The Org Chart May Be Legacy Infrastructure for the AI-Native Organization

The AI-native organization may require us to stop treating the org chart as the natural architecture of work. Org charts, departments and stable roles solved real coordination problems in a world where cognition was scarce, expertise was difficult to access and context had to move through people. AI changes enough of those constraints that preserving the same structure may eventually look like designing roads for horses after inventing the car.

That does not make the road stupid. It makes it historical.

A road built for horses reflects the limits of horses. Width, turning radius, staging, speed and distance all make sense inside that system. Put an automobile onto the same road and the car still creates value. It moves faster and carries more. But at some point the infrastructure becomes the constraint.

I think the same possibility now exists for the organization itself.

Most enterprise AI programs still begin by identifying roles, functions and workflows, then asking where AI can assist them. That is a sensible first step. It is also potentially the point where we smuggle yesterday’s assumptions into tomorrow’s operating model.

The org chart is not only a map of authority. It is also a map of where cognition has historically been forced to live.

The org chart solves human constraints

Modern organizations did not divide themselves into Finance, Legal, Marketing, Technology, Operations and Product because somebody enjoyed drawing boxes. Functional specialization is a rational response to limits.

People have bounded expertise. We have bounded memory. Attention is scarce. Communication takes time. Physical and temporal separation make coordination expensive. Specialist knowledge takes years to develop. Managers can only maintain so many relationships before span of control becomes unwieldy.

As organizations grow, they therefore distribute cognition. Finance accumulates financial expertise. Legal accumulates legal interpretation. Marketing accumulates customer and communication expertise. Technology accumulates technical capability. Management creates mechanisms for allocating work, reconciling priorities and moving decisions between these specialized nodes.

Viewed this way, the org chart is partly a context-routing mechanism. A problem enters the organization, somebody decides who owns it, the relevant function gathers context, work happens, and the result is handed to the next part of the system.

That routing logic is so familiar that we often confuse it with the work itself.

But authority and information routing are not the same thing. Accountability still needs a home. Legal responsibility still needs a home. Budget authority, permissions and risk ownership still need a home. The fact that those controls matter does not mean every piece of knowledge must continue travelling through the same boxes.

This distinction matters because AI is beginning to change the economics underneath specialization. Recent research on knowledge work already shows that generative AI is changing individual work patterns faster than coordinated organizational behaviours. A large Microsoft Research field experiment found substantial changes in independently adjustable work such as email, while meeting time changed far less. That is an early signal of the structural problem: individual cognition can change before organizational coordination does.

AI changes the useful unit of organization

At an AWS event I attended earlier this year, Dr. Ksenia Wahler made a point that stayed with me. She described AI as enabling people to scale vertically rather than only organizations scaling horizontally by adding more specialist teams. The practical implication is not that every employee suddenly becomes an expert in everything. It is that one individual can acquire something closer to polymath-like reach.

A product leader can explore a financial implication without waiting for a full finance cycle. A strategist can interrogate technical architecture more deeply before pulling engineers into the room. A marketing leader can challenge assumptions with market research, scenario analysis and customer-language synthesis without turning every question into a separate workstream.

The relevant economic change is that the cost of temporarily accessing competent specialist cognition falls.

This is consistent with broader evidence about where generative AI is useful. Microsoft Research’s analysis of 200,000 anonymized Copilot conversations found AI being used across information gathering, writing, teaching and advising activities. These are not narrow software tasks. They are cognitive activities that sit inside many different professions.

That starts to undermine an assumption hidden inside the traditional role: capability is bundled relatively statically around the person who holds the job.

If relevant machine capability can instead be assembled dynamically around an objective, then the useful unit may no longer be the employee occupying a narrowly bounded role. It may increasingly be a human judgment node surrounded by dynamically assembled machine capabilities and shared context.

The employee becoming more polymath-like and the organization becoming less functionally segmented are therefore not two separate predictions. They may be the same change viewed at different scales.

At the individual level, cognitive reach expands. At the organizational level, the need to route every problem through stable specialist structures starts to weaken.

AI does not make expertise irrelevant. It makes the location of expertise less fixed.

The first AI-native mistake is to create synthetic departments

I ran into a tiny version of this problem in my own AI operating system.

My first serious setup used specialist GPTs with explicit responsibilities. That was useful because the names and roles gave me understandable handles for complex capabilities. I have written before about why I gave those AI agents ridiculous names and why personification initially improved governance.

Then the structure began producing a strange new management problem. I knew what I wanted, but I was increasingly unsure who should own it. Was an editorial question for the content strategist or the writer? If it affected growth, should the distribution specialist be involved? If the article needed reusable website components, did the work now belong to website development?

The specialists were working. The boundaries were becoming overhead.

That is the point where my earlier argument that AI agents are not digital employees became more concrete. Human-shaped roles are useful scaffolding, but they become a design constraint if every new capability has to be assigned a fictional job, given a territory and handed context through an artificial organizational boundary.

My personal system is not evidence that enterprises should copy the architecture. It is useful because the failure mode appeared at absurdly small scale. If one person can create coordination overhead among a handful of synthetic specialists, a large company can reproduce silos, handoffs and ownership ambiguity at industrial scale.

That is why the emerging enterprise conversation about agentic operating models matters. BCG argues that many organizations are inserting AI into operating models built for a purely human workforce, while McKinsey similarly observes that AI leaders are rewiring how work and decisions happen rather than simply accelerating existing activities. The implication is bigger than automation: the organization itself becomes a design variable.

If the department stops being the context container, what replaces it?

The most important architectural shift in my own experiment was not replacing one set of AI specialists with another. It was separating durable context from the capability acting on it.

I stopped wanting critical knowledge to live primarily inside a conversation or a specialist. Strategies, objectives, evidence, relationships, publishing standards, venture hypotheses and execution references increasingly moved into Superhuman Docs (formerly known as Coda) as canonical state. Skills could then move across that state depending on the objective.

The practical rule became simple:

The context should stop moving. The capability should move to the context.

That principle is now making even some of my carefully separated ChatGPT threads less architecturally important. I still use different threads for focus. But a content-strategy conversation can call a website-development capability when it discovers that the argument depends on reusable components. A venture conversation can retrieve an existing relationship without that relationship being copied into a separate venture memory.

The thread becomes a working surface. The knowledge underneath it becomes durable.

Enterprises already have many ingredients for this: systems of record, stable identifiers, data models, policies, product catalogs, customer records, process definitions, decision logs and access controls. The problem is often that these are organized as repositories rather than a reusable context layer around objectives.

This is one reason I think knowledge graphs deserve more executive attention than the name tends to invite.

Knowledge graphs are less exotic than they sound

A knowledge graph is fundamentally about entities and the relationships between them. IBM describes knowledge graphs as networks of real-world entities connected by relationships. At the formal end of the spectrum, standards such as W3C’s RDF model represent knowledge through subject-predicate-object relationships.

Enterprise implementations can become technically sophisticated very quickly. Ontologies, entity resolution, provenance, permissions, freshness, integration and governance are real problems. I would not describe a collection of linked Superhuman Docs / Coda tables as a formal enterprise knowledge graph.

But the useful conceptual starting point is much simpler: information becomes more powerful when the system can understand not only documents, but relationships.

A person is related to an organization. The organization may be related to an opportunity. The opportunity supports an objective. A piece of evidence may strengthen or weaken a hypothesis. An article may connect to an editorial theme, a relationship and an operating-model concept. A decision may supersede an earlier rule.

Once those relationships are explicit, a capability does not need somebody to manually assemble a folder of documents every time a new objective appears. It can traverse the context that matters.

This is also changing how I think about human-readable and machine-readable organizational knowledge. I increasingly brief ChatGPT on which Coda structures need to be pleasant for me to read and which can simply be efficient for the AI to work with. The machine does not care how far it has to scroll right to find the field it needs. That sounds trivial. It is not.

Most enterprise knowledge systems are designed almost entirely around human navigation. AI creates an opportunity to separate the canonical structure of knowledge from the interface through which different humans and capabilities consume it.

This is not a proposal to replace every enterprise system with a graph database. It is a proposal to stop assuming that the department must remain the primary unit through which context is routed.

Management in the AI-Native Organization Becomes More Important When Routing Becomes Less Important

The easiest way to misunderstand this argument is to turn it into another prediction that AI will eliminate managers.

I think the more interesting outcome is the opposite: management becomes more concentrated around the things that actually require management.

Traditional managers spend substantial effort coordinating scarce human capability. Who owns the work? Who needs to be involved? Who has enough context? Which specialist has capacity? Where does the decision escalate? Which team is waiting on which other team? Should I break down my department into teams because the overhead of managing more than seven people is too big?

If AI reduces some of that routing burden, the managerial center of gravity can shift toward objectives, constraints, permissions, incentives, evidence standards, decision rights and accountability.

That is not less management. It is different management.

The same distinction appeared in my earlier work on AI decision rights. More execution capacity does not remove the need to decide where authority belongs. It makes that question harder because execution can scale faster than meaningful review.

McKinsey’s current work on AI operating models reaches a similar conclusion from a broader enterprise perspective. Its July 2026 research argues that the advantage comes from redesigning how work is done and how decisions are made, not merely from deploying tools. That is an important distinction because operating models coordinate workflows, governance, people, data and decision-making. AI can make some coordination cheaper while making clear governance more consequential.

The system remains socio-technical. Permissions still matter. Incentives still matter. Human behaviour still matters. Legal and reputational accountability still matter. The fact that a capability can cross departmental boundaries does not mean it should cross every permission boundary.

AI may reduce the need to manage the route work takes without reducing the need to manage what the work is allowed to do.

The org chart may become legacy infrastructure before the organization becomes obsolete

I do not expect organizations to disappear. I do expect us to become much more precise about which parts of organizational structure exist because humans need them and which parts exist because the work itself needs them.

That distinction is uncomfortable because many structures serve both purposes at once.

A department can be a useful community of practice and a poor routing mechanism. A management layer can provide essential accountability and unnecessary information transfer. A role can create healthy ownership and artificially narrow the capability available to the person occupying it.

The AI-native organization therefore should not begin by flattening everything. It should begin by asking what each boundary is doing.

Does this boundary protect accountability? Preserve expertise? Manage risk? Create identity and development? Or is it simply the route through which information has historically had to travel because the people on either side could not hold the same context at the same time?

If the answer is mostly the latter, AI changes the design space.

This is why I think the comparison with the horse road is more than a metaphor. Cars did not make destinations irrelevant. They changed what infrastructure made sense between them. They changed distance, settlement patterns, supply chains, retail, commuting and eventually the shape of cities.

AI may do something similar to cognitive infrastructure.

The immediate productivity gain is the car moving faster on the old road. The deeper transformation begins when we redesign the road system around what the car makes possible.

For leaders, that creates a question considerably more demanding than where to deploy the next agent:

If you could redesign the flow of work around the objective rather than the people available to perform it, how much of your current organization would you deliberately put back?

That is the question I would put in front of an executive team before it invests heavily in synthetic employees, agent departments or elaborate AI handoffs.

Because if AI-native work still looks primarily like humans handing tasks between departments, with digital workers now standing inside the boxes, we may have automated the organization we should have redesigned.