Original video in Chinese.
Key Takeaways
- Serial bottlenecks include integrating complex information, making key decisions, coordinating departments, and transferring context. AI exposes these constraints without automatically removing them.
- Let agents prepare evidence, align competing views, test assumptions, and organize objections in parallel, then give people a condensed decision map.
- Human value shifts toward defining questions and standards, managing context, designing evaluations, and taking responsibility for decisions. Small teams can build around that structure from the beginning.
If you want to build an AI-native company or team, remember the idea of serial bottlenecks. It is central to the opportunity and to the constraints traditional organizations struggle to overcome.
Giving every employee the strongest AI tool and connecting many agents to the workflow is not enough. Consider venture capital.
Before AI, an analyst read industry reports and wrote notes, the team discussed them, an investment memo was produced, and the investment committee reviewed it. Each stage waited for the one before it. A project might take two to four weeks.
Now give every analyst powerful AI with ample usage. Ten agents can produce ten reports in parallel. Yet someone still has to read those reports in sequence and make a judgment.
The research stage becomes parallel, while the human consumption and decision stage remains serial.
It resembles early factory owners using steam engines to pump water uphill so it could flow down and drive an old waterwheel. The new technology was attached to the old structure. The real steam era required redesigning power distribution, machine layout, and production.
Likewise, an AI-native organization must redesign itself around AI’s capabilities and continuously reduce the stages people must process one after another.
Where organizations wait
Integrating complex information: a person must read the material, understand it, form a view, and explain it before another can proceed.
Key decisions: an investment, product, or change of direction waits for a central decision-maker.
Cross-department coordination: product waits for engineering, engineering for legal, legal for finance, and finance for the boss. Each stage queues.
Context transfer: sometimes the most expensive asset is the background in someone’s head. Others cannot move until that person explains it clearly.
These sequential dependencies are fundamental constraints. Effort and intelligence alone do not remove them.
What AI changes
AI makes work that can be parallelized dramatically faster and cheaper. Competitor analysis, market scanning, data preparation, and first drafts can happen simultaneously.
That exposes the remaining slow stages. Research and initial analysis may take a day, but the project still waits for someone to read, understand, prioritize, and decide.
The central design challenge is therefore continuous reduction of serial bottlenecks. Both “reduction” and “continuous” matter.
Reduce the preparation required for a decision
Reduction does not mean making everything parallel. Strategy, investment choices, product taste, and content judgment still need people.
But agents can prepare thoroughly in parallel: read materials, collect evidence, align different views, list assumptions, organize opposing arguments, and mark uncertainty.
The decision-maker receives a condensed decision map rather than a pile of raw material. That reduces the amount of sequential processing required before judgment.
Repeat the redesign
AI capabilities keep evolving. A task that needs a human today may be largely assisted tomorrow. Background that requires repeated explanation may later be inherited through memory and context systems.
An AI-native organization behaves like an evolving product. Managers should regularly ask:
- Which tasks still require people to process them in sequence?
- Which parts can agents preprocess?
- Which can be structured?
- Where could better context management reduce communication costs?
The ongoing work is to make previously serial activities parallel, assisted, or automated wherever possible.
The human role
This places greater demands on people, whose value shifts toward five capabilities.
Define the problem. Know what deserves investigation. AI can help find answers, but choosing the most important question remains difficult.
Define the standard. Specify what makes a good report, judgment, product, or piece of content. Without standards, faster agents produce poor output faster.
Manage context. Historical decisions, failures, industry frameworks, and methods must be inherited by the system. Otherwise, each collaboration starts over.
Build evaluation. Decide what can be checked automatically, what needs human review, which errors are low risk, and which are unacceptable.
Make the final judgment. AI supplies information, evidence, counterevidence, and simulations, but people remain responsible for whether to act.
People become network designers, definers of taste, and final decision-makers.
Why small teams have an opening
Large companies have structural serial bottlenecks that a few new tools cannot remove. Small teams can begin with an agent network doing execution, while a few core people set direction, define standards, evaluate results, and make key decisions.
In my framework, each era has a leading organizational form. Huawei represents running the company as a process; ByteDance represents iterating the company as a product; Anthropic represents designing the company as a parallel network.
The first addresses scale, the second iteration speed, and the third serial bottlenecks.
AI gives everyone more power to act. The decisive factor is who can redesign the organization around it.
That is all for this episode. If you want to understand AI, become a more capable independent creator, and meet people with similar interests, join the newtype community. See you next time.