Reading Zeng Ming's "Intelligence": After the Technological Revolution, How Business Reorganizes

Reading Zeng Ming's Smart, I had a moment of sudden clarity.

That feeling came from a thread running through the book: if you put technology back into business history, you realize that the companies, departments, positions, and even the division of labor across industries that we take for granted today all have their own birth eras. They were once good solutions to problems; over time, as they were used for so long, they came to seem like the way business was always meant to be.

AI makes these taken-for-granted arrangements questionable again. Why does completing one thing require so many people and so many departments? Why must a business be undertaken by one complete company? Are some things done this way because they must be, or because there was no other way in the past?

While reading this book, one sequence kept recurring in my mind: new technology emerges—large-scale application—supporting organizational forms gradually grow out of it. Applied to today, that means AI, large-scale application, and the AI-native organization that is still being explored.

When a new technology first appears, we usually use it first for familiar things: making existing work a bit faster and a bit cheaper. Only when it is used more deeply do we discover that some processes are no longer necessary, and that some businesses that could not be done in the past can now be done. Organizations also change slowly in this process. Today everyone is talking about what AI can do; what I care more about is how, once these capabilities become sufficiently widespread, companies will do business and how industries will divide labor.

The history of factory electrification happens to help us understand this period of change.

After electric motors were installed in factories

Economic historian Paul David studied this period of history. Early factories arranged machines around centralized power and drive shafts; where equipment was placed had to accommodate how power was transmitted. After electricity was adopted, at first it could simply be a matter of replacing the power source. Only when machines could be driven by their own electric motors did workshops have the conditions to rearrange themselves according to the production process, reduce handling, and loosen layouts that had previously been constrained by transmission devices. The value of electricity was further released through this kind of supporting change.

Switching to today, one department uses AI to write materials, another department uses AI to review materials; everyone is faster than before, yet the matter still has to go through a full round among several departments. The time consumed by waiting, relaying, and repeated confirmation will not automatically disappear just because materials are written faster.

Factories used electricity but still arranged machines around drive shafts—roughly the same predicament.

If further change is to be made, it is necessary to decide anew which information can be shared directly, which decisions can be made on the spot, and which work no longer has to go through the original departments. Managers who previously relied on reports to grasp the situation may need to look directly at business results; positions previously assessed by how many processes were completed must also be changed to being responsible for results. This involves power, interests, and work habits, and cannot be solved merely by purchasing a set of software. The deeper technology goes into business, the more it encounters these specific organizational problems.

Therefore, from new technology to new productivity, there is often a round of rebuilding work methods in between. Old organizations can continue to operate and can gain some benefits from new tools, but they may not be able to fully realize the potential of the technology. Electricity changed the production arrangements inside factories, and the internet pushed this change further to between enterprises: some businesses that previously needed to be kept within the same company began to be handed to external partners.

At this point, the discussion connects with Professor Zeng Ming's book Smart Business, published several years ago.

The Other Half of Smart Business

Read together with Professor Zeng Ming's earlier Smart Business, this book follows a consistent line of thinking: data intelligence plus network collaboration.

These two concepts are easier to understand when viewed within the same business. Take the apparel business: first judge the style, place production orders in advance, and then sell the goods through channels. Once the judgment is wrong, someone must bear the unsellable inventory. The larger the production batch, the lower the unit cost may be, but the cost of misjudging demand is also higher. Factories pursue production efficiency, while merchants want to hold less inventory. Each makes a reasonable choice on its own, but put together, the result is not necessarily the best.

If consumers' reactions to styles can be transmitted back to merchants in a timely manner, and merchants then adjust replenishment orders accordingly, factories can take on smaller and more frequent orders, and warehousing and logistics can adjust as well. The way of doing business then begins to change. Merchants can first test in small quantities and then add orders based on actual sales; the production arrangements factories receive are closer to demand as it is happening. A business that previously relied mainly on advance forecasting and stockpiling at every level has the opportunity to shift toward selling while adjusting, reducing the waste caused by misjudging demand.

Neither half can accomplish this alone. No matter how accurately sales data is read, if factories still require large minimum order quantities and fixed-cycle delivery, merchants can only watch opportunities pass by. Conversely, no matter how many suppliers there are or how convenient the connections are, if one does not know what to make or how much to make, it is merely passing inventory risk to others. A system continuously adjusts its demand judgment based on sales and inventory, applies that judgment to replenishment orders and production scheduling, and then revises it based on results; only then does data participate in operations. Network collaboration enables merchants, factories, and logistics to absorb these adjustments. Only when the two work together is it possible to change the original business.

Its commercial significance lies in the fact that demand, production, and fulfillment can be continuously adjusted across different enterprises, without having to renegotiate and coordinate every time a change occurs. Small enterprises participating in this network may also gain capabilities that previously only large companies could organize internally. Of course, this requires suppliers willing to change production scheduling, merchants willing to share information, and someone to arrange standards and interests. Network collaboration goes far beyond simply "everyone connecting to the internet."

In Intelligence, this line moves forward another step.

In the past, systems could transmit orders and record inventory, but when demand was ambiguous, conditions changed, or choices had to be made among several options, people still had to understand and coordinate. AI beginning to participate in judgment, task decomposition, and plan adjustment means that more capabilities can be invoked within the network. What was previously connected was information and already-defined processes; going forward, some work that requires ad hoc discussion and experience-based handling also has the opportunity to enter collaborative networks.

Zeng Ming discusses "intelligence compounding" in the book, and I think it should be viewed together with this change. Using the same model, two companies with different business experiences and different problems encountered will ultimately accumulate different capabilities. One merely uses it to handle scattered work, while the other can continuously retain tasks, results, and the correction process; only the latter is positioned to take fewer detours the next time it encounters a similar problem. Beyond the model itself, the depth of a company's contact with real business also affects how far it can go.

But data does not turn into experience on its own. Whether customers are satisfied, why a product did not sell, and where exactly a delivery problem got stuck require someone to follow up. Sometimes feedback is in the hands of external partners; sometimes different departments have fundamentally different understandings of what is good or bad. Only by connecting these links does "the more it is used, the better it gets" have a business basis.

Reading from Smart Business to Intelligence, what I focus on more is precisely this change: as intelligent capabilities become increasingly easy to obtain, how much of a company's original division of labor is still necessary? The way some positions work needs to change, and some business may not even need to remain inside the company.

The Boundaries of AI-Native Organizations

When discussing AI-native organizations, it is easy to think of fewer people: what a hundred people did in the past, ten people can now do, and taken a step further, a one-person company.

A reduction in headcount can of course happen. But if work still unfolds according to the original departments and processes, merely hiring fewer people is not enough to explain the novelty of such an organization.

The shape of a company is related to how it acquired capabilities and coordinated work in the past. Certain capabilities are hard to find externally, and even when found, difficult to coordinate with reliably, so recruitment, training, and placement within departments become necessary. As work grows more complex, more people responsible for coordination are added. Organizations gradually thicken, sometimes becoming bureaucratic, and sometimes genuinely bearing costs that were unavoidable under the conditions of the time.

When some capabilities can be obtained on demand, participants can understand a shared task, and work results can be verified, an enterprise can reconsider: which matters are worth maintaining a long-term team for, and which can be accomplished through external collaboration? The boundaries of a company will change with that answer.

Take a hypothetical consumer goods project. A small team is responsible for judging the target audience and product direction, delegating part of the research, design, and business analysis to AI agents capable of invoking tools and executing tasks, and then connecting to external prototyping, production, and fulfillment services. When demand changes, tasks and collaboration combinations are adjusted accordingly. It does not need to replicate the departments of a traditional company first in order to validate a new idea.

Of course, samples must actually be produced, production capacity must be arranged by someone, and delivery quality cannot rely on a model's verbal assurance. How far an organizational form can go depends on whether these links can keep up. If only internal office work becomes smarter while suppliers and production links remain disconnected, the capability boundaries of a small team remain where they are.

Therefore, when looking at AI-native organizations, focusing only on the internals of a single company is not enough. What we see may be a very small company, yet what supports it is a very large network. Scale has not vanished into thin air; it is simply not necessarily all reflected on the same company's employee roster and balance sheet. How large a business a team can do must be viewed together with the external capabilities it can organize.

The book's discussion of organizational memory then becomes easy to understand. Collaborating members can change, tasks can be recombined, but why a certain judgment was made, what methods were tried, and where failures occurred need to be retained. Otherwise, every time a new group of people or agents comes in, matters must be explained from scratch. The continuity of an organization depends both on trust among people and on the ability of commonly accumulated experience to be understood and used by those who come later.

What Network Collaboration Still Lacks

Having read this far, I still find myself wanting more on network collaboration.

The book discusses supply and demand networks, and also systems in which AI dispatches different capabilities. But after tasks are assigned, why each party is willing to accept them, how much information they are willing to hand over, and whether cooperation can continue sustainably are all areas that deserve much further elaboration.

This is still the same apparel business discussed earlier. The merchant wants the factory to adjust production promptly after seeing sales data, while the factory has to consider: if it reserves production capacity for you and the order ultimately does not come, what then? The factory wants more accurate demand information, while the merchant worries that once customer and sales data are handed over, it might be bypassed by its partner. Both sides know that coordination is more efficient, but that does not mean both sides are willing to bear the same risks.

AI can calculate production scheduling plans in greater detail and detect changes in demand more quickly, but it will not thereby eliminate the divergence of interests between the merchant and the factory. Behind every agent are the people and enterprises that have entrusted it with tasks. Getting several agents to reach a technical arrangement and getting the enterprises they represent to accept that arrangement are two different things.

In the past, one way to solve this type of problem was to bring the relevant business into the same company and coordinate it through internal management; another way was to rely on a platform, with the platform providing transaction rules, credit records and settlement services. The former entails heavier organizational costs, while the latter means that participants must accept the platform's rules. The platform organizes the business and therefore gains the power to decide how traffic is allocated and how transactions are charged.

This is where blockchain has a position worth discussing. Zeng Ming previously discussed how it might promote network coordination. What interests me more is whether, as cooperation becomes more frequent and participants can be dynamically combined, a portion of common rules and transaction records can be jointly verified by the parties rather than being entirely controlled by a single enterprise.

For example, which orders have been confirmed, which payment has been locked, and after what conditions are met settlement can occur — if there is a jointly recognized and verifiable common record, the work of checking with each party one by one may be reduced. For transactions with clear boundaries and verifiable outcomes, execution rules can also be agreed in advance so that settlement occurs as conditions are met. Participants still need to decide whether they are willing to do the business, but part of the confirmation and reconciliation costs in the performance process have a chance of being lowered.

This is especially worth studying when collaboration is broken down into smaller and more frequent tasks. A large order can take several days to negotiate and have dedicated personnel assigned to monitor it; if a large number of small tasks also had to be handled this way one by one, the manpower saved would very likely be consumed again by coordination. AI can lower the cost of doing things, but the accompanying transaction mechanisms must still make cooperation affordable for people, so that fine-grained division of labor can be sustained.

Of course, a ledger cannot replace all commercial judgment. Whether the factory has produced according to requirements and whether the quality of goods is qualified still depend on reliable inspection; whether a design is good and whether a service is satisfactory are even harder to summarize with an automatic condition. Technology can make agreements easier to verify and enforce, but it cannot negotiate a fair distribution plan on behalf of the parties. In many scenarios, handling matters through a platform that everyone trusts may still be more convenient.

Therefore, expectations for network coordination should also include discussion of the rules of collaboration. Whether the credit accumulated by participants can be carried away, whether switching to another platform requires starting from scratch, and in whose hands customer relationships and business feedback will remain — these all affect how much enterprises are willing to invest. Even if blockchain is adopted, one must continue to examine who ultimately controls the entry points, data and rules, and cannot judge that power has been dispersed based on the name of the technology.

When data intelligence and network coordination are considered together, what is being discussed is no longer merely efficiency. Enterprises can accomplish more within a network, but they may also become more dependent on the party that organizes that network. How much space new technology brings to small enterprises and how much power the platform gains from it will ultimately be reflected in these specific arrangements.

Position in the Industry

Looking at AI along this line, there is still a considerable distance between technological breakthroughs and industrial opportunities. What a model can do can only explain part of it. Who puts it into business, who organizes production and delivery, and who reaches customers and obtains continuous feedback will together determine what this business ultimately looks like.

One type of change occurs on the supply side of intelligence: how to make capabilities stronger, cheaper and more reliable. Another type of change occurs in the industries that use these capabilities: which needs were previously not worth serving, which transactions were previously impossible to coordinate, and which organizational costs previously had to be tolerated. The latter must enter specific businesses and handle a large number of matters that have no uniform answer, and therefore it is difficult to judge progress solely by the pace of model releases.

The same AI capability, when introduced into different industries, will encounter very different constraints. Businesses with digital delivery may reorganize work relatively quickly; businesses involving factories, equipment, and offline services still require real-world production and fulfillment to change along with it. In some segments, the cost of judgment has come down, yet scarce production capacity, channels, and customer relationships may thereby become more valuable.

"Intelligence becoming ever cheaper" does not directly lead to the conclusion that existing companies are all worthless. Models can help design more products, but production, channels, and customer trust still have to be earned through operations. Which capability has become easy to buy, and which segment still determines whether a business can succeed, need to be examined separately. The former may quickly become an industry standard, while the latter may gain greater bargaining room.

The "points, lines, and planes" discussed in the book are also easier to understand against this backdrop.

Some enterprises provide a single reliable capability, some organize complete services around demand, and some build networks that enable all parties to cooperate. A very important step in strategic choice is to figure out which layer one intends to occupy, and why others are willing to keep cooperating with oneself.

There is no platform position here that everyone should rush to seize. If everyone wants to keep customers and data in their own hands and leave others to be merely replaceable suppliers, network collaboration will also be difficult to establish on vision alone. Whether a new organization can be formed ultimately depends on participants being able to make their respective accounts work.

"New technology emerges—large-scale application—supporting organizational forms": this line gave me a more personal understanding of business history.

Enterprises in every era arranged the division of labor and coordinated cooperation under the technological conditions of their time. When technological conditions change, previously reasonable arrangements need to be reconsidered. What AI-native organizations will grow into cannot be conceived merely by subtracting some people from today's companies; it also depends on which new forms of cooperation start to make economic sense, and which previously unserved needs start to have a business worth doing.

This is where Zeng Ming's book Intelligence made me feel suddenly enlightened.

Models will continue to improve, but the work of reorganizing industries around models will not be completed by models on our behalf. How factories take orders, how enterprises cooperate with external partners, and what customers are willing to pay for—these changes need to happen slowly within each specific business.

Looking again at AI entrepreneurship, I would be more willing to spend more time understanding these matters. How much capability a small company can actually organize, and which costs a large company used to have to bear can now be shed. Only by thinking these questions through clearly is it possible to know which business one intends to do in the new industrial division of labor.