Department walls are not a culture problem; they are a structure problem: AI is acting as a solvent

Related reading: When AI flattens the organization, every layer has to remake the meaning of its role (the AI Flattening hub)

Introduction|AI is a solvent of structure, not another screw

In most companies, our slogan is the same, the logo is the same, and the everyday is like a set of islands: marketing does not understand the rhythm of sales; sales mocks brand media spend; brand and the supply chain look at each other through a fog; customer service stands at the last mile, fighting fires after the fact. Every island has its own language, KPIs, tools, and rhythm, and can only be bridged by a cross-department meeting, translated by a PM, and barely held together by a deck and an email. Generative AI and multi-agent systems are rewriting all of this: it is not installing one more system, but letting a task, centered on intent, pass through the original human-made boundary, and pushing the enterprise from a “system of divided subjects” toward a true network of collaboration. This piece answers, from mechanism, path, governance, to a plan for landing, a sharp proposition — why will department walls disappear? how, after they disappear, do we not fall into disorder? and how should the next generation of a “wall-less organization” be designed?

NO° 01|Why department walls were born: a by-product of industrial management

The walls of a modern company do not come from malice. They are a necessity of the logic of efficiency. For standardization, a process is cut finely; for specialization, each department cultivates its own cut; for performance, a KPI becomes an independent unit of accounting. This architecture brings stability and controllability, and also a high cost of transaction:

First, a break in information. The same fact has a different standard in different systems: sales looks at the CRM, operations looks at the OMS, the brand looks at a campaign report, customer service holds the FAQ and the tickets. Data is duplicated, the languages do not communicate, the tools are split; any cross-department collaboration is first bound hand and foot by “unifying the standard.”

Second, a lag in decision. A linear process depends on the completion of the previous node; a meeting becomes a conveyor of authority and duty: an intent has to pass layer after layer of sign-off before an action is allowed.

Third, a fragmentation of the customer experience. The customer’s interaction with the brand is cut into “a department’s point of view”: an advertising promise, the in-store experience, and the after-sales policy fight one another, and in the end only customer service is left to tidy up for everyone.

When AI enters, what it first attacks is precisely these rigid structures. Because, for AI, a task is not “handed to a certain department,” but an intent broken into capability nodes that can be dispatched, and collaboration realized with semantics and data as a common interface.

NO° 02|Three ways AI breaks department walls: intent at the center × process modeling × semantic collaboration

1. From “duty” to “intent”: the dispatch of a multimodal agent
AI’s entrance is not “the brand department” or “the content department.” It is the user’s intent. When someone says “I want a skincare plan that fits the change of season,” the intent will trigger a string of cross-domain dispatch: a product database compares skin type and ingredient restrictions; an inventory system checks what can be sold and the expiry; a content engine generates an explanation and a situational image; an operations module calculates a combination of offers and a delivery schedule; a customer-service strategy agent prepares a seven-day follow-up script. There is no “handing the matter to someone.” The task lands automatically on a net of capabilities. At this moment, a department is only a unit of management. Intent is the unit of collaboration.

2. Turn a process into “a work model that can be trained”: a self-learning process net
In the past, a process was written in a Visio diagram, and an update relied on a notice. Under an AI framework, a process is a model that can be measured, A/B’d, and regression-tested: which steps can be automated? which approvals overlap and are redundant? which exceptions are replayed every week? Every node carries data and metrics (latency, error rate, interception rate, first-success rate); iteration is like a neural net “learning the shortest path,” and the process begins to have intelligence.

3. From a data island to a RAG consensus: the integration of semantic knowledge
Knowledge used to be scattered across systems: sales understands the CRM, customer service is fluent in the FAQ, product buries the PRD, marketing guards the campaign. Use RAG and a vector index to open the semantic layer of each system: documents, tables, images, and recordings are sliced, tagged, and given provenance; permission is controlled by a ticket; a citation is automatically attached when it is called. For the first time, a “shared language” across departments is built: it is not a physical concentration of data, but semantics that can be retrieved. When AI can answer “on the basis of which policy, which clause, last updated by whom,” collaboration no longer ends in an argument. It converges on evidence.

NO° 03|Without “departments,” the organization will not fall into disorder; it will be more efficient: from a pyramid to a neural net

“De-departmentalization” is not anarchy. It is an upgrade of the way of management. The old logic emphasized a division of functions, a process orientation, a closed loop of KPIs, and reporting from the top down. The new logic turns to the orchestration of intent, being driven by a task, ecological feedback, and a co-construction of person and machine. The organization is more like a neural net: in the middle is an intelligent hub (an entrance of intent, a RAG corpus, agent scheduling, a measurement panel); around it are capability nodes (a person or an agent); at the periphery is a layer of a task cloud (all kinds of need that start in real time).

On this net, each person no longer corresponds only to one “post,” but to several capability vectors that can be dispatched: today, take part in an intent chain of a new-product launch; tomorrow, support an intent chain of a customer-service peak. You collaborate in parallel with several agents, and, in different task chains, play the role of an evaluator, an adjudicator, or a repairer of exceptions. A mixed mechanism of person and machine becomes the everyday: AI takes the throughput of a large volume of repeated steps; a human is responsible for the definition of the problem, the standard of judgment, and the exception.

To avoid “no wall meaning no order,” the key is three things:

  • Observability — every intent chain has a log, metrics, and a version;
  • A clear boundary — when HITL (Human-in-the-Loop), when HOTL (Human-on-the-Loop);
  • Responsibility that can be traced — citation, decision, and authorization all have provenance and audit. When evidence and rhythm can be seen, speed and order can hold at the same time.

NO° 04|A management paradigm that turns toward an “ecology of intent”: a map, a pool of capabilities, and a mechanism of trust

(1) Replace a post process map with an “intent map”
Take inventory of the enterprise’s high-frequency tasks and of the user’s trigger points (a launch, a promotion, after-sales, a complaint, a stock-out, a recall, a cross-border matter), and settle each type of task into a prompt template + an agent path: the goal and the conditions of success (audience, tone, KPI); a data contract (source, field standard, a permission ticket, a version); primitives of action (retrieve, compare, generate, test, export); a division of labor between person and machine (when a person takes over, the rule of escalation); acceptance and write-back (measurement, citation, a log). A process map is for a person to look at. An intent chain is for the system to execute.

(2) Replace the persona of a post with a “pool of capability vectors”
HR and data governance join hands, and build a capability card for a person and for an agent: the tasks they are good at, the data they know, the tools they can use, the boundary of permission, a memo of risk, and they can be dispatched in real time by an intent chain. Evaluation moves from “annual performance” to contribution to a task chain: first-success rate, interception rate, improvement of latency, coverage of citation, quality of write-back. The flow of talent no longer depends on a department’s approval, but on the pull of a task; an individual therefore has a “multi-role identity.”

(3) Build a “mechanism of trust and transparency”
Trust no longer depends on a shout. It depends on a mechanism. Three basic pieces of work:

  • A grading of permissions: who can see, who can use, who can change. For sensitive fields (personal data, cash flow, medical), forced anonymization and the principle of the minimum.
  • The escalation of exceptions: clearly mark the conditions of “a person must,” and the person who signs off; AI knows how to “throw the ball back to a person.”
  • Observability: RGR (Retrieval Grounding Rate), HIR (human interception rate), IOL (intent-to-outcome latency), CSR (compliance-slippage rate), FCR (first-contact resolution), Autonomy Score (the share of multi-step completion without a person). Let “used a lot” recede, and “used rightly” take the lead.

(4) An update of organizational roles: let the new language have someone to speak it, and someone to be responsible
In a “wall-less” scene, several key roles are indispensable:

  • Intent Architect: turns a business problem into an executable intent chain and a standard of judgment.
  • Ontology Steward: defines the semantics of a field, the standard, and the data contract.
  • Agent Orchestrator: operates the agent team, permissions, and versions.
  • AgentOps / MLOps: watches logs, measurement, regression tests, and recovery from a fault.
  • HITL Guardians: designate the gates that must be decided by a person, and the audit.

Without a clear attribution of responsibility, even a good model will only stop in the demo room.

NO° 05|Closing: departments will fade; collaboration will have more life

AI has not come to destroy the organization. It has come to release energy long suppressed by department walls. When dispatch takes intent as its center, a process can be trained, and knowledge is connected by semantics, the sense of a “department”’s existence will gradually leave the stage, and what takes its place is: a clearer convening of a task (the problem is defined so that it can be executed), a more agile calling of resources (a person and an agent used as needed), a more transparent attribution of responsibility (evidence and measurement leave a trail), and a more resilient collaboration of person and machine (speed and order coexist). The organization will be regenerated from a closed pyramid into a flowing, observable ecology of collaboration — not only smarter, but also more alive.

More deep-dive pieces

The judgment, restated

My judgment is that department walls are not a culture problem; they are a structure problem. AI is a solvent — if you do not change the interface, you will only make the noise louder.

Three things this week

  • Draw one cross-department intent interface (who asks, who answers, and where the data comes from).
  • Set a shared semantic layer: a key standard, a status code, and a boundary of responsibility.
  • Pick one cross-wall agent pilot, and force a write-back to both sides’ systems.

For a deeper rewrite of the layers, see When AI flattens the organization, every layer has to remake the meaning of its role (the AI Flattening hub). You are welcome to subscribe to Future Insights Lab, or to reply and discuss the sticking points of your department walls.

訂閱 Future Insights Lab

不定期寄送 AI × Business 短篇筆記,協助你掌握實務重點。隨時可取消訂閱。