After nearly half a year of my own hands-on practice, I am more and more certain of one thing: in traditional enterprises, AI deployment is in most cases not a technical problem, but a structural impossibility.

The technology has been good enough for a long time. What actually blocks it is the organization’s operating system — power structure, incentives, feedback loops — which is fundamentally incompatible with the way AI demands to run. Without changing that operating system, AI stays stuck at pilot stage forever and never gets to scale.

The core difficulties I have observed

1. Leadership treats AI as a tool upgrade, not an organizational rebuild
The vast majority of enterprises understand AI as “one more piece of software.” They stand up a few scattered point applications and leave processes, responsibility boundaries, and performance logic untouched. The few that actually make it share exactly one thing — they rewrite the end-to-end workflow first, then put AI on top. Do it the other way around and it almost always dies.

2. Power and transparency are inherently in conflict
AI makes decision processes that used to run on experience and relationships visible and traceable. Middle managers feel their professional authority being eroded; senior leaders feel data starting to take the microphone away. This is not “resistance to new technology,” it is rational self-defense of power. Once AI really changes who gets to decide, the resistance turns from “let’s study it a bit more” into systemic foot-dragging.

3. Incentive systems punish the high performers who drive change
The people who can actually land AI are usually the ones who iterate fast, coordinate across departments, and break vague problems down into executable tasks. But in a traditional structure, those people are the first to be pushed to the margins. Because they expose the inefficiency of the existing system. The result: the organizations that need AI most drive out the people most capable of deploying it.

4. Legacy processes are completely out of sync with AI’s iteration tempo
AI needs fast trial and error and a closed data loop. The traditional enterprise process is approval chains, departmental walls, one-shot delivery. The two differ in speed by an order of magnitude. Stuff AI into the old process and it instantly becomes “one more form to fill out.”

5. Data and system integration is a surface symptom
Plenty of AI projects die on “the data isn’t connected.” On the surface it looks like a data problem. At bottom, the organization never treated connecting data across departments as a real priority — because connecting it means exposing accountability and performance.

How East Asian power structures amplify this into something structural

Inside East Asian companies — especially those shaped by high power distance culture — the difficulties above get pushed even further into structure.

High power distance is the default setting: the hierarchy itself is accepted as natural order, and the cost for a subordinate to question a superior is extremely high. Face and harmony come before clarifying facts — if data directly exposes a flaw in a superior’s decision, it is not “useful feedback,” it is a public challenge to authority. Relationship networks often carry more real force than formal process, and information is deliberately controlled, because controlling information is itself the way to hold position.

AI demands exactly the opposite: lateral data flow, low-latency feedback, using results to challenge existing judgment, failure that has to be visible. Inside a high power distance structure, these behaviors are not “a bit awkward” — they directly threaten the core stability of the system. The result is that any move trying to shift decision rights from “people plus relationships” to “data plus explicit goals” triggers a strong immune response — not open opposition, but sustained foot-dragging, dilution, and marginalization.

I have seen it over and over in practice: the moment you try to break the existing hierarchical reporting path with a unified data view, the resistance escalates from a technical problem into an organizational-politics problem. Decentralized nodes — every department running its own agents, managing its own data — are especially lethal under this structure. They do not just fragment the data further; they also give every node a chance to protect its existing power with “our side is special.”

So I argue it more firmly: every workflow inside a company should be extremely centralized. Supernodes are far better than decentralized nodes. This is directly related to data availability.

Decentralization sounds flexible. In practice, under East Asian power structures, it degrades fast into data islands and passing the buck. Without a unified supernode pressing critical execution and data together, however strong the model is, it can only spin inside local blind spots. Only when execution and data availability are compressed back onto a small number of supernodes does the feedback loop actually close and AI get a chance to compound. This is not a preference, it is a physical requirement for guaranteeing data availability and a closed decision loop in a high power distance environment.

The only way out

Not a bigger model. Not another round of training.

Either you confine AI to edge scenarios that do not touch core power and process, and accept that it will never become an organizational capability;

or you change the operating system first — rewrite incentives, tear down the information silos, build a real feedback loop, tilt decision rights toward data capability — and then put AI on top.

The second path is brutally painful under East Asian power structures, because what it changes is exactly what that structure depends on to hold itself together. Most organizations are genetically incapable of it. This is a pessimistic structural reality.

AI will not transform the enterprise on its own. Either the enterprise actively transforms itself to fit AI, or it keeps treating AI as decoration — and then gets eliminated in the next cycle by rivals who actually finished the rebuild. Survival of the fittest.

So maybe FDE itself is a fake problem. Rather than polishing a pile of legacy, tear it down and start over.