For thirty years organizations built walls between departments. Then they spent a decade tearing those walls down. Somehow we ended up with more walls.
I keep coming back to how those divisions affect the work. Roles narrowed. Functions separated. Each person got very good at their own slice, and the job of fitting the slices back together was handed off to structure — hierarchies, processes, the org chart itself. It was a reasonable trade. When understanding the whole system costs more than it is worth, you stop asking people to do it and let the structure carry that load.
Then AI changed both sides of the equation, and almost nobody updated the trade.
The cost of understanding broader contexts is dropping. The value of that understanding is rising. And the people who can move between the strategic view and the operational floor — who see where one decision quietly propagates into three others — are turning into the scarcest resource in the building. Not the deepest specialists. The ones who can see where all the specialists’ work collides.
Specialization Scaled. Visibility Didn’t.
The logic of specialization was always sound. Break a complex system into parts, let people master those parts, and trust coordination to stitch them back into a whole. This works beautifully when execution is slow enough that misalignment surfaces before it metastasizes.
But specialization has a side effect that never makes it onto the org chart: it points everyone inward. Performance reviews ask whether you hit your number, not whether the company did. The result is depressingly predictable. 83% of companies report experiencing organizational silos, and the cause is structural — when performance is measured by departmental output, no amount of collaboration training changes how people work day to day [3].
So you get functional strategies that conflict with or dilute enterprise priorities, each locally optimal and collectively incoherent. Every department is winning its own game. The company is losing the only one that counts.
That was survivable when systems moved slowly. A choice made in supply planning took weeks to reach manufacturing, and somebody usually caught it on the way. AI compresses that window to nearly nothing. When execution accelerates, optimization inside a silo turns into misalignment at the system level before anyone has finished reading the dashboard.
More Data, Less Clarity
You would think the cure for a visibility problem is more visibility. We certainly tried. Dashboards multiplied. Metrics got granular enough to track the color of the carpet. And for anyone who has signed off on a BI budget, here is the uncomfortable part: it did not help.
78% of companies still fail to convert their information into decisions [6]. Not because they are short on data — because they are drowning in it without the context to know which numbers matter. There is a particular comedy to answering a system you do not understand by bolting another dashboard onto it. You now have a higher-resolution view of something you still cannot read.
AI does not rescue you here on its own. It can bridge gaps of time, distance, and scale, but only when it is built to support shared goals and group context. Drop it inside a silo and it simply makes that silo faster — producing locally-brilliant, system-incoherent output at a higher rate than before. The abstraction that once simplified coordination starts hiding root causes, because the system now changes faster than anyone can observe it.
What End-to-End Operators Do
So what is the missing capability? It is worth being precise, because “see the whole system” sounds like the kind of thing you would print on a mug.
It is not omniscience. It is a cognitive pattern. Roger Martin spent years studying how the best leaders think and found one consistent habit: integrative thinkers deliberately broaden what they treat as salient even though it makes the problem harder, reason about multi-directional rather than tidy linear causes, and keep the entire problem in mind while working on individual segments [4]. That last clause explains the capability. They hold the system in view while operating on the part.
In practice, the end-to-end operator slides between levels of abstraction — a strategic metric, the process that produces it, the execution underneath, and back up again — diagnosing root causes by tracing through the layers instead of treating whatever symptom showed up in their own domain. The image I keep reaching for is a restaurant: most people work their station, but a rare few read the whole room — the kitchen, the floor, the suppliers, the look on a guest’s face when the entree lands late.
This is exactly the work AI cannot do but badly needs done. As AI absorbs more specialized tasks, the valuable human work shifts toward judgment, oversight, and orchestration [1]. And the premium is already moving: lower-skilled workers saw productivity gains of roughly 35% with AI, while highly skilled specialists saw minimal gains and sometimes a dip in quality. Deep expertise still matters — but its relative advantage is narrowing, because the routine edge of specialization is the first thing AI floods.
This is also the capability that decides who stays valuable as they grow more senior rather than getting managed out — what turns long experience into an asset instead of a liability, a thread I pull on in Over 50, Underestimated. And it is the cross-system tacit knowledge that walks out the door when senior people leave — the heart of The Expertise Cliff.
The System That Built One: Toyota’s Chief Engineer
The best proof that this capability is real — and buildable — is decades old and lives in a car factory.
Toyota built a role called the Chief Engineer, or shusa, who owns a vehicle program end to end: the concept, the business case, the technical design, coordination with manufacturing and sales, all the way through to launch. The CE leads a team of fewer than ten. The hundreds of engineers who actually build the car report to their own functional managers, not to the CE. The Chief Engineer has almost no formal authority at all.
What the CE has instead is end-to-end visibility and the credibility that earns. They are expected to have driven every competitor’s vehicle, sat with real customers, and understood the manufacturing constraints firsthand. The Lean Enterprise Institute describes the role not as another box in the matrix but as “an operating system that connects across and up and down the organization” — a person doing the job most companies hand to a committee.
The results are not subtle. Toyota delivers vehicles in 12 to 18 months against an industry average of 24 to 36, with fewer late engineering changes and happier owners. The advantage, as Morgan and Liker documented, came not from better specialists but from better integration of specialization [2]. Competitors spent years copying Toyota’s processes and missed the point: the integration lived in a person, not a process map.
The modern version runs on AI. Kraft Heinz spent years rebuilding its planning as one connected system rather than function by function — demand, supply, and material planning unified, computer vision tying quality control to production, customer intelligence finally reaching operations. The payoff: 25% less excess inventory, 50% faster product development, double-digit gains in efficiency and forecast accuracy. None of that came from one function getting better at its job. It came from someone seeing the entire value chain and designing the connections. AI did not replace the operator. It handed the operator reach.
Can One Person See the Whole System?
The cognitive objection is substantial: one human cannot hold a complex enterprise in mind. Working memory is brutally limited; enterprise systems are not. The very abstraction layers I have spent this essay criticizing exist because total visibility became impossible at scale. If “seeing the system” means knowing every detail, I am asking for something cognition cannot deliver.
I am — which is exactly why it means something else. End-to-end visibility is pattern recognition across layers, not a complete model of every dependency. Martin’s thinkers hold the problem, not the data behind it. And this is the part the cognitive objection misses: AI lowers the cost of cross-domain access, so the operator’s edge over the silo-bound specialist gets bigger, not smaller.
Then there is the market objection: specialists still command the highest pay, so the market clearly disagrees with me. Partly true, and worth separating. Individual market value and organizational value creation are two different ledgers. Specialists command premiums on the first. But the second is bleeding — the 83% silo rate [3] and 78% data-to-decision failure [6] are system-level costs no specialist can fix from inside their own domain, because the loss sits in the seams between domains. The market is just slow to price what its org chart was built not to see.
The objection I take most seriously is the last one: AI will simply become the integration layer, making the human operator a transitional role. It might, at the margins. AI is already good at surfacing cross-domain connections. What it still fumbles is judging which connections matter under genuinely novel conditions — the situations with no training data, where someone has to decide what “healthy” even means. So my honest guess is that the role shifts upward rather than vanishing: from spotting integration problems to framing them. But I hold this loosely. Ask me again in two years.
From Individual Talent to Organizational Strategy
There is one more objection, and it is the strongest: relying on rare individuals does not scale, and organizations already built PMOs, enterprise architecture teams, and cross-functional committees precisely to handle integration.
That is a substantial objection. The evidence also raises doubts about how well the existing coordination structures are working. If structural coordination worked, 83% of companies would not still be reporting silos after decades of investment in exactly those structures. They failed where they were needed most. The encouraging part is that the alternative is buildable, because integrative thinking is learnable rather than innate [4], and systems-thinking organizations show 30% faster decisions, a 20% jump in cross-departmental efficiency, and 18% higher innovation success [5]. The market has started to notice, too — demand for AI-generalist roles grew 42% year over year [8].
So treat end-to-end operating as a capability you develop, not a unicorn you hope to recruit. The compressed version:
- Map the abstraction layers. Trace three recent consequential decisions from origin to impact and mark where the signal died between layers. Failure mode: teams map only the layers they already live in, which just redraws the silos in a new color.
- Build cross-layer rotations with real accountability. Move mid-career specialists through three layers with genuine ownership of outcomes. Failure mode: rotation as tourism — exposure without consequences teaches nothing but the location of the coffee machine.
- Redesign incentives around system outcomes. Add system-level metrics that carry real weight in promotion and pay. Failure mode: performative goals with no reward for hitting and no cost for missing.
- Deploy AI as a visibility amplifier. Point it at the seams between functions, so it flags when one layer’s decision distorts another. Failure mode: more dashboards, same paradox — the problem was always context, not data.
What to Do With This
Start where the loss is. Take the last three decisions that genuinely mattered and trace each one from where it began to where it landed. Watch for the layer where the signal degraded — where a strategic intent became a metric became a target became a behavior the strategy never asked for. That gap is where your end-to-end operators live, if you have any. It is also the first thing AI is going to make faster, for better or worse.
The deeper question this leaves me with is not about org charts. It is about what it means to understand something now that understanding has gotten cheap. For a long time comprehension was scarce, so we built institutions to ration it. We are not in that world anymore, and most of our structures have not caught up.
The countermeasure was never to get rid of abstraction. Abstraction is how anything scales; you cannot run a company by staring at every transaction. The countermeasure is to keep reconnecting abstraction to reality — to make sure someone can still read the system that lives underneath the dashboards.
The more abstract the system gets, the harder it is to fix. Unless someone can see through it.
References
- [1] Microsoft Research. New Future of Work Report 2025. 2025. (reference 9: 35% lower-skilled productivity gain vs minimal specialist gain; echoed by Rui Nunes, 2025.)
- [2] PwC. No More Pyramids: Rethinking Your Workforce for the Agentic AI Era. 2025.
- [3] Globalpraxis. The Structural Damage of Organizational Silos: A Functional Deep Dive. 2025.
- [4] Roger L. Martin. The Opposable Mind: How Successful Leaders Win Through Integrative Thinking. Harvard Business Review Press, 2007.
- [5] Mark S. Elliott. Systems Thinking and Leadership: A Strategic Framework for Organizational Success. 2025.
- [6] IT Inside Online. The Paradox of Data: Too Much Information and Too Little Vision. 2026.
- [8] Rui Nunes. The T-Shaped Professional: Why Generalists Won’t Win in the Age of AI. 2025. (42% YoY AI-generalist demand, attributed to Gartner 2025.)
- Toyota vignette: Lean Enterprise Institute; Morgan & Liker, The Toyota Product Development System; HBR, Another Look at How Toyota Integrates Product Development (1998).
- Kraft Heinz vignette: Supply Chain Dive; OMP case study.