How large language models are rewiring the architecture of organizations — not by automating tasks, but by changing the economics of coordination itself.
When electricity arrived in factories in the late 1880s, the first move was to rip out the steam shaft and bolt in an electric dynamo. Productivity barely moved — for three decades. Not because the technology didn't work, but because no one had reimagined the factory floor1.
Steam-era plants were organized vertically around one central source of mechanical power. Every machine tethered to one overhead shaft. Layout followed the constraints of power transmission, not the logic of production. Electricity freed each machine to live wherever the work demanded — but realizing those gains required a wholesale redesign: single-floor layouts, distributed motors, new contracts, new autonomy. The J-curve of general-purpose technologies2 is built from exactly this lag.
We are at the same inflection point. Deloitte's 2026 enterprise survey3 finds that only about a third of organizations are genuinely reimagining their businesses around AI — the rest are optimizing what already exists. McKinsey's parallel data4 shows AI high performers are nearly three times more likely to have fundamentally redesigned their workflows. The pattern is clear. The technology works. The redesign is where the value lives.
Every organization faces a single foundational challenge: making many minds move as one toward a shared objective. Ronald Coase5 observed in 1937 that firms exist because internal coordination is sometimes cheaper than market coordination. AI is rewriting those cost curves on both sides — but the deeper story is about what coordination, internal or otherwise, actually requires.
The overarching principle is intent alignment: every individual, team, process, and system oriented toward a coherent purpose. Many as one, one as many. But intent alignment hides a second, distinct challenge — and the two are routinely conflated.
Does everyone want the same outcome? Are goals, priorities, and definitions of success shared? Misalignment here produces departments optimizing for local metrics at the expense of the whole.
Does everyone see the world the same way? Engineering thinks the bottleneck is technical debt; sales thinks it's pricing; the CEO thinks it's brand. Same goal — different model of reality.
AI affects these differently. For intent, it serves as a tool of articulation — making mission legible, surfacing contradictions between stated priorities and actual resource flow. For understanding, AI is transformative6: it can synthesize across silos and create shared situational awareness, making it harder for different parts of the organization to operate on different versions of reality.
An organization that takes AI seriously must be understood as a coordination system with five interacting layers — and the architecture is only alive when those layers close into a feedback loop.
The intent layer defines what the organization is for: purpose, mission, strategic priorities, constraints, values, the criteria by which the organization judges its own success. It is the smallest layer and the most stable — the things that must not change even as everything else adapts.
AI imposes a discipline of clarity on intent. When strategic priorities must be encoded precisely enough to guide an agent — "optimize for retention over short-term revenue" — ambiguity that was tolerable under human-only mediation becomes intractable. Intent stops being a mission statement; it becomes a contract.
Every other layer depends on something that is rarely named: the accumulated knowledge, context, decisions, and patterns of the organization. Signal processing without memory is stimulus-response. Decision-making without memory is guessing.
Today, institutional memory lives in long-tenured heads and scattered threads no one will ever search. AI changes these economics7: it can serve as a persistent, queryable substrate connecting what the organization knows to what it is doing right now — surfacing precedent, retiring zombie projects, preserving the why behind every decision after the people who made it have gone.
What the organization is able to do — the core competencies through which it interacts with its environment.
AI makes capability formation itself faster and more distributed8. A team that lacked legal drafting, market research, or technical prototyping yesterday may have a functional version of it tomorrow — overnight, via tooling. The capability layer is no longer a fixed asset. It is plastic.
Capabilities are organs. Workflows are the metabolic pathways that cross organ boundaries — the concrete patterns by which capabilities compose into outcomes. "Process a loan application." "Respond to a competitive threat." Each threads through multiple capabilities in a specific choreography.
This is the layer where AI agents operate most naturally. Given a task, a toolkit, and a success criterion, an agent executes the workflow by composing capability invocations in sequence. But not all work is equally suited to agentic execution — and confusing the categories is the most expensive mistake in AI adoption. We return to this in the next section.
The most profound layer is the capacity to observe, evaluate, modify, and stabilize the organization's own operating patterns. Traditionally this loop runs through committees, audits, and reorgs — and takes months or years. AI compresses it dramatically. An agent that executes a workflow can simultaneously observe whether the workflow is producing good outcomes, diagnose why it isn't, and propose or implement modifications inside the same cycle.
This echoes Nelson and Winter9: organizational routines as the genes of the firm. But where biological evolution operates across generations through blind selection, AI-augmented metabolism operates within the organization's lifetime with directed mutation and conscious selection. Without selection and stabilization, mutation becomes churn. The metabolic layer needs its own immune system.
The architecture is alive only as a closed loop. Acts generate signal; signal updates memory; memory shapes capability; capability composes workflows; workflows produce outcomes; outcomes feed the metabolic layer; the metabolic layer rewrites the workflows. Intent — at the center — constrains it all.
AI's most important contribution may simply be that it compresses the cycle time of this loop. As Boyd's OODA framework demonstrated in military strategy10, the entity that cycles through observe-orient-decide-act faster will outcompete one that makes marginally better decisions but runs the loop more slowly. Speed of learning, not quality of any single decision, may be the dominant competitive variable.
The most common mistake in AI adoption: because AI can help with this, AI can run this. The more precise formulation recognizes three distinct categories of work — each requiring a different integration strategy. Drag the chips below to where you think each belongs.
Stable, repeatable, measurable. Clear inputs, defined steps, known success criteria. End-to-end automation; immediate and substantial gains.
Known goal, variable path. Human judgment at key decision points; AI handles surrounding research, drafting, analysis, coordination.
Novel, ambiguous, politically sensitive. No stable structure to automate. AI surfaces information and pressure-tests reasoning — humans run the loop.
The interesting question isn't which category each process belongs to today — it's which way each is migrating. Many processes once emergent have become adaptive as organizations learned to structure them. Many adaptive workflows are becoming codifiable as AI handles variation that previously required human judgment. Monitoring this migration is one of the most important dynamics for leadership to track.
The metabolic layer is the most powerful and the most dangerous element of the framework. An organization that can continuously rewrite its own workflows possesses a remarkable adaptive advantage. But self-modification without governance is, biologically, cancer: uncontrolled replication and mutation without the constraints that maintain systemic coherence.
Governance is not a wrapper around the five layers. It is a constitutional layer that permeates the system — answering four questions no amount of technical capability can answer on its own.
These questions are not peripheral to the AI transformation. They are the central leadership challenge. Organizations that deploy AI agents without answering them will discover the answers through failures — some recoverable, some not. Organizations that answer them in advance will move faster, not slower, because clear boundaries create freedom within them11.
Leadership shifts. The traditional role is to design structures, define processes, allocate resources, and hold people accountable for results. In an AI-native organization, that role does not disappear — but its primary task shifts from designing fixed processes to governing a dynamic system. Leaders still set intent, allocate resources, resolve trade-offs, design incentives, and define what must not change. They now do so in the context of an organization that is continuously adapting its own internal processes.
AI-native organizations will not simply automate workflows. They will develop the capacity to observe, generate, test, and selectively revise their own operating patterns — continuously, at every level. The leadership challenge is to make this self-modification aligned, accountable, and stable enough to compound rather than dissolve into chaos.
The organism metaphor is apt when used with care. A healthy organism is not one that rigidly follows a fixed program. Nor is it one that mutates without constraint. It is one that maintains homeostasis — stable core functions and identity — while continuously adapting its internal processes to a changing environment. The AI-native organization aspires to the same balance: stable purpose, dynamic execution, governed metabolism.
The organizations that thrive will be those that redesign themselves for the new wiring — not by bolting cognitive electricity onto structures built for the steam age, but by building structures worthy of the new source of power.
What changes when it does? Write for yourself; this stays on your device.