Summary

AI is labor, not software. That is the episode’s foundational premise. Rolling AI’s 2 partners, both former BCG consultants and serial entrepreneurs, define an FDE as neither presales nor a wrapper, but someone who “puts a digital employee into a company, helps train it up, and watches it go to work,” much like an HRBP. The concept took off after OpenAI and Anthropic announced $1B-scale enterprise AI JVs on the same day in early May, with both positioning themselves around FDEs—the Forward Deployed Engineer role coined by Palantir.
Rolling AI entered the field in 2022, when GPT-3.5 had just opened its API. Nearly 4 years later, it has served close to 100 companies, all on annual contracts with large enterprise clients, at average account values ranging from the RMB millions to RMB10M, with a team of 60+. Its internal rule is absolute: “Every agent must go live within 15 days,” because “when you hire an employee, you don’t send them to school for 3 months before letting them work.”
The key methodological shift came in the second half of 2025, when Rolling AI gradually stopped relying on top-down interviews with star salespeople to extract best practices. It moved to bottom-up “apprenticeship learning”: place an AI apprentice beside every strong store manager, have it earn the right to be taught by first helping the mentor with scheduling and sales forecasts, then distill the mentor’s “street smarts.” The corresponding incentive principle is that a frontline employee who can train a capable AI can have their value multiplied across 10,000 salespeople—so “you have to incentivize the employees who can teach your AI good intelligence.”
The deliberately provocative call: SOP is a sign of backwardness. Standardization means you have only reached 60/100 across the network; AI can make every store an 85 or 90. Headquarters should shift from control—standardization and internal audits—to enablement. The next generation of management thinking “might be led by Chinese entrepreneurs,” who care more about outcome fairness and getting hands-on; the underlying factor may be whether the entrepreneur is an owner, while overseas companies place more weight on procedural fairness.
AI transformation succeeds in fewer than 50% of cases. The 3 main causes of failure are CEO expectations detached from reality; “letting the IT team lead this will almost certainly fail,” because the initiative must come from the business; and failure to change performance metrics and organizational incentives underneath it—for example, moving sales compensation from 100% outcome-based to 80/20 outcome-and-process based. Technology itself accounts for less than one-third of the transformation. This is an electricity-scale productivity revolution, and Lancashire offers the warning: “They connected to electricity, they went online… Simply connecting to a foundation model cannot stop your company from disappearing.”
The core investment call is clear: every JV partner in OpenAI’s and Anthropic’s FDE ventures is a PE firm. Service fees do not capture the upside: “We genuinely help a company save tens of millions or earn tens of millions more, but we charge only RMB6M.” The conclusion is that FDE firms deserve not only VC funding but VC ownership; every PE or VC portfolio-operations team should become an AI enablement center, with each firm serving only 1 company in a given vertical because chain tea, coffee and snack brands compete with one another.
Middle management whose sole job is passing information up and down—2 or 3 layers in a sales organization—will disappear. Execution-focused knowledge workers who type quickly and remember accurately will also have to move up the value chain. The direction is a shift from the era of knowledge work to an era of “cognitive decision work,” where people make only the calls—whether to proceed and where to go. At a rental platform, the definition of a good employee has already changed: “It used to be typing fast and remembering accurately; now it’s being exceptionally good at providing emotional value.”
The talent market brings both a cold shower and a hot one. The blunt message to new graduates is: “You’ve just graduated—I can’t think of anything you can do that AI can’t.” The joking proposal is, “You pay me for the first 2 years, and I’ll pay you back 2x in the years after that.” But a high-school sophomore intern was “not in the slightest inferior to any mediocre consultant with 5 years of experience.” Commercial sense remains a form of near-mystical taste: “I still haven’t found a particularly good way to develop it.”

Deep dive

1. What FDE Means: An HRBP for Digital Employees, Staying Until They Are on the Job

The story began in early May, when OpenAI and Anthropic announced $1B-scale enterprise AI JVs on the same day. Both said they were building FDE businesses—Forward Deployed Engineers, a role originating at Palantir that embeds engineers inside customers to customize and deploy AI systems from scratch. Koji’s joke captured the distinction: “Silicon Valley is good at inventing concepts. We used to call it a wrapper; they call it an Agent Harness. We called it presales; they call it FDE.”
阿甘’s core premise is the foundation for the entire episode: “Traditional software is a tool that someone has to operate; AI itself is labor.” An FDE works like an HRBP: “put a digital employee into a company, help train it up, and watch it go to work.” That means preparing its workbench, materials and company context. “Don’t think of it as an IT solution. Think of it as a new employee starting a job.”
Rolling AI calls itself a business builder rather than an FDE, but the work is essentially the same. It entered in 2022, when GPT-3.5 had just opened its API, and has been doing this for nearly 4 years. It now serves close to 100 companies, all large enterprise clients on annual contracts, with average account values in the RMB millions to RMB10M range and a team of 60+.

2. Leaving BCG: Chinese Entrepreneurs Do Not Worship Paperwork

Both founders came from BCG, where they worked on projects approaching 9 figures, or roughly RMB100M. Their decision to leave was driven by the twin difficulties MBB firms face in serving Chinese domestic companies: pricing and implementation. Chinese entrepreneurs no longer idolize overseas experts as they did in the late 1990s. “Both the first and second generations are thinking about the long-term health of their own businesses. They don’t care how the phase-one paperwork looks; they care whether you can produce results and change the organization.”
The deliverable is fundamentally different from the old consulting model. Back then, the work meant researching an industry, building a framework and delivering a 200-page PPT. Now, “what I deliver is an agent.” The best ways of working and collaborating laid out in the PPT can be implemented inside the agent. A new business model that once took 6-12 months to launch can now go live in weeks. Rolling AI’s written standard is explicit: “Every agent must go live within 15 days.” When you hire an employee, “you give them at most 15 days of training, then they have to start working.”

3. The First Case: 400,000 Nutritionists for 80M Users

In 2022, a dairy company faced a declining birth rate and a shrinking core dairy market. Its second growth curve—premium products such as protein drinks and probiotics—could not move through the existing distributor network and required a new sales model. The math was stark: China had roughly 400,000 registered nutritionists for 80M people who needed service, at a cost of RMB16 per interaction. With GPT-3 newly available, the company fine-tuned a nutrition and health model. The resulting platform supported 6M online users, operated by “just 1 young woman” managing more than 50 bots; per-interaction cost fell to RMB0.10, or even RMB0.04. “Without AI, this was completely impossible.”
The digital employee still needed a mentor before going live. Ask a generic model, “I want to lose weight,” and it will jump straight to a solution. A real nutritionist starts with: “You’re not fat. Why do you want to lose weight?” Only then does the conversation move to the user’s diet and exercise over the past week and their weight-loss goal. The mentor teaches the digital employee the actual service process, including the emotional support that comes before the advice.

4. From Top-Down to Bottom-Up: Apprenticeship Learning and Distilling the Master

Before 2025, the standard approach was to interview star salespeople and the best trainers, summarize their best practices top-down and put those practices into AI. From the second half of 2025, Rolling AI gradually stopped using that method. “There may be no authoritative, top-down right answer to the best way to do business. There are thousands of ways to succeed.” The new approach is bottom-up: an AI bot reviews the day with each store manager, while an “apprentice” assistant shadows every strong store manager and top salesperson, becoming smarter through observation. The system is called “apprenticeship learning,” which Koji says sounds a lot like “distilling an employee.”
Would veteran employees resist having their hard-won know-how distilled away? The answer is to give before taking. The apprentice first helps the mentor—suggesting schedules, assisting with sales forecasts and taking on useful work. “The mentor has to feel the apprentice is useful before they will teach the apprentice.”
The deeper position is that “we don’t think AI is better than people—most of what we find is AI plus people is better than people, and people are better than AI.” The best frontline employees whose experience can be distilled are a company’s greatest assets. AI magnifies their value: a star salesperson once had a practical limit of making a team 2-3x better through coaching; today, that experience and methodology can be replicated across 10,000 salespeople. Incentives must follow: “You have to incentivize the employees who can teach your AI good AI intelligence.”

5. Street Smarts Beat Headquarters’ Algorithms: Storm Scheduling and Chilled Yogurt

Headquarters’ one-size-fits-all sales forecasts are usually wrong: some stores sit beside nightclubs, some face a new competitor, and some are hit by rain. Rolling AI replaces the central algorithm with an AI deputy manager for each store, feeding it local context such as “there may be a torrential storm tomorrow afternoon” or “a nearby store is running a promotion.” The store manager makes the final call; the system does not dictate the answer. Overall forecast accuracy became “much, much higher.” Koji offered a 7-Eleven comparison: after Japan’s 7-Eleven delegated ordering decisions to store managers, performance jumped. Managers previously lacked the time and analytical capacity to collect and process information; the AI deputy manager fills that gap.
Two examples show what “AI is book learning, while Chinese commercial settings contain an enormous amount of street smarts” means. A sharp store manager who knows a storm is coming calls temporary workers the day before to prepare them. Across 1,000 stores, that can save several million yuan a year—yet the practice could not be implemented when pushed down through layers of headquarters. In the chilled-yogurt case, the AI deputy manager concluded that 7 or 8 nearby stores sold large-pack yogurt well and that this store was underperforming because its customer base was too upscale. The manager corrected it: “A Carrefour supermarket 5 meters away is twice our size, and everything there is cheaper.” The AI immediately changed strategy to focus on supplier-exclusive and premium products, lifting sales by 40%-50%. “AI alone can reach a completely wrong conclusion. The store manager and AI together can reach the right one.”
阿甘’s summary: “AI’s biggest dividend is not headquarters’ top-down strategy. It is putting an ordinary person’s wisdom beside every store manager and salesperson. You couldn’t have 10,000 coaches before. Today, you can have 10,000 coaches accompanying them.”

6. The FDE Profile: A Foreman, Not Presales

阿甘’s metaphor for the role is direct: “We are basically an AI labor-outsourcing company today, and the FDE is the foreman.” The foreman takes a group of “young people from Peking University, Tsinghua and Stanford”—the AI—into convenience stores, marketing departments and HR departments. The job is not to drop them off and leave; the foreman stays until they can perform high-quality work. Only 3 things allow the FDE to leave: business integration, knowledge governance and systems integration. Koji’s reaction: “FDE sounds opaque. Say ‘foreman’ and it immediately makes sense.”
A good foreman needs 3 capabilities. First is the consulting skill to identify whether the real bottleneck is headcount, knowledge, communication or collaboration. Second is native human-machine collaboration: break work into smaller trades, put people in the roles of review, judgment and planning, and let machine employees handle the rest. Third is the ability to build: use AI tools to assemble a working prototype in half a day to 2 days, write agents and understand orchestration.
Can people like this be trained? The honest answer is: “Not necessarily in the short term; perhaps they cannot be trained.” The role requires commercial judgment and a particular kind of taste—a gut feeling about which direction to take next that is “a little bit mystical.”

7. SOP Means Backwardness: The Next Management Wisdom May Come from China

The blunt call is that “we believe SOP means slow; it means backward.” Standardization gives the whole network a 60/100 baseline rather than enabling every person to reach 90. AI can ensure every store reaches 85 or 90. The management shift is from using standardization to protect the floor to using good intelligence at the frontline to provide the best answer. Headquarters’ standardization, internal audit and procurement functions will lose prominence as headquarters becomes an enablement layer. The change will extend beyond management practices into organizational design and departmental structures.
Why was this impossible before? “An organization did not have enough intellectual productive capacity—people who truly understood, could explain clearly and could coach others. Today, once compute is running, you have unlimited intellectual productive capacity.” Koji’s analogy is Douyin’s personalized distribution on the consumer side. Now production is being personalized: every smallest business unit can produce in a highly tailored way.
Why might China lead? “Overseas companies pursue procedural fairness; domestic entrepreneurs pursue outcome fairness, and any path is acceptable if it gets the result.” Chinese entrepreneurs are also more hands-on, building agents themselves. “The underlying reason may be whether the entrepreneur is the owner of the company—that determines the vast majority of it.” This mindset may arise only in a market as intensely competitive and fast-moving as China.

8. This Is an Electricity Revolution, Not a Tech Wave: Electrified Companies Still Die

Technology accounts for “no more than one-third” of enterprise AI implementation. The broader judgment is categorical: “This is not a technology wave. It is a complete productivity revolution, and its impact on society will exceed that of the internet.” The electricity revolution replaced physical labor at scale; the foundation-model revolution will replace cognitive labor at scale. We are still only in the early years of a transformation that could last 10, 15 or 20 years.
Lancashire offers the episode’s most consequential historical analogy. The region once accounted for roughly 70% of global textiles. “Did it use electricity? It did. But it only connected electricity to the same central shaft that had powered the steam engine.” Its production system remained fundamentally steam-era. The US, Japan and Germany rebuilt their production organizations around electricity-native methods—the equivalent of becoming AI-native today.
Taken to its extreme: “Every major productivity revolution causes 95% of companies to disappear. And believe me, every company that disappeared had electricity; every one of them went online.” They did not disappear because they stayed offline. They disappeared because they failed to use the new force to rebuild their businesses from the ground up. Simply connecting a foundation model cannot stop a company from disappearing.

9. Transformation Success Is Below 50%: 3 Traps and a Compensation Overhaul

“The success rate of AI transformation projects should be below 50%.” There are 3 major causes of failure. First, CEOs expect AI to make the company take off immediately. Second, and most importantly, “do not let the IT team initiate the project; let the business team initiate it.” The business team knows how to select the product mix and get customers to buy insurance. “If the IT team takes on AI, it will fail.” In multinational companies, the first fight is usually against IT governance, which puts data security first. Third, companies fail to change the incentives beneath the technology. “Don’t view this as a tool going live. It is a group of employees going live, so the relations of production have to change.”
The concrete compensation fix is to work with the business team because salespeople will not use anything unrelated to their own interests. Change sales incentives from 100% outcome-based to 80% outcome and 20% process, with more activities converted into points that can be exchanged for cash. “Without pushing these things through, implementation is impossible.”
The chairman of Bloomage Biotech made the stance explicit at an AI workshop: “If you’re only going to achieve a 50% productivity gain, don’t do it. Do something 3x, 5x or 10x.” Once the business is laid open to an effectively limitless supply of intelligence, there should always be opportunities to rebuild the business model or at least redesign the workflow.

10. The Rental Manager Case: AI Is Not Cost Cutting but Commercial Restructuring

The client initially asked AI to help rental managers reply faster. Once Rolling AI entered, it found the real problem: urbanization was already high and the rental business itself was declining. The business should be selling home services, pet care and other value-added services, but managers were trapped in low-value disputes over noisy neighbors and leaking air conditioners. The new division of labor puts warm, emotionally attuned care with people and all the routine work with AI. Manager productivity rose from 1 manager per 500 customers to 1 per 1,200, with a target of 1 per 2,000 this year. The department made no layoffs, while service quality, favorability, renewal rates and service sales all improved. “We don’t think AI entering the business should be a cost-cutting exercise.”
A sales-opportunity agent surfaces specific offers: “You can see this tenant has a cat, and they said they’ll be traveling for a few days. You should sell them at-home pet feeding.” The person in charge preserved the shift in talent standards: “The definition of a good employee has changed. It used to be typing fast and remembering accurately; now it’s being exceptionally good at providing emotional value—‘Don’t rush, I’ll take care of it.’ Making people happy and relaxed is what human talent can do.”

11. Psychological Counseling Is the Core of Consulting: The Role Disappears, Not You

阿甘 restated his old view that “the essence of consulting sales is psychological counseling for the chairman and CEO.” In the AI era, management is fundamentally about releasing goodwill. If your intent is to manage, constrain and restrict, “its energy is 10x”; if your intent is to empower people and help them earn more, that can also be 10x. “If you are trying to control people, replace all of it with robots.” The conversation with a CEO should therefore focus on where their intent to empower lies, then encode it in the AI deputy manager.
The heavier psychological work now falls on frontline employees. The message is: “This is not about replacing you. What disappears is not you, but the outdated role.” Employees should not seek development inside a backward role; they should look for new roles that fit them better.
Which roles disappear? “Any manager whose only job is information transfer—passing instructions up and down—should disappear.” In a sales network running from headquarters to provincial heads to regional managers, 2 or 3 layers exist mainly to transmit information. They used to clean up information—turning a story such as “this customer is an idiot” into a structured account—and aggregate it layer by layer. AI now does that better. Execution-focused knowledge workers face the same choice: learn to drive agents or develop expert and street knowledge that foundation models cannot retrieve. The broad direction is that “humanity will move from the era of knowledge work into the era of cognitive decision work: I make the decisions—whether it works or not, and where to go. I no longer write proposals, draw designs or write code.”

12. The New-Graduate Paradox: Pay to Work, Equal Footing with a High-School Intern

The episode recounts an actual interview in which an excellent new graduate asked what value they could bring to the company. 阿甘’s honest response was: “You’ve just graduated—I can’t think of anything you can do that AI can’t.” The joke was: “You pay me for the first 2 years, and I’ll pay you back 2x in the years after that.” The serious concern is that pure desk-research consulting no longer offers enough room to train young consultants. “We are also worried that this industry will develop a generational gap.”
The paradox runs the other way for young people with the right traits: on day 1, their capability can match that of a mid-career consultant because anything they can write, you and your AI can write too. The company’s youngest intern is a high-school sophomore, and it “does not feel at all that he is less seasoned than any mediocre consultant with 5 years of experience.” Selection is shifting from age and tenure to judgment and curiosity: some people lower their heads when they see a problem, while others’ eyes light up. But how do you teach commercial sense? “I personally have not found a way to develop it. Some things are innate.” People from families that run small businesses and those with strong powers of observation can be more useful. The counterintuitive advice for graduates is to start at a larger platform and develop commercial sense and judgment there.

13. Why OpenAI and Anthropic Are Entering Directly—and Why Every JV Partner Is PE

There are 2 drivers. Public-domain data is limited, and foundation models hit a shortage of industry data and domain knowledge when they enter verticals. Anthropic can train models across many tasks, but experts can charge RMB10,000+ per hour. Rather than hire them one by one, it is more efficient to go deep into industries and solve the problems directly. And B2B is fundamentally a services business: “The real heavy lifting does not come from plugging in a model. It comes from the operating model—the workflow, the organization and the people.”
The most tradable investment insight is the structure: every JV partner in these OpenAI and Anthropic ventures is PE. Service fees do not capture the upside. “We genuinely helped a company save tens of millions or earn tens of millions more, but we charge only RMB6M.” Only by entering alongside PE and transforming portfolio companies can the FDE provider capture the capital-side value. The implication is that AI will be the largest corporate growth enabler over the next 10-15 years, and every PE or VC portfolio-operations team should have—or become—an AI enablement service center.
The more aggressive conclusion is that FDE firms “deserve not only to be funded by VCs, but to be owned by VCs.” They should not be treated as portfolio companies; they should be treated as part of the investor itself, serving its portfolio. There is an exclusivity constraint: chain restaurants, tea brands, coffee chains and snack retailers compete with one another. “Once we serve one, we cannot serve another.” Each vertical can work with only 1 PE or VC. Rolling AI had served PE portfolio-operations teams at BCG, including its largest client, “大马西.” The difference now is confidence that strategy can reach the frontline: “At least we can say with confidence that we will definitely bring upside to your business, so we have the confidence to bet with you.”

14. Service as Software: Results Are the Service

The historical case for MBB is that it solves the CEO’s core problems, operates closest to revenue and power, and “still demands results.” But “the best era of management consulting in China has not yet arrived.” That era requires companies to be standardized and leaders to be professional managers; it may still be another 30 years away. Either MBB firms will scale up in China, or Chinese consulting firms will reach MBB scale. The criticism of the old model is clear: PPT businesses have limited growth; once you turn PPT into agents serving the frontline every day, “it has to become more substantive.”
The naming moment came when an investor said Rolling AI was neither consulting nor SaaS. “Calling everything SaaS now feels a little insulting.” The answer was Service as Software: the company does deliver software—AI deputy managers and AI salespeople—but the core product is a service. Koji’s closing formulation was: “Result as a service. Results are the service.”
The business discipline is equally strict. Rolling AI speaks only with the client’s top decision-maker and usually reaches an agreement within roughly 3 meetings. At the start of the year, it encountered software-buying deals worth around RMB6M-8M. “No one could clearly explain the business objective, there was no owner, and the target kept moving. That is the most dangerous kind.” The organization is 60+ people and extremely flat; the 2 partners still write code every day. The team operates like a SEAL unit, forming temporary squads around a specific objective. FDEs must stay on the frontline and work inside the stores.

Verification Notes

“大马西” (transliteration uncertain): the raw captions do not provide enough information to uniquely identify the entity, so the original name has been retained.

“Pessimists Are Right, Optimists Succeed”: A Conversation About the AI Industry with My Friend 亚婷 Uber CEO on AI, Autonomous Vehicles, and the Future of Transportation