Authored by: Deep Thinking Circle

A few days ago, I came across a long tweet by Freda Duan on X, where she investigated the AI implementation situations in companies of various sizes and found that each one is stuffing AI tools into existing processes, but hardly anyone is asking why these processes are structured the way they are.

A typical scenario: the company bought Copilot, granted licenses to everyone, and the CTO said at an all-hands meeting, "We need to embrace AI." Three months later in the review—code generation has sped up significantly, documentation is slightly smoother, meeting notes are automated, but the ROI hasn’t truly been proven. Why? Because today’s organizational structure determines that AI can only do some minor enhancements; real ROI requires restructuring the organization.

The Real Function of Hierarchies

The textbook definition of organizational structure is about power dynamics—who reports to whom, who has approval authority. But this is just the surface. The real problem that hierarchical systems address is information routing.

In a relatively large company, it’s impossible for everyone to see the whole picture. Therefore, you install managers at various layers to do two things: aggregate signals from the front line, refine judgments, and pass them upwards; translate high-level strategic intentions into executable actions and distribute them downwards. Weekly meetings, daily stand-ups, QBRs, steering committees, cross-department alignment meetings—these are all information routing mechanisms.

However, there is a rarely discussed structural paradox here: the existence of departments and hierarchies is meant to solve the limitations of individual skills and productivity—one person cannot complete everything, hence the need for division of labor. But division of labor and hierarchies themselves create new bottlenecks. Information degrades every time it passes through a layer of management, and culture is diluted every time it crosses a departmental boundary. The larger the organization, the more severe the degradation, which necessitates more meetings, more processes, and more intermediaries to compensate for the degradation. More intermediaries create even more degradation. This is not about management capability; it’s a vicious cycle at the structural level.

Over the past few decades, all management innovations—Agile, OKRs, flattening, and matrix structures—have essentially been local optimizations within this vicious cycle. None have truly broken it.

AI breaks the cycle itself. When the cost of information routing approaches zero, the organizational structures that existed to compensate for information degradation lose their foundational premise.

The Real Bottleneck is Translation Costs

Let’s look at the delivery process of a medium-sized product feature: the PM spends two to three weeks writing a PRD. The designer receives the PRD, understands the PM's intent, and translates it into a visual draft. The engineer receives the visual draft, understands the design intent, and translates it into code, providing an "eight-week" timeline. Then the requirements change, and the PRD is rewritten. Development takes two to three months. QA receives the code, understands the expected behavior, and translates it into test cases. GTM prepares launch materials and trains sales. End-to-end takes three to six months.

The apparent bottleneck is speed. But the real bottleneck is translation cost. The ideas in the PM’s mind get encoded into documents, which the designer decodes and re-encodes as visual language; the engineer then decodes and re-encodes it as code; QA decodes and re-encodes it into testing logic. Each time a translation occurs, fidelity is lost, each translation requires alignment meetings, and each translation generates wait times. It’s not because people are slow; it’s that making one person’s understanding consumable for another person is extraordinarily difficult.

AI is collapsing these layers of translation. PMs can create interactive prototypes in a day with AI, compressing the translation layer between PM and engineering to nearly zero. AI generates tests while writing code, eliminating handoffs between development and QA. An intelligent layer synthesizes customer signals and business metrics in real time—middle managers who used to manually aggregate this information weekly must redefine their value sources. This doesn’t mean that every role is getting faster independently. It means the gaps between roles—translation layers, handoff queues, alignment meetings—are evaporating.

The real change occurs at the workflow level: it’s not about speeding up each stage, but about end-to-end reconstruction of the entire chain. The difference isn’t a matter of degree; it’s a difference in paradigm.

A startup founder I recently spoke with described a particularly interesting chain reaction. Their engineering team used AI to compress a three-month development process into two weeks. The first reaction was excitement. The second reaction was realizing that the QA, which originally had a two-week review cycle, suddenly became a bottleneck as long as development—therefore, QA was eliminated, and testing was embedded in development. Next, the month-long finalization process back-and-forth between PM and design was exposed as a new bottleneck—the PM team only retained the most versatile individuals. Then, the three-to-six-month preparation cycle for GTM seemed absurd next to the two-to-three-week product cycle—GTM was largely AI-automated, running in parallel with development. The entire organization shrunk by 80%, and end-to-end delivery was compressed from nearly a year to one or two months.

The point of this story isn’t that "AI makes people faster." The focus is on the chain reaction of bottlenecks revealed after translation layers disappear: every time a translation layer is cut, the next slowest link reveals itself as a new bottleneck. This process won’t stop until the entire sequential chain is flattened into parallel, small team end-to-end processes. If you only deploy AI at one stage, the benefits you see will be minimal because the bottleneck has just shifted to the next translation layer. You must reconstruct end-to-end; otherwise, you are merely adding a high-pressure pump in front of the narrowest pipe.

Where Most Companies Get Stuck

If we look at a three-stage model—

Stage 1: Same old tasks, same old method, just a new tool. This is where the vast majority of companies are now. Corresponding to the role of AI in the organization: AI is a tool at the bottom, helping employees do work, while the organizational structure remains unchanged.

Stage 2: The same old tasks with a new method, the processes are restructured. The story of the aforementioned founder represents Stage 2. The product remains the same, transitioning from sequential to parallel, large teams to small squads, and translation layers are eliminated. The role of AI has moved to the middle layer—beginning to take on routing information, making integrated judgments, and cross-functional coordination, which were previously tasks for middle managers. The organization begins to flatten.

Stage 3: Doing things that were previously impossible. Jack Dorsey shared an example—when a restaurant's cash flow begins to tighten before a seasonal low, the system detects the pattern, automatically packages a short-term loan and adjusts the repayment plan, sending it to the merchant—before they even thought about funding. No PM decided to create this feature. The system identified the moment, combined existing capability modules, and spontaneously generated a new product. AI is at the center, no longer assisting human decision-making, but participating in demand identification, solution combination, and resource allocation. The organization rearranges itself around AI.

Most companies are stuck between Stages 1 and 2, and the reason isn't technical—the technology is already ready. The reason is organizational inertia. Restructuring workflows involves moving existing positions: middle managers will lose their monopoly on information routing, functional departments will lose their independent reasons for existence, and approval chains will be drastically shortened. Every step disrupts the existing power structure. This is why the most successful AI transformations can only happen in founder-led companies—it's like a startup again.

The Skeleton of the New Organization

When we break down organizations to their core, three elements remain: information, decision-making, and action. Traditional organizations process information through hierarchy, handle decision-making through approval chains, and execute actions through departmental divisions. AI rewrites the cost structure of all three, so the organizational skeleton must be rebuilt.

From a relay race to a basketball game. Sequential delivery—PM → design → engineering → QA → GTM—gives way to teams of three to five people, covering all skills and progressing synchronously. The vast majority of decision-making happens within small teams, with only directional bets increasing.

The underlying logic is: AI drastically expands the coverage of individual capabilities. A sufficiently skilled PM + AI can accomplish what used to require a PM + a designer + a junior engineer. Individuals become long-range players—covering longer chains. When individuals are long-range, organizations can be short-range—fewer steps, fewer handoffs, faster end-to-end. A military analogy: from navy to navy seal. It’s not a larger army; it’s small elite teams where each person is highly capable.

From departments to capability atoms. Teams aren’t formed by functions but broken into independent, combinable capability units—risk control scoring, identity verification, collections, savings—each self-contained, each with clear APIs and data interfaces, ready to combine freely.

When capability atomization is complete, the system can generate its own roadmap. Returning to Dorsey’s example—the system combined existing capability modules for loans, repayment adjustments, and push notifications to automatically produce a product. The PM's role shifts from translator to architect—defining the boundaries and quality standards of capability atoms rather than transporting information between people.

Quality shifts from pitfalls to guardrails. QA is no longer an independent review phase after development but an embedded constraint throughout the entire process.

Releases transition from major versions to continuous flows. There will no longer be "launching v2.0 in Q3." Instead, there will be daily small improvements, replacing the jumpy cadence of major version releases with quiet, continuous delivery.

AI as a Super Employee: The Overlooked Second-Order Effects

What has been discussed so far is still about process-level changes. The deeper impact is that when AI starts to produce substantial output—not just assisting but actually creating items—organizational software must also be rewritten, not just the hardware.

The production relationships have changed. Traditional teams consist of human collaboration. When AI becomes a core output node, managers face human-AI hybrid teams. Who is responsible for the output quality of AI? When AI writes 90% of the code (the current situation with Anthropic), who is the subject of code review?

The units of resource allocation have changed. Traditional resource planning is headcount-driven—how many people and months this project needs. When the output of two people + AI equals what used to require twenty people, headcount is no longer the correct measure of investment. Zuckerberg's own words: "Projects that used to require large teams can now be done by one sufficiently capable person."

OKRs might become even more important. This is an intuitive judgment. AI allows individuals to do ten times more, but the gap between "can do" and "should do" has also widened tenfold. In the past, one person could push three projects in a quarter, and a slight directional shift resulted in limited losses. Now, one person + AI can push thirty items in a quarter, and a directional shift results in ten times the loss. Ensuring that everyone is doing the right thing has become the most critical bottleneck in the AI era, making OKRs, as a directional alignment mechanism rather than a performance evaluation tool, more valuable than ever.

The cultural shock is the most subtle. When individual output can be five to ten times what it was, the traditional promotion ladder, title system, and salary bands become inadequate. An IC who produces ten times more with AI and a manager who oversees twenty people but with equivalent team outputs—whose value is greater? Traditional organizations lack the framework to address this issue.

Large Corporations: Never This Big/Many Changes; Yet Not AI-Native

One investment "secret/trick" has always been to select stocks from companies undergoing organizational restructuring—typically, after a large reorganization, there are positive surprises in growth and margins. The market tends to overestimate the chaos of restructuring and underestimate the efficiency released by it. Today, the number of companies undergoing reorganization has never been higher; the changes have also never been greater. From an investment perspective, it’s fair to say there are "many potential candidates," but as of now, I haven't seen a truly AI-native architecture that stands out.

Meta has promoted a 50:1 engineer-manager ratio and has undergone numerous reorganizations within a year: integrating AI from a federalized structure into MSL. Launching Meta Compute for centralized computational planning. The organizational focus is thoroughly shifting.

Nadella said 220,000 employees are "a huge disadvantage in the AI race." Three AI-related reorganizations in 18 months. Cutting middle managers and functions, unifying Copilot architecture, and merging internal model development. Microsoft's employee costs are about $55 billion to $65 billion a year, and if AI can improve everyone's output even by 50%. The most recent one was in March 2026, unifying Copilot architecture, merging internal "super-intelligence" model development, and promoting younger executives to lead Copilot—this is certainly a significant move.

Shopify saw eight executives leave or be replaced over the past year, with the general counsel promoted to COO. The product was restructured around merchant data and AI-driven checkout. The shift from geographical dimensions to vertical industry dimensions—a signal in itself: when AI allows you to understand the unique needs of each vertical industry more deeply, dividing by geography is no longer the optimal method of information routing.

Apple not only experienced Cook's retirement but also drastically cut the entire AIML organization and moved Siri under Federighi's software engineering organization. The AI leadership now reports to the iOS/macOS delivery team. Design has been re-anchored to hardware engineering. Apple's signal is the clearest: AI is a delivery tool, not exploratory research. This constitutes a massive reorganization.

A shared pattern: systematically compressing information routing layers. But frankly, these are still large companies struggling to move from Stage 1 to Stage 2. Truly AI-native organizations may not yet exist.

The Boundaries of Organizations Are Blurring

Up to this point, the discussion has been within the framework of "how to reorganize within the company." But the impact of AI goes beyond this—AI affects not only internal processes but also communication between organizations and the outside world.

When AI agents can automatically discover services, compare options, complete transactions, and process payments—the translation costs between "company" and "user" are also collapsing. In the past, you needed sales, customer service, and marketing to interact with users—explaining value, addressing queries, and completing conversions. In the agent era, most of these stages are automated.

This means the boundaries of organizational design are extending. It's not just about internal structure; there's also: can your service be discovered and called by agents? Where do you rank in the agent's discovery layer? These questions will become as crucial as "where do you rank in Google searches"—even more important, as agents not only display options but directly complete transactions for users, with conversion rates several times those of search ads.

The Migration of the Moat

For the past decade, the core narrative of competitive advantage has been execution speed—who can deliver better products to users faster.

Now the moat has shifted from execution speed to learning speed—how quickly an organization can absorb the new possibilities brought by AI and restructure itself around them.

The vast majority of companies are merely using AI to make existing structures run a bit faster. It’s valuable, but it doesn’t touch the fundamental issue. The real question that creates a gap is: if starting from scratch today, knowing what AI can do, how would you build this company? The answer won’t be "existing organization + AI tools."

The answer is a shape we haven’t yet seen—individuals are long-range, organizations are short-range, capabilities are atomized, information routing is automated, and products emerge spontaneously. The path to get there isn’t a one-time reorganization but a constant inquiry into the same question: does this step still require a person to translate? If not, why do we still retain it?

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。