Individual productivity tripled, but company delivery only went from 20 days to 17
AI armed everyone, but the organization didn't get faster — because the weak link in the barrel isn't people, it's process. The real dividing line is whether you can transform process and organization, not whether you hand out tokens to employees.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Individual productivity tripled, delivery only sped up three days
NetEase ran the numbers last year: a software requirement, from the product manager receiving it and producing a design doc to the feature going live, averaged 20 working days. After AI coding crossed the inflection point, many engineers doubled or even tripled to quintupled their personal efficiency, some tenfold. But for features of the same scope, the whole process only shrank from 20 days to 17. The reason is that a team completing a complex task is like a barrel: of 20 staves, 10 are engineers, 3 are product, 3 are testing. The engineer and product manager staves shot up fast, but some staves aren't people — they're process — and they didn't improve at all. So AI adoption has to clear three hurdles: cognition, landing on individual short-range tasks, and transforming process and organization. The third is the hardest.
— Ruan LiangCancel the weekly meeting: AI syncs information, humans only solve problems
Ruan Liang asked friends who run companies whether they still hold weekly meetings; 99% said yes. Weekly meetings usually serve three purposes: sync information, discover problems and opportunities, and make decisions. It used to be 2.5 hours a week, but the time spent on that meeting could be ten times 2.5 hours — every department prepares materials layer by layer, and the information sync itself isn't done well, because no one listens attentively the whole time and may miss a sentence relevant to them. After the reform: the general manager and management agree on what information is needed, everyone fills in a template and stores it in the knowledge base, AI scans and integrates it into a focused, interactive HTML report where people can leave comments and follow-up questions; an AI agent produces a preliminary list of problems and opportunities, which human brains confirm and supplement; problem-solving and decision-making become dedicated meetings.
— Ruan LiangSaaS isn't dead, but ordinary SaaS will die
Ruan Liang doesn't think SaaS is dead: SaaS is essentially Service, and Service always exists, it just changes form — AI as a Service and Agent as a Service are themselves software. But ordinary SaaS may really die, while truly good SaaS that has already entered a company's critical processes will only get better, because it becomes internal infrastructure. In the past, humans used software, limited by time and energy; in the future, a large number of agents will call databases, APIs, security APIs, audit APIs, HR APIs, CRM APIs — call volume will far exceed the number and actions of humans, it will grow enormously, and the more it's called the more entrenched it becomes. The premise is that this Service is critical enough inside the company.
— Ruan LiangDistill top salespeople into skills and distribute them to everyone
Bosses have two headaches: the sales team has only one or two strong performers and the rest are mediocre; or a salesperson is too strong and what if they leave. There used to be no solution, but now you can distill or extract the actions, scripts, and behaviors of good salespeople into skills. Two uses: distribute them to other salespeople to use directly; and have AI turn elite salespeople into teachers to train ordinary salespeople. What training needs to distill isn't the salesperson, it's the customer — take the customer interaction data the salesperson has had, from calls, WeChat, email, etc., and have AI simulate various customers to train the salesperson. Online sales is purely digital and on record; for offline sales, recommend or even require a recording card — after a visit, record a two-to-three-minute voice memo while memory is fresh, because salespeople dislike CRM (they think CRM is for managing them), but they don't mind a recording card, because it's also useful for their own customer review.
— Ruan LiangFewer middle managers, strong individuals lead agent legions back to the front line
Ruan Liang judges that middle management will become fewer and fewer, but not necessarily eliminated — rather, they'll go to the front line, and the front line can also lead a bunch of agents. Middle managers with management responsibilities will become fewer and fewer, while elite individuals with strong skills will become more and more. He gives the example of an engineering team: previously, socialized division of labor split software development ever finer — frontend, backend, algorithm, test engineers; in the last half year it has clearly reverted to full-stack, everyone becoming full-end engineers. In the future, organizations will look more and more like a distributed structure, with countless highly capable engineers, a mother peak commanding some engineers, and countless agents fighting a battle. He imagines a future interview negotiation scenario: the candidate says I'm not one person, I bring an AI team, a summary of all my past experience and skills, and the premium the company has to pay might be some kind of OPC.
— Ruan LiangAI value formula: quality times time saved
Ruan Liang gives a formula referencing a paper: AI value = quality of AI completing the task (as a ratio of human quality) × time saved × the company's economic value per unit time. For example, a human makes a cup of coffee at eight points, AI also makes it at eight points, so the quality coefficient is 1; a human takes 20 minutes, AI takes only 10 seconds, saving 14 minutes 50 seconds; multiply by the company's economic value per unit time, and you can calculate ROI. Prioritize four types of tasks for AI: time-consuming, high-frequency, unavoidable, and tasks where different people produce roughly the same result. Once calculated, you can immediately decide whether to use AI. He also suggests first using the most expensive AI (Claude, GPT) to prove the task can be done, solidify the process, then switch to cheaper ones — if the most expensive can do it, within two months the cheaper ones can too.
— Ruan LiangToken consumption leads by a landslide, output also leads by a landslide
NetEase started a points system in 2024, trying to quantify engineers' work tasks into points with work value. By the AI era, they found a group of people whose token consumption was astonishing, and whose work output and work value were also astonishing, forming a statistical pattern: people whose token consumption leads by a landslide are most likely building systems for themselves, saving their own work time and increasing thinking time, so their work quality and work results also lead by a landslide. Many bosses worry about AI abuse causing waste; Ruan Liang's hot take is that truly excellent or elite people won't waste — the company has audits but doesn't tell colleagues, and through statistics they found elite colleagues don't waste. He also says small companies are more likely to fully embrace AI, because there are fewer organizational relationships, the boss uses it first, there are fewer layers, and the chain is shorter.
— Ruan LiangThe core of a decision hub is decision history, not the model
Building an AI decision hub essentially requires a high-quality knowledge base, and what really matters is the historical decision processes and decision rationales. This year NetEase preserved historical decisions, incorporated them into the knowledge base, combined the two types of knowledge to train an agent or skill, then fed back the original hundreds of cases for the agent to judge, achieving over 90% accuracy. Ruan Liang also mentions that Bridgewater gave researchers and analysts agents they built themselves, precipitating decision processes and decision history into a decision knowledge base; what Bridgewater did better was making the agent a learnable, evolvable part — after a human adjusts it, the confirmed adjustment actions enter the next round of reinforcement, keeping pace with the times, essentially human-machine collaborative self-evolution.
— Ruan LiangIn their own words · checked verbatim
Across our whole process, for software features of similar scope, from product requirement to launch, it only shortened from 20 days to 17.
我们整个流程下来 同样的类似力度的软件功能 从产品需求开始到上千 从20天只缩短到了17天
Ruan Liang17:07
The third and hardest hurdle to cross is actually transforming the process, your organization. Right — once you transform the process, it inevitably means transforming the organization.
第三个也是最难越过的一个坎 其实是要改造流程 你的组织 对 一旦改造了流程 就势必意味着要改造组织
Ruan Liang19:09
A two-and-a-half-hour meeting — it's impossible to be that focused every second. That would take a god.
一个两个半小时的会 不可能每一秒都那么专注 那是神仙
Ruan Liang22:12
Middle managers with management responsibilities will become fewer and fewer, but elite individuals with strong skills should become more and more.
实行管理职责的中层会越来越少 但是拥有强悍技能的精英人士 应该会越来越多
Ruan Liang45:33
If your token consumption leads by a landslide, your work quality and work results are also most likely leading by a landslide.
你token如果消耗断压式领先 你的工作质量工作成果也大概率是断压式领先
Ruan Liang1:00:41
Sometimes it's not that the boss wants to push and can push — because many organizations have ten thousand reasons to make sure this thing doesn't get done.
有时候不是说老板想推动 就能推动 因为很多组织 他有一万种理由 让这个事情办不成
Ruan Liang1:04:43
It made the agent into a learnable, evolvable part.
它把agent做成了一个 可学习 可进化的一个部分
Ruan Liang1:12:51
Figures
| Average software requirement delivery cycle (before AI) | About 20 working days | 16:07 |
| Average software requirement delivery cycle (after AI) | 17 days | 17:07 |
| Engineer personal AI coding efficiency gain | Doubled to tripled to quintupled, some tenfold | 17:07 |
| Weekly meeting duration | About 2.5 hours per week | 20:11 |
| NetEase Group token consumption | This year, January and February or one month in Q1 equals last year's entire year | 53:37 |
| Programmer monthly AI cost | Several thousand RMB | 54:37 |
| Decision agent accuracy | Over 90% | 1:11:50 |
Glossary
- Vibe Coding
- Using natural language to have AI generate code, so people who don't know programming can build applications.
- Agent
- An AI program that can autonomously call tools and execute multi-step tasks.
- Skills
- Distilling and extracting human capabilities into reusable AI skill modules.
- OPC
- One Person Company, where one person leads an AI legion to complete company-level work.
- Robustness
- The ability of a system to continue operating normally when some components fail.
How to listen
CEOs, CTOs, and HR leaders at companies pushing AI adoption, especially managers trying to figure out why individual productivity went up but the organization didn't get faster.
The first 4 minutes of education-scenario setup can be fast-forwarded; go straight to the barrel weak-link case at 16:07.