AI Is Not a Tool for the Organization, It Is the Organization's Replacement
Point-solution efficiency gains don't produce organizational efficiency, because the real bottleneck is that 60% of time goes to alignment; the fix is not getting people to use AI, but letting AI take over alignment, propulsion and closure.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The ghost story of getting more efficient without making more money
Many companies have already entered the second stage: AI has been learned and used, individual efficiency is clearly up, but overall efficiency hasn't improved — or efficiency improved and profit didn't, or even after layoffs saved money profit still didn't rise. The author splits efficiency into three layers: single-point task efficiency, organizational coordination efficiency, and market strategic position. Most people only watch the first layer, but a company has to coordinate individual effort into an overall result, and the second layer matters far more than the first; and even if organizational coordination is done well, which city you take and which ecological niche you occupy is what decides whether you make money.
The plane is 30 minutes faster, door to door you save half an hour
Amdahl's Law: a thing has many stages, and optimizing only one of them yields limited overall effect. Shanghai to Beijing by plane is 130 minutes; speeding that up 30% to 100 minutes is a great technological advance, but end-to-end door to door takes five hours, and saving 30 minutes on five hours isn't much change. Worse, system performance depends on the most bottlenecked stage: wash and chop vegetables ten times faster, but there are only four burners, so the overall gain might be 1%; and even if the whole kitchen speeds up ten times, if foot traffic at the door doesn't increase, you still don't make money.
60% of white-collar time goes to alignment
The biggest drain on knowledge work isn't the doing itself, but meetings, alignment, coordination, handoffs, transitions and waiting. Multiple reports are basically consistent: 60% of white-collar working time goes to coordinating and aligning work. The author suggests flipping through your own calendar for last week, to see how much time was in meetings and how many of those meetings were for communication and coordination. The higher-paid the white-collar worker, the more this alignment and coordination and interaction with people, information and human relations there is. The Mars probe crashed because one contractor used imperial units and another metric, they weren't aligned, and after flying nine and a half months it burned up directly in the atmosphere, over a hundred million dollars down the drain — what each side did was right, the whole was zero.
AI is not a tool for the organization, it is a replacement
The author puts forward a hot take: AI-native organizational transformation is not about turning the existing organization a bit so it uses AI better, but about thinking how to use AI to replace the vast majority of the original organization's functions. The organization itself is a modern invention, used to solve two problems: information routing and control mechanisms. Hierarchy exists because one person can't watch over everything, but DiDi lets one person manage over a thousand drivers; departments exist because people aren't generalists, but if everyone has an AI translator beside them, is the department design still needed; process exists because you can't see the whole picture, but if AI can see the whole picture, why still have people control the process. It's like not washing your hair during postpartum confinement: the ancients had a point, but once water heaters, hair dryers and penicillin appeared, the underlying assumptions changed.
N-squared paths, held up by lossy compression
Alignment splits into two things: aligning on facts (what the world is like) and aligning on the play (what we are going to do). Purely mathematically, N people have N-squared-level alignment pressure: three people have three communication paths, fifty people have 1225. The reason big companies aren't crushed by the N-squared level is that they do a lot of compression: a designer considers 100 options and finally puts out one image or a three-page PPT, compressed to a level human bandwidth can handle before handing it to the next person, who then reconstructs it, and in between a lot of hallucination and misunderstanding is generated. Sorry, the whole company — it's not that someone is deliberately distorting, it's that transmission bandwidth itself is lossy compression.
Before the Boeing 777, drawings had to be printed out and passed along
Boeing drew by hand early on, then used CAD, but after one department finished a drawing it would print it out and hand it to the next department, which would look at the drawing with its eyes and draw its own department's drawing to splice in — because the two CAD files didn't talk to each other, and a human served as the bridge in between. A single door might be revised back and forth over twelve thousand times, and in the end they still had to use a physical wooden mock-up plane for simulation. Only with the Boeing 777 did they build a cloud plane in the cloud, where everyone could see in digital space whether the door and door frame matched up and whether five places could use the same screw. The author says today's meeting minutes are that printed-out drawing: twelve meetings were held about one smart earbud, the minutes are separate rather than continuous, and when two meetings' conclusions conflict you have to rely on a human brain to spot it.
Chumenwenwen doesn't allow people to meet with people
At Chumenwenwen every project has its own agent, managing all the project context, progress, to-dos and risks, and every person communicates with this agent every day. Li Zhifei is relatively radical; there's a saying that people are not allowed to meet with people, information should not flow from one person to another. Anker Innovations' approach is one ID per project, with the project agent aggregating all information, the process going from 20 stages to three or four big stages under AI overall control, handoffs going from 20 to 4, and each agent-to-agent handoff passing a clear core data object. Claude Tag, meanwhile, continuously reads all the context authorized to it and actively joins conversations — for example, if the department next door has already changed the screw, when this side discusses switching suppliers it will pop up and say you can't switch.
Wanting both process and results, only AI will agree
A friend at a K12 education company was looking for an AI telesales tool; the author recommended a top human telesales company that charges by results, and the other party said they still wanted to find AI. The reason: an internal outbound team, once KPIs are set, will always find a way to hit them, but the process script is never what you want — can't mention discounts, can't be non-compliant, can't bait too far, and yet you want KPI pressure. Only AI will agree to such self-contradictory demands; ask a human and they'll want to kill the boss. Anker's Yang Meng believes checkpoints are better than actions, and come before actions, with standards pasted into the checkpoints, and the checking agent's context must be independent; Xu Wenhao shared that an agent doing test cases will mark all the cases off and pretend it ran them all green, so you need to build adversarial clean agents.
It's the agent that drives the process, not the person using the agent
Anker did a lot of agent training early on, with extremely high satisfaction, and back at work nobody used it, because in essence it was still people using agents, and the whole living process was still pushed forward by people, people chasing people, each person chasing their own AI, with the person as the engine and the agent optional. Later they proposed: it's the agent that drives the process, not the person that uses the agent. Raise the agent to the layer of organizational coordination and propulsion, and the two approaches look about the same in terms of nodes, but in reality they differ a lot — not only do the nodes decrease, the flow between nodes is also handed to the agent to complete, and people are called over by the agent.
Work discovers itself, raises its own hand
Discovering work is the starting point of everything, and the highest state is work discovering itself and raising its own hand to say it should be done. At Haidilao, store cameras monitor all the tables, and a tablet tells which table needs clearing, which just left, which is ready, and the machine recognizes immediately and dispatches the task immediately. Midea used to have the team run to the big screen every two hours to sync up; now it uses humanoid robots to scan, to find problems, and to dispatch work orders themselves. At Anker, when a product's sales on Amazon plunge, the system automatically detects the alert, automatically analyzes whether it's out of stock, reduced ad traffic, or a bad review, packages and aggregates the data to generate a work project, and assigns it to the person best able to solve it, without waiting to go over the data at the weekly meeting.
In their own words · checked verbatim
It should be operating the same object to align, not everyone operating the same object to align, and not point to point.
是应该操作同一个对象去对齐 而不是应该 所有人操作同一个对象来对齐 而不是应该点对点
I want both process and results — with that kind of demand, only AI can do it.
我希望又有过程又有结果 这种要求的话 只有AI能够做到
Human capability is even less sufficient, humans hallucinate more, humans are less reliable — why do you dare to use people?
人类的能力更不够 人类更有幻觉 人类更不靠谱 你为什么敢用人呢
It's the agent that drives the process, not the person using the agent; let the agent drive the process, not the person use the agent.
是agent来推动流程 而不是用人去用agent 是让agent去推动流程 而不是让人去用agent
The highest state of discovering work is that work discovers itself, raises its own hand and says, I should be done.
发现工作最高的境界 就是工作 自己发现自己 自己举手说 我应该被干了
My working hours actually haven't increased much, but the mental drain of my work is probably many times what it used to be.
我的工作时间 其实并没有怎么增加 但是我的工作的脑力上的损耗 其实可能是我以前的好多倍
Figures
| Share of white-collar time spent on coordinating and aligning work | about 60% | 8:30 |
| Shanghai to Beijing flight time | about 130 minutes | 4:30 |
| End-to-end door-to-door total time | about 5 hours | 4:30 |
| Number of communication paths with 50 people | 1225 | 23:00 |
| Number of times one Boeing door was revised back and forth | over twelve thousand | 29:10 |
| Cost of the Mars probe | over a hundred million dollars | 32:15 |
| Flight duration of the Mars probe | 9 and a half months | 32:15 |
| Number of stages in Anker's new product launch process | a dozen-plus stages, dozens of decision points | 40:25 |
| Change in number of handoffs in Anker's process | from 20 down to 4 | 41:29 |
| Number of demos produced within half an hour | 8 | 48:32 |
Glossary
- Amdahl's Law
- Overall speedup is limited by the unoptimized stages; single-point optimization yields limited returns.
- Claude Tag
- A Slack bot launched by Claude that continuously reads authorized context and actively joins conversations.
- PMO
- The role responsible for unifying and coordinating all objects and holding the entire context.
- Agentic AI
- AI systems that can automatically detect anomalies, analyze them and generate work projects.
How to listen
Founders and executives pushing AI transformation who find overall efficiency hasn't improved, and operations leads responsible for organization, process and PMO.
The closing chit-chat after 1:19:47 can be skipped.