By Jacques Bughin
As AI begins to coordinate work, not just automate it, firms may evolve from static hierarchies into computationally adaptive organizations.
Artificial intelligence is often presented as a productivity tool. That is true, but incomplete. The more important shift is that AI may begin to alter the internal logic of the firm itself. For more than a century, companies have scaled by adding people, managers, reporting layers, and coordination routines. The modern corporation emerged as a solution to a basic economic problem: coordinating human activity at scale is expensive, and internal hierarchies were a practical response to that cost.[1]
If execution itself becomes a source of learning, then firms may begin to operate less like static hierarchies and more like computationally adaptive systems.
AI now introduces a different possibility. Instead of merely helping individuals complete isolated tasks, AI systems are increasingly being designed to interpret goals, formulate plans, trigger actions, supervise workflows, and learn from operational outcomes. That change is important because it moves AI from the edge of work into the coordination layer of organizations. If execution itself becomes a source of learning, then firms may begin to operate less like static hierarchies and more like computationally adaptive systems. This is what I call recursive capitalism: a system in which organizations improve their own coordination through the data and telemetry generated by their own activity. [2]
Why firms exist
The classical theory of the firm rests on coordination costs. Firms exist because some activities are cheaper to coordinate internally than through markets. But as firms grow, the costs of internal coordination rise too. More scale means more information flows, more approvals, more conflict resolution, more governance, and more managerial layering. Industrial capitalism therefore evolved around a specific organizational solution: combine labor specialization with hierarchical coordination. [3]
That model worked because humans were the primary coordination layer. Managers interpreted ambiguity, synchronized departments, allocated authority, routed workflows, and resolved exceptions. Even after the digital transformation, that basic structure remained intact. ERP systems digitized operations, CRM systems structured customer interactions, cloud computing centralized infrastructure, and internet platforms accelerated transactions and communication. Yet the organization still depended on people to translate information into coordination
This is why AI matters. It does not merely digitize existing routines. In some contexts, it begins to participate in coordination itself. That is a qualitatively different development from earlier automation waves. The steam engine, electrification, and enterprise software transformed production and information processing, but they did not directly assume the role of coordinating the firm. AI increasingly can [4].
From software to learning systems
The key concept is recursion. In older software systems, usage generated information. In recursive AI systems, execution generates organizational learning. Every workflow, approval, exception, and interaction leaves a trace that can be used to improve future execution. Over time, the firm accumulates not only data but operational memory embedded in its coordination systems. [5]
That is why infrastructure matters so much. Databricks and Snowflake are increasingly central not just because they store and analyze information, but because they help firms manage metadata, governance, lineage, and semantic consistency. Palantir has pushed this logic further by connecting workflows, permissions, operational logic, and decision structures into executable environments. These systems are best understood as attempts to encode organizational memory into the operating layer of the firm.
NVIDIA’s “AI factory” framing adds another dimension. It suggests that computation is no longer only a support function or even a productivity layer. It is becoming part of the enterprise production substrate itself. ServiceNow and Salesforce are moving in the same direction by embedding AI into workflow orchestration, customer operations, and internal resolution systems. In each case, the strategic shift is from point solutions to coordination architectures [6].
That is the real significance of recursive capitalism: the organization does work, the work generates telemetry, the telemetry improves the system, and the improved system changes how the organization works next time.
What evidence already shows
There is now enough evidence to argue that this shift is not just theoretical. In software development, AI-native environments such as Cursor are compressing development cycles by allowing engineers to supervise multiple coding agents. The role of the developer is changing from manual producer to orchestrator and reviewer. This is significant because software engineering is one of the most coordination-intensive activities in the knowledge economy [7].
The same pattern is visible in enterprise operations. ServiceNow is positioning itself as an AI control center for enterprises, with tools that benchmark, govern, and orchestrate agents across workflows. Its partnership with NVIDIA is aimed at combining reasoning models, infrastructure, and governance so that AI can be deployed at scale with more transparency and control. Salesforce’s Agentforce similarly reflects a move from assistant-like AI toward autonomous workflow execution
Academic research points in the same direction. Work on dynamic capabilities argues that competitive advantage depends on a firm’s ability to sense, seize, and reconfigure in response to change. That framework becomes even more relevant when AI is embedded in the operating model, because the organization’s ability to learn and adapt becomes part of the competitive asset itself
Studies on technological frames and ambidexterity also show that AI adoption depends heavily on how managers interpret and organize technology, not just on the technology’s technical features.
The OECD’s recent work on AI adoption in firms reinforces this point. Adoption is shaped by infrastructure, skills, governance, and integration capacity. In other words, the limit is often organizational rather than technological. Firms do not simply buy AI; they redesign operating routines, decision rights, and control systems to make AI usable. [2]
How competition changes
If AI changes coordination, it changes competition. Historically, firms competed through scale, labor efficiency, manufacturing capability, distribution reach, and information advantages. In recursive capitalism, firms may increasingly compete through learning velocity, orchestration quality, semantic consistency, and operational memory. [7]
This has two effects. First, the half-life of competitive advantage may shorten. If execution itself becomes a mechanism for improvement, firms that learn faster can adapt faster. Second, the distribution of advantage may become more complex. AI may centralize intelligence while decentralizing execution. Smaller AI-augmented firms can punch above their weight because coordination costs fall, while larger firms may accumulate deeper telemetry and richer learning systems because they execute more workflows.
That is why the emerging AI economy does not look like a simple winner-take-all market. It looks more like a federated system of coordination powers. NVIDIA dominates compute infrastructure, hyperscalers control cloud execution, frontier model providers shape cognition, data platforms govern memory and semantics, and workflow platforms orchestrate enterprise action. These firms increasingly compete and cooperate at the same time
For enterprise leaders, the strategic implication is clear. Dependence on a single platform is no longer just a procurement issue. It is a governance issue and, increasingly, a sovereignty issue. If a firm centralizes workflow authority, operational memory, and semantic interpretation inside one ecosystem, it may create new forms of strategic dependence embedded directly in the architecture of work
What managers should do
The managerial response should not be blind adoption. It should be organizational redesign. Executives should ask four questions.
First, which coordination tasks should remain human, and which can be delegated to agentic systems? AI can already support many managerial routines, but not all tasks should be automated. Human judgment still matters in ambiguity, escalation, and accountability.
Second, where does organizational learning reside? If a firm cannot capture workflow traces, exception patterns, and resolution histories, it will not build operational memory. AI without memory is simply faster software
Third, how exposed is the firm to single-platform dependence? Interoperability, portability, and governance should now be treated as strategic design principles, not technical preferences
Fourth, how should managerial roles evolve? Leaders should spend less time supervising routine execution and more time designing the systems that will coordinate execution over time. In recursive organizations, management becomes an architectural function. [8]
This does not reduce the importance of leadership. It increases it. As organizations become more computationally adaptive, the quality of judgment, trust, and institutional coherence matters more, not less. AI can scale good decisions quickly, but it can also scale flawed coordination just as fast. The role of management is therefore to ensure that the organization learns the right lessons from its own execution.
Conclusion
It is still early, and the constraints remain real. Reliability, cybersecurity, integration complexity, regulation, trust, and liability continue to shape enterprise AI adoption. Most firms will move incrementally rather than transform overnight.
AI can scale good decisions quickly, but it can also scale flawed coordination just as fast.
But the broader direction is increasingly visible. Industrial capitalism optimized labor. Digital capitalism optimized information. Recursive AI capitalism may optimize adaptive coordination. If that happens, the most successful firms will not simply be those with the largest workforces or the broadest software stacks. They will be the organizations that learn operationally faster than their competitors while preserving enough trust, interoperability, and resilience to sustain performance over time. That would not be just another technology cycle. It would be a new organizational architecture for capitalism itself.
Jacques Bughin
