89. NetEase’s 阮良 on AI-Native Organizations and Enterprise Agents
Summary
The real AI divide is not whether you use it, but how well you use it. Individual productivity gains still have to pass through processes and convert into organizational throughput. At NetEase, top engineers have seen 3–5x productivity gains from AI coding, with some reaching 10x; product managers are about 50% faster. Yet the time from submitting a software request to launch has fallen only from 20 working days to 17. 阮良’s conclusion: “Process and organization are still the short plank in this barrel.”
Enterprise AI has to clear three successive hurdles—awareness, individual adoption, and process redesign—with the third being the hardest. Lengthening one plank does not increase the barrel’s capacity. A true AI-native transformation removes unnecessary steps and may even require “changing the barrel.” That is why small companies with simpler organizations and hands-on founders often transform faster than better-funded enterprises.
Agents are not the end of core SaaS; they may be demand multipliers. Critical SaaS products are already embedded in enterprise knowledge bases, CRM, approvals, security and other workflows. Their future users will shift from humans with limited time and energy to Agents that are more numerous and act more frequently. 阮良’s view: “Truly good or core SaaS will only get better.” Ordinary, non-critical software may be eliminated.
The core asset in enterprise AI is not a general-purpose model, but a high-quality knowledge base, a record of historical decisions and callable business processes. NetEase turned past decisions and their rationales into Agents and Skills, then had them reassess hundreds of old cases, achieving accuracy above 90%. The end goal is not to let AI decide autonomously, but to write human corrections back into the system and create a feedback loop in which “people stand on AI’s shoulders, and AI stands on its own shoulders.”
阮良 expects the ranks of middle managers carrying out execution and coordination duties to shrink, while frontline elites equipped with AI armies become more common. Teams of 2–3 or 3–4 people will be able to direct large numbers of Agents and draw on company platforms and resources. Frontend, backend and testing roles are also beginning to recombine into full-stack jobs. High-end talent may eventually negotiate not only individual compensation, but also on the basis that “I’m not alone—I bring an AI team.”
Token costs should be judged by output value, not by the absolute size of the bill. NetEase observed that heavy Token consumption is probably correlated with high performance. In January 2026, one month of Token usage could equal the entire previous year’s consumption, yet a programmer’s AI bill of several thousand yuan per month remains cheap relative to the productivity gains. The internal rule of thumb: “If your Token consumption is leading by a cliff, your work quality and results are probably leading by a cliff too.”
AI replaces work steps, not the person or their creativity, courage and curiosity. 卫诗婕 asked where people would build judgment and conviction if AI took over basic labor. 阮良 acknowledged that old technical work will disappear, but believes new technical work will continue to provide a training ground. Companies that treat AI only as a layoffs tool are “fundamentally thinking about it the wrong way.”
Investors assessing enterprise AI realization cannot look only at employee adoption; they need to track end-to-end cycle times, knowledge capture, decision feedback and organizational redesign. The best tasks to automate first are time-consuming, frequent, unavoidable and likely to produce similar results across employees. 阮良 recommends validating feasibility with the strongest and most expensive model first, then cutting costs, because “once the most expensive AI can do it, the cheaper AI will definitely be able to do it within two months.”
Deep dive
1. AI’s Potential Is Far Greater Than Mobile Internet Because It Can Understand and Execute
阮良 has lived through 3 waves—Internet, mobile Internet and foundation models—and is more excited in his 40s than he was in his early 30s. He says he still feels “like a young person in his 20s, fresh out of college,” with his curiosity, fighting spirit and ideals restored.
His distinction between the 2 technology waves is clear. Mobile Internet produced point innovations through sensors such as location, cameras and gyroscopes; AI not only receives data, but understands it, identifies intent and executes, allowing it to enter the “capillaries” of personal life and organizational workflows.
This is not another interface or endpoint. It is the first time software has gained the ability to generalize across contexts. What 阮良 sees is not a handful of new applications, but a geometric expansion of what is possible.
2. Personalized Education Shows AI’s Most Direct Public Benefit
Education has long established that one-to-one teaching tailored to a student’s stage and interests works better. It was never broadly available because the cost was too high. AI makes it possible for every child to receive learning content calibrated to their ability, interests and even preferred game formats.
He contrasts this with the arcade quiz games of his youth, which relied on fixed question banks. Old software could only execute preset patterns; AI can continuously generate personalized, entertaining content and deliver frequent positive feedback. Even people who cannot code can now build simple web applications.
The strongest example he heard involved a Silicon Valley engineer who used Claude to scan textbooks and past exam papers, build a study plan and generate repeated practice questions, earning a difficult accounting or tax credential in under 2 months. 阮良 compares this form of knowledge equalization with the printing press weakening the clergy’s monopoly over copying and interpreting the Bible.
3. A CEO Who Left Coding for a Decade Was Pulled Back to the Front Line by AI
阮良 wrote code for 7–8 years after graduation, then moved into team, product and company management and went nearly 10 years without writing a line. He had concluded that he no longer knew how to program and had no time to catch up with changing languages and tools.
The turning point came through an investor friend who had also left programming more than 10 years earlier. The friend used AI to write tens of thousands of lines of code—more than he had written in total during 5 years as a programmer. 阮良 realized that the “software-defined world” was opening up again to people outside the engineering front line.
He then used Claude Code to build a dedicated CEO system. The first functions were not grand strategic decisions, but approvals, information aggregation and contextual search—the persistent drains on managerial attention that make “focus on the big and let go of the small” impossible.
4. The First Value of a CEO Agent Is Reclaiming Attention from Trivialities
阮良 divides approvals into 3 categories. The Agent handles large volumes of low-risk items automatically, presents items requiring judgment in a concentrated view, and for the highest-risk approvals traces the relevant background information and asks the submitter for more justification when the materials are incomplete.
The value is not simply opening fewer OA, email and enterprise systems. A CEO’s job is often to identify opportunities and risks in a web of relationships, while human cognitive capacity, energy and context are limited. AI can extend an approval outward to related matters.
He treats time saved as the recovery of organizational opportunity cost, while emphasizing that rest is itself productive: “Watch a short drama, look at the scenery for a while, or have a cup of coffee” is not wasted time. A person’s state can matter more than their IQ to the quality of the next task.
Agents beat Chat through compounding. Fixed information does not need to be requested from scratch each time; the system continuously captures, organizes and executes. “You don’t need to go back to Chat.”
5. Individual Productivity Rose 3–5x, but Organizational Delivery Improved by Only 3 Days
When GPT-4 appeared, NetEase briefly worried that its intelligent customer-service business would be directly displaced by general-purpose models. Practice showed that models still require extensive tuning, business knowledge and process integration. Individual AI and enterprise AI are different problems.
Individual tasks often close over a short loop. Enterprise tasks involve multiple people and long chains across product, engineering, testing, supply chain and other functions. If even one node in the chain remains unchanged, spectacular local gains are absorbed by the overall process.
NetEase previously took about 20 working days on average to move a software request from product-manager intake to launch. AI coding made top engineers 3–5x more productive, with some reaching 10x, and turned frontend and backend staff into full-stack engineers. The end-to-end cycle nevertheless fell only to 17 days.
Product managers became about 50% faster at writing PRDs and prototypes. The unchanged “plank,” however, may not have been a person at all, but approvals, handoffs or other process steps. 阮良 therefore identifies 3 hurdles: organizational awareness, short-cycle individual applications, and process and organizational redesign.
6. AI-Native Does Not Lengthen Every Plank; It Removes the Planks
阮良’s correction is blunt: “Some of the planks in this barrel are unnecessary. Just pull them out.” If every role interface and management action from the pre-AI era remains in place, automation merely lets the old organization generate queues faster.
NetEase Zhiyun started with the weekly meeting. The meeting lasted about 2.5 hours and was expected to synchronize information, identify problems or opportunities, and make decisions. But the time spent by departments preparing materials at each level could be 10 times the meeting itself.
Information synchronization was unreliable in its own right. Irrelevant updates caused people to zone out, while the one sentence that mattered to them was easy to miss. Even a general manager cannot stay “that focused every second” for 2.5 hours—only a deity could.
7. The Weekly Meeting Was Split into AI Reports and Human Decision Sessions
The new process began with the general manager and management team defining the information they needed. Departments then submitted materials using a template, and an Agent scanned everything before the meeting to produce a concise, focused report capable of surfacing cross-department connections.
The report appears as interactive HTML, allowing comments and follow-up questions. AI also lists potential problems and opportunities in advance for people to confirm and supplement. Once synchronization and discovery are complete, the standing weekly meeting disappears; teams convene only for focused sessions on problems that actually need to be solved.
卫诗婕 cited ByteDance’s pre-meeting document system and asked whether mature organizations could already improve meeting efficiency, with AI merely exposing old problems. 阮良 acknowledged the importance of organizational foundations, but warned that document culture itself can evolve into information overload.
8. The Easier It Is to Produce Documents, the More Attention Becomes the Bottleneck
NetEase used Feishu for about 1.5 years starting in 2020, particularly valuing the practice of spelling out problems before meetings, commenting in documents in advance, and discussing only disagreements in the meeting. During the pandemic, Feishu documents also became an important tool for distributed collaboration.
Once the tools became too easy to use, documents multiplied and grew longer. Transparent collaboration began sliding into information explosion. AI is now accelerating information production again; without a filtering mechanism, an organization merely shifts the old bottleneck to reading and judgment.
阮良 also acknowledges the individual side effect. Seeing a long document now triggers the instinct to “ask AI to summarize it,” which can produce intellectual laziness. When the team once encountered an abnormality with its Claude account, people even felt “the sky had fallen and we couldn’t work,” as though a crutch had been taken away.
9. SaaS Will Not Die as a Whole; Agents Will Amplify Critical Services
阮良’s response to “SaaS is dead” is that the essence of SaaS is Service. The form can evolve from Software to AI as a Service or Agent as a Service, but enterprise demand for service will not disappear.
He draws a boundary. Ordinary SaaS that has not entered a critical workflow may genuinely die; products that have become enterprise infrastructure may gain more frequent and deeper usage through AI.
In the past, software users were humans constrained by time, energy and context. In the future, users may be Agents that far outnumber people and act at much higher frequency. API calls to databases, security systems, audit tools, human resources and CRM could all increase dramatically.
NetEase Zhiyun therefore does not need to abandon its existing To B business. It needs to use AI to rebuild the parts of that business that represent truly irreplaceable Service.
10. Private Knowledge and Process Orchestration Are the Fuel for Enterprise AI
Take intelligent customer service. A general-purpose model cannot deliver high-quality service on its own because it lacks product knowledge, customer profiles and the company’s operating process for resolving issues. The private knowledge base where those elements are accumulated is the critical asset.
The stronger AI becomes, the more frequently it can call a high-quality knowledge base. 阮良 says AI has “given the knowledge base a pair of wings.” The other genuinely important capability is calling and orchestrating internal information systems, workflows and APIs.
This is why he remains bullish on core CRM products and Salesforce-type platforms. Their value lies not only in the interface, but in their position inside business workflows, their accumulated structured knowledge and their ability to be repeatedly called by a new Agent layer.
11. The Real Source of a Skill Is Excellent Behavior and Real Customers
阮良 summarizes the evolution of enterprise capability this way: it first resides in people, then becomes software, and is now being “distilled” by AI into Skills. Sales is the clearest example. A boss worries both that the team has only one star salesperson and that the star may leave.
Companies can extract highly relevant actions from a top salesperson—for example, what to do after a customer visit to shorten the sales cycle or raise the probability of closing. They can also extract how that person responds to competitive challenges without mechanically copying an entire script.
To train ordinary salespeople, the thing worth distilling is not the star but the customer. Historical calls, WeCom messages, emails and other interactions can generate simulated customers of different types, allowing new hires to repeatedly “do practice questions” and compensating for the fact that mentors may not be willing to share everything they know.
Online sales are naturally on record. For offline key-account sales, the salesperson can record a 2–3 minute voice memo after a visit. Salespeople often resist CRM because they see it as merely “managing people,” but are willing to leave recordings that help them review their own performance and close the next deal.
12. Total Surveillance Is Not the Answer; the Customer’s Voice Is More Valuable Than Employee Monitoring
阮良 agrees that “the boss often gets the least information.” By the time information reaches the CEO, it has usually been packaged and stripped down at multiple levels. NetEase therefore requires meetings to retain source materials, while offline meetings use recording cards or enable online recording.
But when employees—especially high performers—resist being “distilled,” he favors accommodating human nature rather than forcing everyone to send data back. NetEase is more interested in Voice of Customer, returning customers’ real voices to product and engineering than in tracking every step employees take.
His priority is clear: “Focus on customers, focus on the market,” and only then focus on competitors if necessary. If the direction is wrong, no level of organizational efficiency matters. Existing online data and voluntary records are often sufficient; compulsory full capture only provokes resistance.
13. The AI Era Requires a New Debate over Ownership of Experience and Talent Premiums
卫诗婕 raised a sharp objection. The premium an employee receives after promotion reflects years of knowledge and experience. If the company takes that capability in full and copies it for others without paying extra, “it feels a little unfair.”
阮良 did not deny the tension, but expects the pricing of talent to change. A future candidate may say: “I’m not alone; I bring a team.” The team would not consist of people, but of an AI army formed from the candidate’s accumulated experience and Skills.
The premium for high-end talent may therefore not fall, but expand. The company would be buying more than one employee’s working hours; it would also be buying the individual AI organization that person can direct and the capabilities it has accumulated.
14. Fewer Middle Managers Does Not Mean Fewer Elites; Organizations Will Return to Small Squads and Full-Stack Roles
阮良 expects the number of middle managers responsible for execution and coordination to fall. Organizations will look more like amoebas or special-forces units: teams of 2–3 or 3–4 people directing large numbers of Agents and drawing on company platforms and resources to complete complex work.
He draws a distinction between “execution and coordination duties” and whether people themselves are eliminated. Middle-management positions will shrink, but their occupants may not be fired; they may return to the front line as deeper-skilled, higher-leverage elites.
Software teams are already moving in this direction. As requirements and system scale grew, the industry created separate frontend, backend, algorithm and testing roles. AI is now broadening everyone’s horizontal capabilities and bringing teams closer to the full-stack era when 张小龙 wrote Foxmail independently and 邱伯军 built large software systems on his own.
阮良 describes the structure as a beehive. Many “worker bees” are highly capable in their own right, with deep vertical expertise and AI filling in their horizontal abilities. What will truly be scarce is the depth of T-shaped talent.
15. AI Replaces Steps, Not Innovation, Courage or Responsibility
阮良 has seen many companies over the past 2 years try to use AI to replace people directly, and act on that intention. His judgment is clear: “They are fundamentally thinking about it the wrong way.” AI is good at executing end-stage tasks, but cannot fully replace people in entering deep waters and driving product and organizational evolution.
An excellent customer-service representative does more than answer questions. They identify recurring product defects, opportunities and risks. AI can receive inquiries, aggregate information and analyze it, but people are still needed to combine resources, propose new solutions and take responsibility for pushing them forward.
卫诗婕’s counterquestion is worth preserving: people’s convictions, judgment and courage often come from doing tedious work early in their careers. If beginners no longer perform basic tasks themselves, will the future still contain enough “courage content”?
阮良 is not pessimistic. He believes new work will emerge after old technical work disappears, giving young people another chance to train, accumulate understanding and develop an appetite for risk. The Industrial Revolution should first have amplified creativity, he argues, with new demand then driving a more powerful surge in productivity.
16. AI ROI Can Be Reduced to Quality, Time and Unit Economic Value
NetEase’s framework is that the value of an AI task equals the ratio of AI output quality to human output quality, multiplied by the time saved. Multiplying that by the company’s economic value per unit of time converts it into cash-based ROI.
阮良 explains with coffee. If human quality is 8 points and AI quality is 7, the quality coefficient is 7/8. If the human takes 15 minutes and AI takes only 10 seconds, the remaining 14 minutes and 50 seconds represent the value of time saved.
The tasks that should go to AI first have 4 characteristics: they are time-consuming, frequent, unavoidable and likely to produce similar final output regardless of the employee’s education or intelligence. Their value is easiest to calculate and they are best suited to rapid decisions.
His implementation rule is to use the strongest and most expensive model first to validate capability and quality, then find a more economical model to institutionalize the workflow. Establish the quality bar and capability ceiling first; saving Tokens too early may prevent the team from proving feasibility at all.
17. Heavy Token Consumption May Be a Productivity Signal, Not Evidence of Waste
At the beginning of 2026, one month of NetEase’s Token consumption could equal the entire previous year. 阮良 still does not consider the bill expensive: the company’s existing personnel costs and task values are high, and a programmer’s monthly AI spend of several thousand yuan is cheap relative to the productivity gain.
Non-R&D roles may also produce more value with AI. Before visiting a customer, a To B salesperson no longer prepares only a PPT, but may build interactive HTML and a customer-specific demo. HTML and Markdown are easier to distribute and better suited than PPT for entering an AI knowledge base.
NetEase has used a points system since 2024, converting work value through measures such as effective code volume. In the AI era, the team has observed that people whose Token consumption is “ahead by a cliff” are often building systems for themselves, reducing execution time and creating more time to think. Their output leads accordingly.
阮良 also argues that truly excellent people do not waste Tokens. For standardized, long-cycle tasks, NetEase builds task selectors that break work into subtasks and automatically match each one with a model that can complete it at a relatively economical cost.
18. Multi-Agent Decisions Need Competition, Redundancy and a Final Human Call
阮良 subscribes to most major models on the market, including Claude, GPT, Grok, Zhipu, Kimi and DeepSeek. He assigns different models to different Agents and gives them names such as “诸葛亮.” This heavy setup is reserved for major decisions, institutional design and proposal development.
The workflow is to have each Agent develop a solution independently, then pit them against one another like red and blue teams. A central coordinator synthesizes the results, and 阮良 still makes the final call. It is “a little more expensive,” but he believes the incremental value for important decisions far exceeds the model cost.
Multiple instances also provide resilience. Early on, OpenCloud was unstable; if all 6 sub-instances depended on the same entry point, one outage could disable them all. Distributed across 6 Agents, the system would still have 3 available even if half went down.
Each Agent has a different Harness and produces different results. 阮良 compares the redundancy to retaining spare talent in an organization and designing software for robustness: “You have to make sure it stays alive.”
19. Small Companies Realize AI More Easily; Large Companies Must Break the Old Workflow
In a small company, the boss usually adopts AI first. Once the value is confirmed, it can be pushed to sales and design. Fewer layers, shorter chains and simpler relationships let awareness and resolve translate quickly into action.
Large companies often encounter AI earlier, but do not necessarily unlock its full value faster. Existing socialized divisions of labor turn people into “screws”; even if AI improves one step, it does not automatically remove the handoffs and waiting between departments.
阮良 uses the “turning rice paddies into mulberry fields” policy in The Ming Dynasty in 1566 as an analogy. The policy logic looked complete, but implementation became entangled with food allocation, grassroots survival and interests at every level. “Once you actually execute it, it becomes a pot of porridge.”
His reform path is to choose a small department most willing to change, build a closed loop, win a small victory, spread it widely and then expand the coalition. Even if the scope is limited, it should ideally be a top-leader project: the boss must use AI personally and enter the arena.
20. The Most Striking Enterprise Case Compressed the Long Chain from Inspiration to Production
A toy company’s old process began when the boss saw a new artistic symbol while traveling and brought back a photo. He explained it to design, design worked with the supply chain to produce a draft, and the company then made molds, produced samples and tested sales. A long series of handoffs separated the discovery of an opportunity from a product launch.
The NetEase team demonstrated a different path. AI moved directly from the concept to a prototype design, advanced to a CAD drawing, and simulated the manufacturing cycle and potential suppliers, compressing the previously fragmented chain.
The boss immediately called the heads of design, manufacturing and HR to the screen: “Take a good look at what this AI can do now.” What 阮良 remembered was not the demo itself, but the shock and excitement of a traditional-industry boss seeing for the first time that an entire workflow could be rewritten.
21. An AI-Native Organization Needs Talent Standards, Safety Guardrails and an Evolvable Decision Core
If he were building an organization from day 1, 阮良 would first establish “faith in AI,” then evaluate talent using NetEase’s updated standards: taste, standards, structure, judgment, insight and self-directed learning. Those qualities determine whether a small number of people can direct Agents to build genuinely good products.
Enterprise Agents must also be controlled. The probabilistic nature of Transformers creates unpredictable outputs. NetEase’s security products put guardrails around AIGC content and limit Agent permissions and actions, preventing production risks such as unauthorized access and accidental data deletion. Large enterprises have an especially strong need for this.
The foundation of a decision core is not simply adding another model, but retaining why a decision was made. NetEase puts decisions and their rationales into a knowledge base, generates Agents and Skills, and backtests them against hundreds of cases, achieving accuracy above 90%.
Videos related to Bridgewater gave 阮良 the idea of feeding human final corrections back into the Agent so the system keeps learning instead of freezing old experience. AI will still make mistakes; people will continue to judge, decide and execute. But every confirmed adjustment becomes capability for the next round: “People stand on AI’s shoulders, and AI stands on its own shoulders.”