Why is your company still using 20th-century organizational structures to do business in the AI era?

By: Deep Thought Circle

A few days ago on X, I saw a long post by Freda Duan. After researching the AI implementation of companies large and small, she found that: every company is putting AI tools into existing workflows, but almost nobody is asking why this workflow has ended up in this form.

A typical scenario: a company buys Copilot, issues licenses to everyone, and the CTO says at an all-hands meeting, “We need to embrace AI.” Three months later, they review the results—code generation is much faster, documentation is a bit smoother, and meeting minutes are automated—but the ROI hasn’t really been proven. Why? Because the organizational structure today only allows AI to provide small-scale enablement; real ROI requires restructuring the organization.

The real function of hierarchy

A textbook definition of organizational structure is a power structure—who reports to whom, and who has approval authority. But that’s just the surface. The real problem that hierarchical systems solve is information routing.

In a company above a certain size, you can’t let everyone see the full picture. So you install layer upon layer of managers, and they do two things: aggregate signals from the front line, refine judgments, and pass them upward; translate high-level strategic intent into executable actions and distribute them downward. Weekly meetings, daily stand-ups, QBRs, steering committees, cross-department alignment meetings—these are all information-routing devices.

But there is a structural paradox that is rarely discussed: the existence of departments and hierarchy is meant to solve the limitations of individual skills and capacity—one person can’t do everything, so division of labor is needed. Yet division of labor and hierarchy itself create new bottlenecks. Each time information passes through a layer of management, it degrades once; each time culture crosses a departmental boundary, it dilutes once. The larger the organization, the more severe the degradation becomes. So you need more meetings, more processes, and more middle layers to compensate for the degradation. More middle layers create even more degradation. This isn’t a management capability problem—it’s a death loop at the architectural level.

Over the past decades, all management innovations—Agile, OKRs, flattening, matrix-style structures—essentially have been local optimizations within this death loop. None has truly broken it.

What AI breaks is the loop itself. When the cost of information routing approaches zero, the organizational structures that exist to compensate for information decay lose their premise.

The real bottleneck is translation cost

Look at the delivery process for a product feature at a medium scale: the PM spends two to three weeks writing the PRD. The designer receives the PRD, understands the PM’s intent, and translates it into visual drafts. The engineer receives the visual drafts, understands the design intent, translates it into code, and provides a schedule—“eight weeks.” Then 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 the launch materials and trains the sales team. End-to-end, it takes three to six months.

The bottleneck on the surface is speed. But the real bottleneck is translation cost. The ideas in the PM’s head get encoded into documents; once the designer decodes them, they re-encode them into a visual language; the engineer decodes and re-encodes them into code; QA decodes and re-encodes them into test logic. Each translation loses fidelity; each translation requires alignment meetings; each translation creates waiting time. It’s not because people are slow—it’s because turning one person’s understanding into a format that another person can consume is inherently extremely difficult.

AI is collapsing these translation layers. With AI, a PM can go from idea to an interactive prototype in a day; the translation layer between the PM and engineering compresses to nearly zero. When writing code, AI generates tests in sync, and handoffs between development and QA disappear. An intelligent layer continuously synthesizes customer signals and business metrics. Middle managers who used to manually aggregate this information every week now have to redefine the source of their value. This isn’t that each role gets faster on its own. It’s that the gaps between roles—the translation layers, handoff queues, and alignment meetings—are evaporating.

The real change happens at the workflow level: it’s not about speeding up each step individually, but about reconstructing the entire chain end-to-end. The difference between the two isn’t a matter of degree—it’s a matter of paradigm.

A startup founder I recently talked with described a particularly interesting chain reaction. His development team used AI to compress a three-month development process into two weeks. The first reaction was excitement. The second was discovering that QA’s original two-week review cycle suddenly became a bottleneck that was as long as the development timeline—so QA was eliminated, and testing was embedded into development. Next, the one-month back-and-forth process to finalize things between PM and design turned out to reveal a new bottleneck—the PM team would only keep the most all-round people. Then the three-to-six-month GTM preparation period looked absurd compared with the two- to three-week product cycle—GTM was mostly AI-ized, running in parallel with development. The entire organization shrank by 80%, and end-to-end delivery compressed from nearly a year to one or two months.

The point of this story isn’t “AI makes people faster.” The key is the chain-bottleneck effect exposed after translation layers disappear: each time you remove a translation layer, the next slowest link reveals itself as the new bottleneck. This process won’t stop until the whole serial chain is flattened into a parallel, extremely small-team end-to-end workflow. If you deploy AI in only one link, the gains you see will be very small, because the bottleneck simply shifts to the next translation layer. You must reconstruct end-to-end; otherwise, it’s like putting a high-pressure water pump in front of the narrowest pipe.

Where most companies get stuck

If you look at a three-stage model—

First stage: old things, old ways—just swapping in a new tool. This is where the vast majority of companies are now. Corresponding to the role of AI in the organization: AI sits at the bottom as a tool—helping employees do work—while the organizational structure remains unchanged.

Second stage: old things, new ways—workflow reconstruction. The founder’s story above is an example of the second stage. The product is still the same product; serial becomes parallel, large teams become small squads, and translation layers are eliminated. The role of AI moves to the middle layer—starting to take on information routing, integrated judgment, and cross-functional coordination, tasks that used to be done by middle managers. The organization begins to flatten.

Third stage: doing things you couldn’t do before. Jack Dorsey once shared an example—before seasonal lows hit, a restaurant’s cash flow starts to tighten; the system detects the pattern, automatically packages a short-term loan, adjusts the repayment plan, and pushes it to merchants—before they even think to seek financing. No PM decided to build this feature. The system recognized the timing, combined existing capability modules, and emerged with a new product. AI is at the center, no longer just assisting human decision-making, but participating in requirement recognition, solution assembly, and resource allocation. The organization reorganizes around AI.

Most companies are stuck between stage one and stage two. The reason isn’t technology—technology is already ready. The reason is organizational inertia. Reconstructing workflows means moving the “moving pieces” in terms of power: middle managers lose their monopoly on information routing; functional departments lose their independent rationale for existence; approval chains get shortened significantly. Every step moves the existing power structure. That’s why the most successful AI transformations can only happen in Founder-led companies—that’s like restarting entrepreneurship all over again.

The skeleton of a new organization

Break an organization down to the underlying layer, leaving only three elements: information, decision, and action. Traditional organizations handle information through hierarchy, handle decisions through approval chains, and execute actions through departmental division of labor. AI rewrites the cost structures of all three, so the organization’s skeleton must be rebuilt.

From a relay race to a basketball game. Serial delivery—PM→design→engineering→QA→GTM—gives way to small teams of three to five people, with full skill coverage and synchronized progress. Most decisions are completed within the small team, with only directional bets escalating upward.

The logic behind this is: AI dramatically expands the coverage of an individual’s capabilities. A sufficiently good PM+AI can do what used to require a PM+designer+junior engineer. Individuals become long-distance runners—covering longer chains. When individuals are long-distance runners, the organization can be short-distance—fewer links, fewer handoffs, faster end-to-end delivery. A military analogy: from navy to navy seal. It’s not that there are more people and a bigger mass army; it’s that each person is an extremely capable elite squad member.

From departments to capability atoms. Don’t form teams by function groups; instead, split into independent, composable capability units—risk scoring, identity verification, collections, savings—each self-contained, with clear APIs and data interfaces, freely combinable.

Once capabilities have been atomized, the system can generate its own roadmap. Returning to Dorsey’s example—by combining existing capability modules like loans, repayment adjustments, and notification pushes, the system automatically brings forth a new product. The PM’s role shifts from translator to architect: defining the boundaries and quality standards of capability atoms, rather than moving information between people.

Quality shifts from checkpoints to guardrails. QA is no longer an independent audit step after development; instead, it becomes embedded constraints that run throughout the entire process.

Releases shift from big versions to a continuous stream. No more “Q3 launch v2.0.” Ship small improvements every day. Replace the leapfrogging rhythm of major releases with the quiet cadence of continuous delivery.

AI as a super employee: the overlooked second-order effect

What we discussed above is still change at the process level. The deeper impact is: once AI starts producing substantive output—no longer just assisting, but actually making things—the organization’s software also needs to be rewritten, not just the hardware.

Production relationships change. Traditional teams are collaborations between people. When AI becomes a core output node, managers face human-AI hybrid teams. Who is responsible for the quality of AI output? When AI writes 90% of the code, who is reviewing the code?

The unit of resource allocation changes. Traditional resource planning is headcount-driven—how many people and how many months a project needs. When two people + AI produce the same output as what used to require twenty people, headcount is no longer the correct unit for measuring input. Zuckerberg’s words: “Projects that used to require large teams can now be done by a single sufficiently excellent person.”

OKRs might actually become even more important. This is a counterintuitive judgment. AI enables each person to do ten times more—but the gap between what one “can do” and what one “should do” also widens tenfold. In the past, if one person could push three things in a quarter, and the direction was off, the loss was limited. Now, with AI, a person + AI can push thirty things in a quarter, and if the direction is off, the loss is also ten times as large. Ensuring everyone is doing the right things in an AI era isn’t less important—it becomes the most critical bottleneck. OKRs, as a mechanism for aligning direction rather than as a performance appraisal tool, are more valuable than ever.

Cultural impact is the most hidden. When an individual’s output can be five to ten times higher, traditional promotion ladders, title systems, and compensation bandwidths all start to look inappropriate. Between an IC who uses AI to produce ten times the output and a manager who manages twenty people but delivers comparable team output—which is more valuable? Traditional organizations don’t have a framework to handle this problem.

Big tech: unprecedented scale and change; but not AI-native yet

An “investment secret/technique” has always been to pick stocks in companies that are undergoing organizational restructuring—after large reorgs, there are often positive surprises in growth and margin. The market tends to overestimate the chaos of reorganization and underestimate the efficiency released by it. Today, there are never this many companies undergoing reorganization; the changes are never this big. From an investment perspective, you could say there are “potential candidates everywhere,” but so far, no one has seen a truly AI-native architecture that is genuinely eye-catching.

Meta has a 50:1 ratio of engineers to managers. In one year, it has reorganized countless times: integrating AI from a federated architecture into MSL, setting up Meta Compute for centralized compute planning, and shifting the organizational focus entirely.

Nadella said that 220,000 employees are a “major disadvantage in the AI race.” Three AI-related reorganizations in 18 months. Cut middle management and functions, unify the Copilot architecture, and merge internal model development. Microsoft’s annual employee cost is about $55-65 billion; even a 50% increase in individual productivity due to AI is huge. The most recent move in March 2026 was to unify the Copilot architecture, merge internal “superintelligent” model development, and promote younger executives to lead Copilot—an enormous shift.

Shopify saw eight senior executives leave (or be replaced) last year, and its general counsel was promoted to COO. The product was rebuilt around merchant data and AI-driven checkout. Cutting from a geographic perspective to a vertical industry perspective—this in itself is a signal: when AI enables you to understand the unique needs of each vertical more deeply, geographic-based routing is no longer the optimal information-routing approach.

Beyond Cook’s retirement, Apple also made a large-scale cut to the entire AIML org and moved Siri under Federighi’s software engineering organization. The AI leadership reports to the iOS/macOS delivery teams. Design is re-anchored to hardware engineering. Apple’s clearest signal is that AI is a delivery tool, not exploratory research. It’s a massive reorganization.

A shared pattern emerges: systematic compression of the information routing layer. But frankly, these are still big companies suffering through moving from stage one to stage two. Truly AI-native organizations may not exist yet.

Organizational boundaries are blurring

Up to now, discussion has been framed around “how to reorganize within a company.” But AI’s impact doesn’t stop there—AI affects not only internal structures, but also communication within and outside the organization.

When AI agents can automatically discover services, compare options, complete transactions, and handle payments, the translation cost between “company” and “user” also collapses. In the past, you needed sales, customer service, and marketing to translate value to users, handle questions, and complete conversions. In the agent era, most of those steps get automated.

This means the boundaries of organizational design extend. Not only internal structure, but also: can your services be discovered and called by agents? Where do you rank in an agent’s discovery layer? These questions will become as important as “your ranking in Google search”—or even more important—because agents don’t just present options; they directly complete transactions for users. And the conversion rate is many times higher than search ads.

Moats migrating

Over the past decade, the core narrative behind competitive advantage has been execution speed—who can deliver a better product to users faster.

Now, the moat is shifting from execution speed to learning speed—the speed at which an organization can absorb new possibilities brought by AI and restructure itself around them.

What most companies are doing is using AI to make existing structures run faster. Valuable, but not fundamental. The real gap is: if you started from zero today, knowing what AI can do, how would you build this company? The answer would not be “existing organization + AI tools.”

The answer is a shape we haven’t seen yet: individuals are long-distance runners, organizations are short-distance, capabilities are atomized, information routing is automated, and products emerge. The path to get there isn’t a one-time reorganization; it’s a continuous process of asking the same question: does this step still need people to translate? If it doesn’t, why do we still keep it?