AI Has Entered the Building. Your Company Is Still Running on Steam Power.

In 1882, Edison built a power station on Pearl Street in New York and lit 400 electric lamps. To promote it, he gave away six months of electricity. Today, electricity is basic infrastructure. At the time, even Edison could not imagine how it would change the world.

The electricity story has a second half.

For the next decade, electricity entered factories, but productivity did not take off. Early factories replaced the steam engine with one large electric motor. The central drive shaft stayed. Belts and gears stayed. Power still traveled through the structure left by the steam age.

Only when small motors could sit beside individual machines did factories escape the central shaft. Machines no longer had to line up around one shaft. Equipment layout, logistics, process, and factory design could change. The Second Industrial Revolution began.

AI Coding today is that electric lamp.

The model is inside the company. Engineers write code with AI. Sales finds customers with it. HR filters information with it. But approvals do not shorten because the model improved. Department walls do not fall because tokens got cheaper. When one role becomes ten times faster, the next step may only get endless overtime and delays.

I. End-to-End Delivery Did Not Speed Up

At the end of 2024, Alibaba Cloud CIO Jiang Linquan’s team began using AI Coding tools at scale. He saw a pattern: AI generated more code, but end-to-end delivery of a customer requirement did not get much faster.

The team broke down the R&D process.

Developers spend about 20% of work time writing code. The rest goes to requirements, design, testing, review, launch, operations, and coordination across roles. Even inside code work, the hardest parts for AI, legacy compatibility, core logic, security, and production stability, are only 30% or less. Those tasks still sit with the most experienced people and take 80% of their time.

Sometimes the result is worse. A young engineer uses AI to produce in half a day what used to take several days. The senior engineers who can review code get hit by twice as much average code. Their work slows down. If AI only raises output, it can hurt end-to-end time. More code without value or quality control also makes the existing system heavier.

“The Mythical Man-Month” described a famous problem: adding people to a late software project often makes it later. New people need context. Experienced people must split tasks, explain, and train. Larger teams add communication links fast.

In the AI era, two conditions change.

First, adding a few agents to one engineer does not add several human coordination links. It raises single-node output and cuts communication nodes. Second, AI lowers the cost of learning adjacent skills. Moving from one role to another used to take long training. Now the learning cost can fall to 10% or even 1%.

In the past, product, design, frontend, backend, testing, operations, and architecture were separate roles because skills were scarce. One person could not master many of them. With AI, individual capability expands, and the division of labor changes.

The first changes in Jiang Linquan’s team came in testing and operations. Developers used to avoid writing tests because the investment was high. After AI Coding, checking coverage, adding cases, and generating logs and observability code became faster. Developers also had the fullest business context. Some testing and operations work moved back into R&D.

Requirements changed too. Product managers used to write a PRD, then hand it to design and R&D. A document cannot make several people hold the same product in their heads. Now product, design, and even customers can build a demo first. A mismatch shows up immediately. Rework that used to appear late in development moves to the far left.

Product managers, design, and frontend merged into the PDFE role, responsible from user intent to user interface architecture. Backend and architecture merged into the ABE role, responsible from data structure to system stability. The two sides connect through API contracts. This is the Half-Stack talent structure. It shortens the interaction chain.

Jiang Linquan found that the organization is no longer centered on skills. AI makes it easy to get something done, so skills have inflated. Taste has deflated. What is scarce is the ability to define the problem. He encouraged the team to shift left culturally and go to customers for value and answers.

A new question followed: a CIO’s R&D group can work this way, but can a large organization change on the same model?

II. Super Individuals Do Not Equal a Super Organization

At the AI Native Organization Forum at the Apsara Conference, Alibaba Group Chief People Officer Jiang Fang asked whether AI Native organization is just a hot concept or an inevitable direction.

Her answer: it is inevitable, but companies react differently. That has become a management consensus.

From a year of observation and practice, Jiang Fang found a pattern. Companies move fastest when the top leader understands AI and gets hands-on, when the work is already data-driven, when the team is small, and when the core team trusts one another.

Alibaba is a company of more than 100,000 people. Its businesses differ widely, collaboration is complex, and many decisions never enter the data chain. Drawing a future organization at headquarters and asking 100,000 people to switch together is not realistic.

Before trying to change, Alibaba felt another gap. Jiang Fang found that AI had produced a group of super individuals. But putting them together does not create a super organization.

That was the core question of the forum: from super individuals to super organizations, how far?

The answer is not to give everyone a tool. Business systems were designed for old ways of working. Human speed is limited. If a process is awkward or data definitions differ across systems, people use experience to keep going. Agents do not.

After general-purpose agents spread this year, Jiang Linquan found that business systems had a new group of users. An agent opens a browser, fails, retries, fails, retries again. When many agents work at once, call frequency and operation volume far exceed human use. Some systems that ran fine were quickly blown up.

Permission and security problems appeared. If an agent does not inherit the identity and permissions of the employee behind it, it can complete data operations far beyond human scale in a short time. That creates security risk.

Starting in February, Jiang Linquan’s team prepared business systems for the spread of general-purpose agents. Opening a few more MCPs would not solve it. If two systems define the same object differently, people can talk it through. Automatic agent orchestration can create chaos. Processes that were never online, data scattered in different places, and steps that relied on employee judgment all become exposed when machines take over.

The CIO team turned the process backward. It returned to the business, sorted workflows, and unified business concepts. It made experience explicit and online. Then it turned those capabilities into skills agents can call.

After this foundation was rebuilt, call data from opened cloud business systems showed that in four months, general-purpose agent calls passed GUI interactions and first-party agents. General-purpose call volume grew 1,000 times.

More business people began building agents with natural language. They came to the CIO team and asked, “Can this system give me an MCP?” Those needs had always existed. They had never entered the IT demand pool. Once systems opened, demand poured out.

The CIO team’s job changed. Simple needs closed inside business departments. The CIO team focused on complex projects. It also had to fix the foundation: check data consistency, identity, permissions, and agent audit trails. It had to find offline processes and personal experience that were never put online. The team called its second role “developer operations.” Business teams were the new developers. Each business team got a developer operations person to help with development problems.

After the internal work matured, the experience became products. Alibaba Cloud had been running 28 types of digital employees, equal to more than 2,000 headcount. Some methods and technology were refined into the Rui series, including Ruiyibao for translation and Ruihubao for outbound calls.

Ruiyibao handled full translation at this year’s Apsara Conference. It came from Alibaba Cloud’s globalization business. It used hundreds of thousands of internal technical documents and a large amount of GTM content, iterated over two years. It matches a senior translator, reaches 95% accuracy in simultaneous interpretation, compresses GTM delivery from weeks to 10 minutes, and costs 1/5 to 1/10 of human cost.

Ruihubao came from Alibaba Cloud’s telesales business. Under the same business metrics used for human agents, it matched or beat humans in renewal reminders, trial-to-paid conversion, and lead cleaning. Its cost is 1/5 of a human agent. Jiang Linquan’s standard for a digital employee is simple: it must act like a person and deliver results that meet business metrics.

III. A Large Organization Cannot Move at Once, So Let Some Teams Become Special Zones

Jiang Fang told a story.

This year, a business team went to Silicon Valley. One product leader did not want to go. He was watching key projects and felt he could not leave. He applied to stay. After Silicon Valley, the first thing he did was hand over the things he used to watch personally. Then he picked seven or eight young people and studied with them how AI could join the work: what information goes to AI, what steps can be deleted, where people should stay.

After the leader changed his own work, the team began to check old processes.

Jiang Fang made “leader gets in the water first” the first rule of AI Native change. If a manager has not used AI to complete a full piece of work, usage rates, token counts, and demos from below will not tell him which actions were merely accelerated and which processes should be removed and rebuilt.

Alibaba’s scale means not every manager can get in the water at once. Her second rule is “open special zones.”

The Qoder team is one sample. Five people, seven days, QoderWork. Plans were discussed, decided, and changed the same day. The team no longer pulled a circle of people into meetings for things it could decide itself. Ideas that take months to test in the industry were run in a week, then iterated.

How do you find teams for special zones? Yuan Yimin, HR Vice President of Alibaba Cloud Intelligence Group, says: find businesses that come with their own referee.

In the Alibaba Cloud international team, the online and telesales channel is one example. To test a commercialization strategy, the team can run placement, traffic, and upsell through the channel. The full reach chain has a data loop. ROI and conversion can be measured by the hour. Once the channel works, the lessons can spread.

How do you turn special-zone results into normal operations? Bao Chenxing, HR Vice President of Alibaba Cloud Intelligence Group for public cloud, says: make employees cross role boundaries.

In February, the team started a 100-day AI maker competition. Everyone had to join. The work did not have to relate to their jobs. People built family companion robot dogs. A post-00s management trainee built an AI video platform that turns your photo into a short drama lead. Someone built a crosstalk generator and sent the result to a famous crosstalk performer.

The point was to break limits, make people more full-stack, and blur role boundaries. Only then can handoff steps and unnecessary processes disappear.

IV. As Skills Get Cheaper, People Get More Valuable

Roles can be rebuilt. Systems can open. More execution work can go to digital employees. In the end, the problems return to people.

Jiang Fang ended her talk with people. Two months ago, another group of Alibaba colleagues went to Silicon Valley. Two things stood out.

NVIDIA still cares about employee values. Anthropic trains internal interviewers to judge technical ability and to judge whether a candidate’s values match the company’s mission.

In the AI era, a skill that is scarce today may become common after the next model upgrade. AI can lower skill thresholds. It cannot automatically create trust.

Bao Chenxing first doubted the “manager digital avatar” product. He thought it could only imitate a manager’s tone. Later, the team distilled the manager’s past materials and work experience into it, then used Harness engineering to put the avatar into live scenarios. Hundreds of employees began talking to it each day: What should we do when a customer hits this problem? How should we judge when a project is stuck?

A manager still has 24 hours. For the first time, part of the experience that depended on the person could be called on by more people at once.

Bao Chenxing tried it himself. He had to write OKR feedback for more than 20 employees in a quarter. In the past, time was short, and feedback was one or two vague lines. This time, he had Qwen Office read employee work materials and help organize them. He checked the result and sent it to colleagues for confirmation.

An employee told him it was the best feedback since joining Alibaba. Past OKR feedback was vague. This time it had data and specific things specific people were doing. Frontline work was seen.

Jiang Fang split AI Native organization into organization, talent, and culture.

At the organization level, do not stuff a tool into every role. Design end-to-end workflows first, then rethink where people should appear. People will become more generalized and full-stack, so role boundaries will keep blurring. At the culture level, accept uncertainty and learn to move on. Do not wait for tools to mature. Do not assume a method will work forever because it just worked once.

Different industries will not get the same organization chart.

The forum roundtable included guests from Kunlun Digital Intelligence, Yihai Kerry Arawana, and Galbot, representing energy, traditional enterprise, and embodied intelligence. Their data foundations, scale, and business forms differ. Some Alibaba lessons may not apply. But the way they find bottlenecks and manage organizations has common points.

A large organization cannot change all at once. Let the top leader get in the water. Let a small team run. Use results to decide what to spread. Every organization will find its own AI Native path.

V. Put the Small Motor Next to the Machine

Back to the question: from super individuals to super organizations, how far?

If you look at tools, the answer is close. One engineer plus a few agents can become a super individual. A five-person team can build a prototype in seven days. A business person can generate an agent with natural language. Skill thresholds fall. Individual capability expands. Organizational possibilities grow.

If you look at the organization, the answer is far.

A super individual is a point. A super organization is a surface. Efficiency at a point does not become system capability across the surface. If one role gets ten times faster, the next step can jam. If one team runs AI Native, another may still wait for approval in an old process. If one agent retries automatically, an old system can still be blown up.

A super organization is not a sum of super individuals. It comes from redesigning workflows, systems, permissions, data, culture, and talent structure around AI.

The task is not to distribute tools. It is to answer a few questions again:

Which processes should be deleted, not accelerated?
Which roles should be merged, not kept?
Which systems should be opened, not protected?
Which experience should go online, not stay in people’s heads?
Which people should appear in key positions, not all positions?
Which trust should be rebuilt, not replaced by AI?

There are no standard answers. Different industries, scales, and data foundations will produce different organization charts.

Models can improve. Tokens can get cheaper. Agents can get smarter. If the organization does not change, AI is just a large motor dropped into an old factory. The drive shaft remains. Belts and gears remain. Power still travels through the structure left by the steam age.

Only when small motors sit next to each machine can the factory escape the central shaft. Only when AI enters the workflow can an organization move from super individuals to a super organization.

The small motor is already next to the machine. The factory just has to be rewired.