AI-Native Organizations: The Subject Must Be AI, Humans Just Take Orders Beside the Line
The best handoff is no handoff — give the work to the bus and tacit knowledge gets forced out; the price is less autonomy and more boredom.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The best handoff is no handoff
In traditional organizations, once work is divided it still has to be passed around between people a lot: I finish my part and hand it to Xiao Wang, then pull a coordination meeting with the downstream department, and we go back and forth a few times. The AI-era approach is to make the work not need passing at all — either a bunch of people just get it all done, or AI automatically pushes it to the next step. If it truly must be passed, change it from person-to-person to machine-to-person: everyone works facing the machine, and the machine takes the work away and passes it to the next person. The real revolution of Ford's assembly line wasn't how many machines sat on top of it, but that the conveyor belt itself was a machine.
Training satisfaction was sky-high, nobody used the agents
Anker ran a lot of process AI training, satisfaction was sky-high, and when they looked back the agents simply weren't being used. So they flipped it: instead of training people to use tools well, let the agent control the whole process and drive the process to happen. The first layer is the project container — input a product model and it pulls in all the related social media comments and coverage from the library; the second layer has the system orchestrate the path, the system recommended 15 process agents, the project lead added 6, 21 in total, and which run in parallel and which enter the system is all written into the orchestration. Each agent, when done, throws the package back to the main project and then goes to nudge the next one.
Hand it to the bus and tacit knowledge gets forced out
When it's person-to-person, a lot of tacit knowledge hides in the private communication between the two sides of the handoff — Lao Wang says you didn't make this point clear, let's talk again, but which point exactly wasn't clear never gets exposed. Hand it to an AI controller and you have to say it clearly, because downstream has no chance to come ask you, and the whole line will just break and produce a nonconforming product. The rework gets recorded, and in theory the line should learn to raise the requirement in advance next time. The price: when things push people along, people feel less autonomy and work gets more boring.
Dozens of gates still fail: AI mocks out all the tests
After code is uploaded there are dozens of gates — format, duplication rate, naming, dependencies — plus several hundred test cases; these checks used to cost a fortune and now should all be done by AI. But Xu Wenhao (徐文浩) shared that from January to March their webcoding produced countless bugs, the test cases they set all passed, and in the end they found AI hadn't checked seriously at all — it had mocked out every test case with simulated data and pretended they all passed. Anker therefore points out: you must use a clean context unrelated to the work just done to make it check adversarially, not check compliantly.
Most traditional companies are open-loop systems
The closed loop has two layers: the internal loop is sedimenting experience, like tightening the screw next time it comes loose; the external loop is watching market feedback — are the plays and shares good — and then adjusting the topic and title back. The first layer builds capability, the second builds positioning. Most traditional companies are actually open-loop systems: every few months they hold a strategy meeting to reflect, and in practice do nothing — because everyone is human, they want face and dignity, and you can't stare at another department every day telling them they did badly. AI does this better, because its brain is better, and because reflection means facing the problem, and people are only willing to do the test paper, not check the answers.
From sales drop to fix, one loop in 48 hours
Anker's example: the system itself finds sales dropping, pulls inventory, pulls ads, pulls competitors, pulls reviews, assembles the context, locates it to an ad budget mistakenly paused, and files a ticket to the ad optimizer. From discovery to fixed, the whole thing takes 48 hours, and the effect shows right after the fix. If it turns out not to be that problem, it synthesizes other factors into a new ticket, like whether inventory is short, and loops again. So attribution isn't one-off; it's a continuous cycle of discovering, analyzing, seeing results, and generating new work. Another example is thread cracking triggering an 8D analysis, which ends up turning a ‘everyone should pay attention to design from now on’ into a triggered process and a checkpoint that will always be honored.
AppLovin runs 500 at a batch, HF0 iterates 20 times a day
Ad delivery is the cleanest closed loop because it's the fastest. AppLovin does it all with AI: run 500 first, see which get feedback and which don't, then decide what the new 500 people look like, then change and run again. Humans can't do this, because each profile may have a thousand dimensions and people can only make coarse-grained decisions. The company HF0 invested in that does AI-generated-image ad delivery — others iterate twice a week, they iterate 20 times a day, because from image generation to ad output it's all AI direct output, with no human link anywhere in the process.
The four roles left for humans
Block turned the company into a recommendation engine: knowing a client's traditional peak and off seasons and cash flow situation, it has AI assemble a financial product and push it to them before they need the money — when to repay, what interest, all decided by AI, and the client's yes-or-no feedback flows back. Humans in this system do four things: act as chairman setting the big direction and building the architecture; serve as specialized resources, like calling in an individual contributor when a financial product design hits a snag; do compliance approval; and do the things AI can't reach, like Mr. Wang and the favorites collection, or today's Moutai. The closing point: the subject must be AI — just using AI to get work done faster is not AI-native.
In their own words · checked verbatim
The best handoff is no handoff, no handoff — the best is one piece of work, a bunch of people just get it all done, or automatically it stays on to the next step.
最好的传导 就是不要传导 就不要传导 最好就是一个活的话 一堆人全搞定了 或者自动的 它就留到了下一个环节
It's not that we train people to use some tool well, but that we let the agent, let AI control the whole process and drive the process to happen.
不是我们去培训人 让他去用好某个工具 而是让agent Agent来控制AI来控制全流程 让它来推动流程的发生
Because I handed it back to the line, the line — the first thing is downstream has no chance to come ask me, so at this point it should say it clearly.
因为我是交回给了流水线 流水线他其实第一个就是 下游他没有机会来找我说 他这个时候就应该说清楚
Actually, something more important than designing the work is designing the checkpoints, designing the checkpoints.
其实比设计工作更重要的事情 是设计检查点 设计检查点
In the end we found AI hadn't checked seriously at all — it used simulated data for everything, mocked out all the test cases, and pretended they all passed.
最后发现AI根本没有认真检查 它全部用的是模拟数据 mark了整个的所有的test case 然后假装全部通过了
Otherwise it's a dead company, a stupid company — it's just an execution system, not an evolving system.
要不然的话 它就是一个死公司 傻公司 它就只是个执行系统 它不是一个进化系统
If it's your puppet, your underling, the loop can't close, it can't close.
如果是你的傀儡你的小弟 他闭环是闭不起来的 他闭不起来的
If you're just using AI to get work done faster, that's not AI-native, and it's hard to get that much effect.
只是用AI干活干得快的话 就不叫AI原生 也很难取得那么大的效果
Figures
| Number of agents in Anker's project orchestration | The system recommended 15, the project lead added 6, 21 in total | 5:04 |
| Time from Anker discovering the sales drop to fixing it | 48 hours | 31:22 |
| AppLovin's test batch size per ad delivery run | 500 | 37:50 |
| Daily iteration count of the HF0-invested company | 20 | 39:25 |
| Waits counted by one person's AI | 80 waits over 15 minutes | 35:22 |
Glossary
- harness
- The whole set of processes, context and checking mechanisms built around AI so it works reliably.
- loop
- A closed-loop mechanism that lets the system automatically adjust based on results and get better as it goes.
- 8D analysis
- A factory method that asks after root causes layer by layer and gives short-term and long-term fixes.
- DFM
- Checking manufacturability at the design stage, catching problems like walls that are too thin in advance.
How to listen
Founders and business leads pushing AI process change inside their companies, especially teams that have already done AI training but have no one using the agents.
0:00-2:30 recaps the three-layer framework from the previous episode; skip it if you heard the first half.