Summary

朋新宇, CEO of Alibaba’s Lingyang, says the fundamental difference between FDE and rebranded field operations is that “the deliverable has shifted from functionality to a business outcome.” He rejects the idea that “changing the title can upgrade the salary” as inconsistent with business logic: OpenAI’s mid-level FDE roles in San Francisco pay a base salary of $160,000 to $280,000, while traditional solutions engineers earn roughly 60% more than that role. The premium pays for a scarce combination of 3 capabilities—AI sharpness, industry depth and data breadth—with outcomes that can be measured and verified.
泓君’s macro framing is that the AI battlefield is moving from consumer applications back to enterprise, and the $6 that customers spend implementing every $1 of software is being reallocated. OpenAI has set up a dedicated deployment company; Anthropic has joined Blackstone and Goldman Sachs in a $1.5B joint venture; Microsoft has created Microsoft Frontier Company with $2.5B of investment and 6,000 engineers, saying it aims to move beyond the FDE concept itself; and Amazon is investing tens of billions of dollars in a similar organization. 朋新宇 likewise says he sees “the entire AI battlefield moving from C back to B.”
The “chase shipment” use case runs through more than 260 internal steps, which is where the real difficulty of enterprise AI lies. The 3 types of shipment inquiries—missing items, missing accessories and undelivered gifts—cut across internal and external systems and multiple e-commerce platforms, with more than 260 steps covering over 95% of cases. Lingyang’s customer-service Agent outperformed traditional bots from launch, and in postmortems came “close to, and in some cases exceeded, the average performance of humans at their best.” The rule for selecting use cases is to target wherever labor, money or time consumption is highest. Customer-service costs can equal 2% to 3% of sales at some companies and 3% to 5% at others; AI only works if it delivers a better business outcome than the existing process at the same cost.
The right analogy for AI implementation is not delivering a system but hiring a new employee—one that arrives with 5 to 10 years of job skills and only needs permissions, goals and discretion. The product takes a “4+X” form: 4 preset Agents for marketing, sales, customer service and operations, plus X for the company’s data assets and context governance. There are 2 commercial models: seat-based pricing, potentially replacing labor at an investment equal to 80% or 90% of the human cost; or a share of the gains above the human conversion rate. Outcome-based pricing “definitely exists,” and its customers are continuously renewing and receiving ongoing service.
Unlike the Palantir-style customization route, Lingyang is betting on using standardization to solve personalized problems and sees the marginal return on person-day customization as low. The best bidder made more than 10,000 bid adjustments in 3 months, while an average bidder might make only a little over 100 in 100 days; one bidder made 1 operation a day, while another averaged 100 operations a day. Lingyang digitizes and models benchmark employees’ behavior, then uses Harris Engineering to learn and evolve autonomously. When a customer asks for “1,000 person-days” of custom development—“from the chicken farm to raising chickens, collecting eggs and scrambling them, with every task completed”—the marginal benefit for both sides is low.
The biggest obstacle to deployment is production relations, not technology: top-down projects work, while bottom-up initiatives can hit the wall between IT and business budgets. In one successful case, the chairman of a manufacturing company personally delivered more than 15 AI training sessions within a year. After a pilot, the company’s least advanced sales province became its most advanced sales region in the first half of this year. Employees are splitting along a “K-line evolution”: internal FDEs who understand both AI and the business gain value, while opponents lose value—“you can oppose a tool or a project, but you cannot oppose this trend.”
Token bills and model selection are deliberately taken out of the customer’s line of sight: one Lingyang AI project could have generated RMB3M of revenue against RMB12M of investment at the end of the 1st quarter, then gradually reached near breakeven by the end of this quarter. The standard is whether the economics work. A private deployment of a large model with a parameter footprint above 2TB may require an upfront investment of several tens of millions of yuan; the “building blocks” are now strong enough that nobody talks about post-training anymore. When an Agent calls a large model, the model itself does not absorb the data; data may instead be collected by Agents used for programming or web coding. Consumer products such as ChatGPT may also collect relevant data to optimize the product, and users can turn off the “use for training” setting. Model upgrades require human validation and confirmation rather than automatic switching: enterprises want stable work output.
Enterprise leaders’ attitudes have changed sharply this year, jumping straight from “how do we use it, and can we use it?” to “when can you help me put it to work?” 朋新宇 sees the inflection point after Lunar New Year, with the rise of “Lobster,” followed by Harris Engineering and current trends such as 路普. Attention has shifted from model parameters and benchmark scores to whether AI can execute tasks. His view is that “entrepreneurs are the smartest and most sensitive group of people.” China has no independent SaaS companies, but its e-commerce platforms carry SaaS capabilities; that is the foundation for Chinese-style FDE to “stand on the shoulders of giants,” and the biggest structural difference from the Salesforce and SAP path in the US.

Deep dive

1. The $6 Business Is Being Reallocated: Tech Giants Rush Into Enterprise

泓君 opened with an old rule of the enterprise software market: for every $1 a customer spends on software, it spends another $6 putting that software to work. In the AI era, that $6 is being reallocated—“whether AI is a bubble will not be answered in the chat window, but in the enterprise ledger.”
Her chain of evidence: OpenAI has established a dedicated deployment company; Anthropic has formed a $1.5B joint venture with Blackstone and Goldman Sachs; Microsoft has created Microsoft Frontier Company, backed by $2.5B of investment and 6,000 engineers, and says it intends to move beyond the FDE concept itself; Amazon is investing tens of billions of dollars in a similar organization. 朋新宇 says he sees “the entire AI battlefield moving from C back to B.”
The guest is 朋新宇, vice president of Alibaba Group, CEO of Lingyang and founder of Alibaba’s data-middle-platform methodology. He has led a team working on this for 5 years, “longer than the term FDE has been popular”; Lingyang’s one-line definition is “an enterprise-grade growth Agent.”

2. What the FDE Premium Buys: AI Sharpness, Industry Depth and Data Breadth

On compensation of $160,000 to $280,000 for OpenAI’s mid-level FDE roles, versus traditional solutions engineers earning roughly 60% more, 朋新宇 rejects a game of labels: “Changing the title can upgrade the salary—this does not conform to business logic or market logic.” The underlying issue is scarcity, with the premium buying time for development.
The line between FDE and traditional operations is that operations only guarantees that a system remains healthy, while FDE must determine whether “today’s operation is moving in a better direction according to my business goals, or getting worse,” then adjust skills, data corpora, context and memory. These tasks either did not exist in traditional delivery or were never encountered there.
He reduces the required capabilities to 3: AI sharpness, industry depth and data breadth. A mature company must either have people who can span all 3 or, like a large customer, rely on an external team for ongoing operations and support.

3. From Delivering Functionality to Delivering Business Outcomes: 3 Principles of Chinese-Style FDE

In response to the objection that FDE is merely on-site development, 朋新宇’s answer is concise: “The fundamental change is that we used to deliver functionality, and now we deliver a business outcome.” The outcome must be verifiable through cost reduction, efficiency gains or expansion into a new business.
The 3 principles are: orient the work around business outcomes, with a business objective, numbers and historical records that can support an evaluation set and metrics; use enterprise data as the foundation, protecting privacy while reaching the upper limit of large-model intelligence; and set job benchmarks as the default, such as a chief customer-service Agent, a top bidder or a top sales Agent—“the starting point may be a bidder or customer-service Agent with 5 or 10 years of experience.” In older language, the goal is customer success management: making the business “better and better.”
The C-to-B divide is straightforward: consumer products need to be “new, novel and unusual,” but can become ordinary after 1 or 2 weeks; enterprises want “more, faster, better and cheaper,” and require a reference point. Individuals may use AI to search for the upper bound and push beyond the edge, while enterprises often want to constrain the edge.

4. The 260-Step Shipment Chase: Customer Service Is About Solving Problems, Not Attitude

The episode’s hardest case involved a 3C electronics customer whose shipment inquiries fell into 3 categories: a customer ordered 2 air conditioners but received only 1, an accessory was missing, or a promised rice cooker had not arrived. The after-sales process runs through more than 260 steps across internal and external systems and platforms including JD.com and Tmall, covering over 95% of cases; only about 5% produce a genuinely new problem.
The customer-service Agent was “better than traditional bots from the moment it went live.” In postmortems, its performance came “close to, and in some cases exceeded, the average performance of humans at their best.” 泓君’s realization is worth preserving: customer service is traditionally understood as a matter of communication and attitude, but this case shows that “the most important thing in customer service is the ability to solve problems”—which requires connecting the full business process and its data.
The methodology for choosing use cases runs through the entire episode: target wherever labor costs are highest, including calls and complaints; money costs are highest, including compensation, loss prevention, ad spending and creative testing; or time costs are highest, including order checks that take hours or days, after which a customer may ask for a refund. Customer-service costs can represent 2% to 3% of sales at some companies and 3% to 5% at others. AI only works when, at the same cost, it delivers a business outcome better than the old process with the same level of performance; when a delay exceeds 1 hour, a customer may ask for a refund.

5. AI Implementation Means Hiring an Employee, Not Buying a System

The core analogy is that “if AI implementation is compared with delivering a system, the comparison is not accurate enough; it is more like hiring a new employee.” The employee arrives with skills equivalent to 5 to 10 years of market experience: “Give it permissions, goals and discretion, and it can go to work.” That is fundamentally different from buying a tool.
The prerequisite is a data foundation. Even a highly capable hire will be lost if “the company’s basic information and knowledge base are a mess.” FDE’s critical task is therefore organizing and governing enterprise data—data for Agent. Lingyang has deployed data platforms for several hundred to more than 1,000 mid-sized and large customers over the past 4 or 5 years, making it easier to build business applications on top of that foundation.

6. The Best Bidder Made 10,000 Adjustments in 3 Months: Modeling Benchmark Employee Behavior

The case involved a company buying traffic across multiple platforms, where the gap between bidders was enormous. Over 3 months, the top bidder made more than 10,000 bid adjustments; an average bidder might make only a little over 100 in 100 days. One bidder made 1 operation a day, while another averaged 100 operations a day. “This difference is very hard to learn by having a master train an apprentice,” which is where Harris Engineering’s self-learning and autonomous evolution come in.
The operations are still physically performed by people, although bidders may use some tools. AI learns from those behaviors. The work is broken into 4 elements: learning, or data collection; modeling; execution; and iteration. “Digitizing, measuring and learning from all the behavior of the previous 3 months—AI will certainly summarize, extract and discover more comprehensively and deeply than people can.”
The environment is generalizable, but the details of each company are not. Agent One covers the full chain from data collection and model building to organizing work skills, task execution, evolutionary iteration and improvement. The major e-commerce platforms have standardized creative, pricing, systems and data, accounting for 80% of the data preparation and system-learning capability needed to train a company’s bidders. That is the basis for FDE to work “on the shoulders of giants.”
The structural point specific to China is that “China does not have an independent SaaS company; its platform companies carry the capabilities of SaaS companies.”

7. Will Universal AI Adoption Erase Competitive Advantage?

泓君 asked whether differences would disappear if AI outperforms people and everyone uses the same tools. 朋新宇 answered in terms of the 3 elements of production: everyone is upgrading the means of production, expanding the boundaries of individual productivity, and AI will eventually create new productive forces. But “time and sequence mean there cannot be an absolute competitive advantage forever; there will always be a window.”
During that window, whoever can better own and operate digital employees may lead. The resource that can extend the lead is still the company’s own core means of production: digitization. “We used to build data platforms so people inside the company could look at data. Today, data systems are for AI. Whether you build them well creates a difference here.”
The final measure of “good” is a concrete business outcome: whether the system generates positive economic value. In the past, delivery ended when the data platform was delivered. Today, delivery of the Agent is only the beginning; the company must put the digital employee to work, and ultimately pay based on business results.

8. Top-Down Projects Work; Bottom-Up Initiatives Hit Production Relations

In one counterexample, a business unit wanted to build an Agent but had no budget, while IT had a budget. Yet “the advanced technology IT wants is not necessarily the outcome the business wants, and the outcome the business wants is not necessarily the most advanced technology today.” The solution is to design a business plan that satisfies IT’s need for technical advancement while delivering the business’s actual outcome.
The positive example was a traditional manufacturing company whose chairman made clear that it “must embrace AI and use AI well today.” Once the effort began top-down in 1 sales region, finance, IT and the sales organization all moved, and the Agent was quickly replicated nationwide.
The hardest transition is ideological. One chairman spent a year training all employees and management internally, personally learning about and explaining AI in more than 15 sessions: “I have to learn it first and then explain it to employees; only then can they understand my thinking and use these systems well.”
Companies will also develop along a “K-line.” An IT team only needs to move a little closer to the business to become an internal FDE, identifying the areas that consume the most money, time and labor. Another group will move downward: “You can oppose a tool or a project today, but you cannot oppose the trend.”
Customer-service and sales employees also have an upward path. Rather than taking calls or handling complaints themselves, they can become the owner of a customer-service AI Agent—handling exceptions, reviewing performance, assigning tasks and teaching the Agent new skills. The goal is “not to handle 100 increasingly difficult problems yourself, but to find a way for the 10 Agents working with you to cover work that previously exceeded your workload and capabilities.”

9. Business Models: Seats or Outcome Sharing; “Betting” Recast as Evaluation

There are 2 broad charging models. The first is seat-based: a company that previously needed 500 people to handle a year of customer service might invest 80% or 90% of the cost of 500 people and achieve the same result. The second is outcome-based: for example, sharing the portion of the conversion rate above the historical human conversion rate. Outcome-based pricing “definitely exists,” and its customers are continuously renewing and receiving ongoing service.
The most candid line in the pricing philosophy is: “Everyone who does business is smarter than I am. Only if they can make the math work will they renew and continue the service.” The point is not to wrap a concept in a premium price.
PUC may now be called MVP. There is no need for extensive conceptual argument: the digital employee can be experienced and used immediately, with minimum viable value tested against the customer’s own business and data units. Once the customer accepts the result, it continues using the product.
Some companies are willing to experiment and even use an outcome bet. “We don’t call it betting; we call it evaluation,” measuring how far the Agent can move the business up the performance curve.

10. The Anti-Palantir Route: Standardization for Personalization, No Person-Day Business

The technical model has 3 terms: customization, standardization and personalization. In the past, companies used customization to satisfy standardized enterprise processes. If the process was not understood deeply enough, only 30% of the problem could be solved; 60% to 70% of new problems would then appear every day, turning the project into a “long and protracted system-delivery exercise.”
Harris Engineering can now standardize the company’s existing workflows, data flows and context, while using autonomous evolution to personalize the result for different customers. “In AI, the future may rely more on standardization to solve personalized problems.”
泓君 drew the comparison with Palantir: she recalled that roughly 46% to 47% of its revenue comes from commercial customers, with engineers entering companies to transform them from zero to one. She sees OpenAI and Microsoft as potentially pursuing a similarly pure-customization enterprise model. 朋新宇’s distinction is that Palantir starts by helping customers identify valuable problems, rather than implementing a listed requirements sheet containing 100 features.
On requests to “buy 1,000 person-days,” 朋新宇 says this kind of customization will become less common: “from the chicken farm to raising chickens, collecting eggs and scrambling them, with every task completed.” The marginal return is low for both customer and service provider. Lingyang focuses on 3 growth areas—sales, customer service and marketing—connecting them one vertical at a time. “We cannot serve every enterprise; we look for the scenarios inside enterprises where these problems actually exist.”

11. The Sales Agent Case: The Least Advanced Province Became the Sales Champion

Offline-channel sales in manufacturing face 3 major pain points. First is route planning: a salesperson may have 20 to 30 customers but can visit only 2 or 3 in a day. Second is the meeting plan: what solution to present, what topics to discuss and how those topics relate to internal tasks and goals. Third is post-meeting entry: when a customer raises a task outside the preset plan, the salesperson still has to record the action in the enterprise system. Salespeople may spend half of each day preparing materials, then another roughly 2 hours at night reviewing and entering data.
What is missing is a sales AI that connects before, during and after the meeting: plan the route, prepare customer materials, write up the meeting afterward, connect with the system and push the information to the people responsible for follow-up.
The company started with a pilot in the province that had previously ranked last in sales. In the first half of this year, that province was named the company’s most advanced sales region and took the national sales championship.
Growth can be divided into 3 areas: pre-sales marketing, including pre-campaign budget allocation, channel analysis and audience discovery; in-campaign creative generation, management and optimization; and post-campaign review, channel selection and optimization. The second is in-sales activity, including new customers, new purchases and referrals—actions such as test drives, tastings and fittings are also data problems. The third is after-sales service, including fulfillment, complaints and long-term repurchase.
The product takes a “4+X” form: 4 preset Agents for marketing, sales, customer service and operations, each representing a job benchmark with 5 to 10 years of skills in the relevant industry. X is the company’s data assets and context management, including data permissions, fund approvals, advertising permissions, customer-service inbound permissions and the ceiling on compensation payments.

12. More Than a Year of Deployment, Then One Week of Time-Sliced Go-Live

朋新宇 is candid that the work takes more than a year despite sounding simple in 2 sentences. Familiarizing the team with the job accounts for more than one-third of the time; modeling comes next, followed by roughly another one-third spent tuning the AI mentor group. Actual implementation, execution and system launch are relatively fast. The core difficulty is changing the organization and production relations, making a shift in mindset essential.
Go-live is not a conventional traffic canary but a time-window cutover. On day 1, the Agent handles 0:00 to 2:00, when demand is low and the impact is controllable. After review and adjustment, day 2 shifts the full peak flow from 12:00 noon to 2:00 p.m. “Only when the system runs at full volume do we know the boundaries of the exceptions”: which glitches will make customers complain or lose their temper, and where the Agent performs better than people. On day 3, the window expands to 8 hours; within a week, the Agent meets the company’s chosen benchmark.
His understanding of the role has changed: “We used to treat it as a tool. Now I genuinely treat it as a companion.” Once trained, it can fully take over the work when people are absent or during overnight hours, rather than merely executing a single instruction.
What persuades a chairman is not a pitch but an MVP. After the team used AI tools to build a prototype quickly, the chairman said: “Before I saw it, I didn’t know what I wanted. But this is exactly what I wanted.”

13. FDE Is Not One Person but an Organization: BA, AI Architect and Coaching Team

“It is extremely difficult for one person to have all these capabilities.” Lingyang divides the work into 3 roles: the BA, who is oriented around business outcomes and understands industries including beauty, appliances, home furnishings, pets and automobiles, can orchestrate more than 200 customer-service steps and set scientific evaluation standards; the AI architect, who translates business problems into model size, points for human intervention and points for machine takeover while keeping up with new models and Agents; and a business-expert coaching group made up of chief customer-service and top-sales personnel.
During tuning, Lingyang steps back and takes a lighter role, selecting the best customer-service and sales staff from the customer to form an AI expert or AI coaching group.
The long-term view is that large enterprises, and companies with strong platform, IT and technical capabilities, should eventually have these roles in-house. For now, they can use Lingyang to quickly understand the best level in their industry and avoid unnecessary detours.

14. Token Bills and Model Selection: Enterprises Do Not Need to Worry About the CPU

Lingyang has seen this firsthand. At the end of the 1st quarter, one AI project generated RMB3M of revenue against RMB12M of investment; by the end of this quarter, it was close to breakeven. That early investment is why Lingyang designs Agents so customers do not become overly sensitive to Token usage but focus on what problem was solved. “Can the math work? If it works, most bosses do not care which Tokens are behind it.”
There is no model dogma. Lingyang uses the models available on Alibaba Cloud, including Alibaba models and open-source models such as Kimi and 智谱, selecting the best cost-performance ratio for each scenario. As with buying a computer, customers do not necessarily need to debate the CPU, chip and memory; they mainly care about the final result.
Private deployment involves 2 separate calculations. The “building blocks” are now strong enough that nobody talks about post-training anymore. Large Alibaba models and models such as Kimi have parameter footprints above 2TB, and private deployment may require an initial investment of several tens of millions of yuan.
Data anxiety must also be separated into the system layer and the model layer. When an Agent calls a large model, the large model does not absorb the enterprise’s data. Data may instead be collected by Agents used for programming or web coding. Consumer products such as ChatGPT may collect conversations to optimize the product, and users can turn off the “use for training” setting. Lingyang keeps the data inside the enterprise’s own Agent; the system can be deployed internally, making it observable and controllable.
Model upgrades do not trigger automatic switching. An Agent remains stably bound to a specified model, and any upgrade requires human confirmation. Only after testing shows that a model can generate gains will an upgrade plan be made. Enterprises are also bound to processes, permissions and management boundaries; what they actually want is stable work output.

15. Enterprise Leaders’ Attitudes Have Changed Sharply, and China-US FDE Is Not Directly Comparable

The biggest change this year is that “people used to ask how to use it and whether it could be used; now they skip that phase and ask when they can put it to work.” 朋新宇 sees the inflection point after Lunar New Year, with the rise of “Lobster,” followed by Harris Engineering and current trends such as 路普. Attention has moved from model parameters and benchmark scores to whether AI can execute tasks. “Entrepreneurs are the smartest and most sensitive group of people.”
The China-US structural gap has 2 parts. The US has Salesforce and SAP to establish workflows and data standards; China is relatively short of that foundation, so “everything has to be done, from driving the first piles and building the structure to the finishing work.” Payment habits also differ: large Chinese enterprises generally have their own IT teams, so a pure delivery model runs into both talent scarcity and market saturation.
In the US, FDEs may come from traditional SaaS platforms such as Salesforce and SAP or from internet platforms such as Facebook and Google. In China, the role requires greater customer-side orientation and coordination across internet platforms including Taobao, Tmall and WeCom. Pure function integration and delivery easily become a price-comparison business; people from internet platforms understand traffic, internet operations and platform coordination, giving them a natural advantage.
朋新宇 declines to say whether China and the US are ahead, synchronized or behind: “I don’t have the data, so I cannot compare whether we are better or worse than others.” Customers will ultimately express their judgment through orders and renewals. For now, “so far so good.”

16. People, Work and the Game: What Skills Should AI-Era Workers Build?

As CEO, 朋新宇 focuses on 3 things: people, work and the game. People means helping customers and employees grow faster and better through AI; internally, the company holds a monthly employee forum on AI best practices, while on the customer side it holds a “data peers” session monthly or annually. Work means finding tasks with momentum, so customer-service, marketing and sales Agents can grow organically inside customers and spread through word of mouth. The game includes the evolving landscape of AI, technology, engineering, models and business architecture.
Hiring requires a combination of 3 capabilities: technology, business and data. “The difficulty of moving from technology toward the business side may be lower,” and business specialists do not need to be generalists. Data fluency is not about looking at dashboards every day; it requires 3 basic skills: year-on-year analysis, such as whether an RMB1.8M increase is good or bad; comparison, with oneself, competitors, another brand or another product; and share, such as whether RMB1.8M represents one ten-thousandth or one-tenth of the market. Data provides the reference point and target.
The “data scientist” title eventually dropped the “scientist” suffix to emphasize scientific methods rather than a person’s position. Industry experience remains a barrier in To B, although capabilities that once took 10 years to learn may now take 1 or 2 years. Young people should grow by standing on others’ shoulders rather than starting from zero.
Enterprises have 2 paths. “Add AI” adds AI to existing systems for data analysis, data governance and risk control, potentially shifting software from being used by people to being used by AI and reallocating value. “AI-plus” redefines work in customer service, marketing, shopping guidance and ad bidding. At the core of the enterprise, the areas consuming the most labor, money and time are the easiest places to explore “AI-plus.”

Everyone Is Still Undersizing the AI Market | Eric Vishria AI Is Learning at the Wrong Level of Abstraction — Matthieu Wyart