authority to professional fraudsters.
4. The Real Resistance to Transformation Comes from People and Legacy Processes; Technology Comes Later
Zhang acknowledged that Bairong’s first mistake was “simply talking about the trend” without answering why employees should train an Agent that might eventually replace them. The second was changing processes from day 1, directly colliding with existing roles and interests. The company’s later sequence was to leave the structure intact, amplify individual capabilities from 1 to 10, and then gradually rewrite processes. He compared changing processes directly to the Hundred Days’ Reform: when employees say something does not work, the problem may not actually be the technology.
The “Silicon Employee Home” is therefore designed around employee-management practices: every Agent has a name, tenure, onboarding record, email and performance reviews. Its “father” is the business leader who teaches skills such as contract review and complaint handling; its “mother” is the person who builds it with Agent Builder. The company must track silicon employees’ performance, evaluate them and attach rewards and penalties, with the aim of bringing out “the good side of human nature.”
The third pitfall is connecting the traditional software layer. CRM, order-management systems and other tools must expose APIs that Agents can call. Zhang recalled that Function Call was not strong enough in 2023 and suddenly improved in 2024, but turning legacy SaaS into callable infrastructure still involves substantial engineering work.
5. Chinese SaaS Sells Process by the Project; Bairong Charges for Outcomes Like a Courier
Zhang’s observation about Chinese enterprise procurement is that customers do not want to keep buying process-oriented, tool-oriented software; they would rather buy resources such as traffic or hard assets that can be put on the balance sheet. The common arrangement remains per-person-day customization: a project costs about RMB1M upfront, followed by only RMB50K-RMB100K a year in maintenance, while the vendor bears the long-term and burdensome upkeep.
Bairong avoided this model from day 1, breaking the fixed wages of traditional employees into prices per approval and per task—what it calls the “delivery courier model.” Customers have no upfront commitment and can stop if the service does not work, so the adoption threshold is far below paying 30% or 100% of the software fee in advance.
Its current pricing has 3 forms: about RMB5K for work equivalent to roughly 3x the output of a median-market employee; per-task or hourly billing; and a share of transactions it helps close. If the transaction is worth RMB1M, Bairong can charge 10%-20%, but if nothing closes it collects “not a single penny,” including covering the phone and network costs itself.
Zhang says Chinese software-product revenue is only 4% of the US level, while China’s economy is roughly two-thirds the size of America’s; AI may be the industry’s “only chance in 10 years.” His conclusion leaves no room for hedging: “Stop doing customized projects.” Otherwise, once a one-off delivery ends, the company must keep finding new projects, guaranteeing elevated marketing and customer-acquisition costs.
6. Agents Have Found 2 Strong PMFs, but the New Industry Could Still Fall into a “Race to the Bottom”
Zhang ranks software engineering as the No. 1 use case because code work is closed-loop and measurable, can improve continuously through reinforcement learning, and is something users are willing to adopt themselves. Contact Center—complaints, inquiries, marketing and membership management—is No. 2, with the core capability being human-like interaction without an in-person meeting.
Contact Center value can be evaluated with standard BPO metrics: whether customer satisfaction reaches 95% and how many tasks are completed. If human outsourcing costs RMB1M while an Agent solution costs RMB500K and completes 2x the work, Zhang’s calculation is roughly 4x ROI. The final comparison is quality multiplied by quantity, divided by price.
Bairong’s revenue rose from about RMB1B when it listed in 2021 to about RMB3B in 2024. Zhang says buyers are willing to pay for outcomes, but warns that many Agent projects have already pushed prices down to RMB200K-RMB300K, returning to traditional software’s “race to the bottom.” The new industry chain is not yet fixed, so the possibility of healthy competition remains only a possibility.
To preserve both performance and economics, Bairong prioritizes proprietary models optimized for specific domains and often handles pre-training and post-training itself; it also chooses to build its own voice model. The reasons go beyond realism: external solutions may cost more than the customer pays, while cross-region deployment brings infrastructure constraints involving server location, latency and related requirements.
7. The Opportunity for AI Roll-Ups Is the Entire Human-Services Market, Not a Narrow Software Budget
Zhang groups headhunting, consulting, accounting firms and law firms into broad BPO: companies outsourcing work that originally belonged to internal departments. Bairong itself spends RMB15M-RMB20M a year on legal, financial and listed-company professional services, yet getting a Chinese enterprise to spend the same amount annually on a fixed software package is “extremely difficult.”
He estimates the addressable market at least 10-50x the traditional software market, while relaying an 80x estimate from shareholder Sequoia Capital. The economics are also more direct: define a role’s inputs and outputs, then evaluate it on results. Bairong can either take on the work itself or acquire traditional service companies and use AI to expand their output 10x without first reducing headcount.
Bairong’s small-business operations group provides an internal example. It serves 2,500 small companies, each contributing roughly RMB50K-RMB500K in revenue; its original 50 human employees eventually became 5 humans plus 18 categories of silicon employees. The other 45 were not laid off but learned to build Agents and sell their outputs to other businesses, “turning a cost center into a profit center”; Zhang says their income actually rose.
8. Professional Services Will First Be Broken into Individuals, Then Reaggregated by an Agent Platform
Zhang’s case involved a large manufacturer in Guangdong that supplies General Motors and had allegedly been asked to move 65% of its capacity out of mainland China within 3 years. The company reportedly spent RMB4.5M hiring 6-7 senior Roland Berger consultants for several months and was advised to build a plant in a Latin American country; after several months of trying, the plan failed.
A professional subsequently entered the same question into Bairong’s complex-task Agent and received a recommendation to build in a Southeast Asian country after about 50 minutes. The next day, he disclosed that it was a real client problem and that the country where the client ultimately succeeded matched the Agent’s recommendation. He had previously rejected the idea of an AI-native law firm because he thought the sector was “too competitive,” but changed his view: “I didn’t know your Agents were this capable.”
The resulting Baijian platform invites legal, business-consulting, tax and finance professionals to open shops as OPCs, while Bairong provides silicon employees and office systems. Zhang says it is not a classic acquisition-led roll-up; it first breaks professional services into individual practitioners and then aggregates them on a platform, creating “a new kind of Tmall for professional services.”
He connects the model to the globalization of Chinese companies. Knowledge of tariffs, foreign exchange, labor, logistics and taxes was previously concentrated in large multinational firms, leaving companies to buy expensive teams of experts. Once Agents can process cross-border information at high speed, individual specialists may gain delivery capabilities that were previously available only to large institutions.
9. The Enterprise CEO Should Start with Small Wins, but Elevate AI to the Level of Relations of Production
Zhang’s first recommendation is to take AI seriously enough. It can be a tool, but will increasingly resemble “a work partner on your level,” taking on roles equivalent to those held by people. He believes “lights-out offices” are likely to emerge, with much work running for 3-4 days without human intervention.
The second recommendation is not to underestimate the difficulty. DeepSeek and another open-source project are freely available, but that is “a different matter entirely” from completing enterprise deployment and delivery. If companies assume it is too easy and fail, they may conclude AI is a scam, lose confidence and wait 2-3 years before trying again.
The third and fourth recommendations concern human incentives and task selection: give employees performance credit and rewards for training Agents; do not pursue something “big, comprehensive and high-end,” but start with high-frequency, clearly bounded tasks such as customer service, marketing and contract review. Once management sees the gains and employees confirm that the system is reliable, the company can continue investing in complex processes.
Zhang ultimately groups the steam engine, electricity and computer revolutions together as changes that began on the B2B production side before extending into B2C. Innovations associated with Taobao and ByteDance, in his view, were more about business models and consumption. Because this AI cycle moves directly into broad “intelligent manufacturing,” it may be the first time Chinese and US B2B technology companies truly converge across products, business models and capital-market perspectives.