The rapid proliferation of agentic artificial intelligence—systems designed to not only generate content but to execute multi-step tasks autonomously—has sparked a global debate regarding the future of human cognitive agency. As AI vendors and consultants forecast an era where autonomous agents handle everything from software development to legal analysis, a growing cohort of academics and industry analysts are warning of a structural "con" reminiscent of the management consulting industry’s historical trajectory. The concern is that the current trajectory of AI adoption may lead to a permanent "unlearning by not doing," where individuals and organizations forfeit the very skills required to evaluate the technology they have become dependent upon.
The Structural Parallel: Lessons from The Big Con
The current AI landscape bears a striking resemblance to the dynamics described by economists Mariana Mazzucato and Rosie Collington in their 2023 work, The Big Con. Their analysis suggests that the management consulting industry often extracts value in excess of what it creates by fostering a cycle of dependency. Organizations that outsource strategic functions eventually lose the internal expertise necessary to judge whether the advice they receive is sound. This "information asymmetry" forces continued engagement with external firms, regardless of the quality of their output.
AI vendors are currently following a similar playbook. By offering subsidized pricing and seamless integration, they have compressed what would normally be a decade of institutional adoption into less than four years. Since the launch of ChatGPT in late 2022, the narrative has shifted from AI as a "copilot" to AI as an "agent." This shift implies a transition from human-led work to system-led execution. As organizations integrate these agents into core functions—such as financial auditing or contract drafting—they risk shedding the tacit, institutional knowledge required to spot errors or biases in the AI’s probabilistic outputs.
A Chronology of the Agentic Shift (2022–2026)
The transition from generative tools to autonomous agents has moved through several distinct phases, marked by both technological leaps and regulatory friction.
- November 2022: The public release of ChatGPT initiates the "Generative Era," focusing on text and code assistance.
- 2023–2024: Major consultancies, including McKinsey and BCG, launch massive AI advisory wings. McKinsey forecasts that generative AI could add $2.6 trillion to $4.4 trillion annually to the global economy, driving a wave of corporate "AI-first" transformations.
- 2025: The "Agentic Pivot" occurs. Software-as-a-Service (SaaS) providers begin replacing human-in-the-loop workflows with autonomous agents capable of "closing the books" and managing supply chains without direct oversight.
- May 2026: Academic studies and professional reports begin highlighting "cognitive debt." MIT professor Micah Nathan publishes a widely cited account of students surrendering creative authorship to AI, sparking a debate on the "atrophy of endurance" in human cognition.
- June 2026: Geopolitical tensions manifest in the AI sector. The U.S. government orders Anthropic to suspend access to its "Fable 5" and "Mythos 5" models for foreign nationals, highlighting the vulnerability of nations dependent on a single provider’s infrastructure.
Data and Economic Realities: The Cost of Overdependence
Despite the promise of cost savings, the economic reality of AI adoption has proven more complex than initial forecasts suggested. In 2026, several early adopters reported significant budgetary strain due to the "compute-to-labor" cost ratio.
Internal data from Microsoft and Uber in early 2026 revealed that AI coding licenses were being cut or restricted after annual budgets were exhausted in less than six months. An Nvidia executive recently acknowledged that in high-stakes environments, the cost of the compute required for agentic AI can exceed the cost of the employees it was meant to replace.
Furthermore, a 2026 survey of 2,500 global firms found a concerning "value gap." For every dollar spent on AI tokens, only 18 cents generated direct user-facing value. Approximately 44 cents of every dollar were spent on "debugging" or fixing errors introduced by the AI systems themselves. This phenomenon has led to "tokenmaxxing," where employees inflate AI usage metrics to meet management KPIs, despite decreasing actual productivity.
The Human Cost: Cognitive Debt and Atrophy
The individual impact of delegating cognitive tasks to AI is becoming increasingly measurable. Recent randomized controlled trials have introduced the term "cognitive debt" to describe the long-term effects of AI reliance. These studies found:
- Impaired Independent Performance: Participants who used AI for executive functions performed significantly worse when later asked to perform the same tasks unaided compared to a control group.
- Neural Connectivity: Researchers observed lower neural connectivity in areas associated with sustained attention and original thought among frequent AI users.
- Hedonic Adaptation: As ready-made answers become the default, the "cost" of independent thinking feels higher to the individual, leading to a psychological resistance to unassisted work.
Professor Micah Nathan’s observations in The Guardian (May 2026) underscored this, noting that writing and coding are not just about production but are "training for endurance by way of sustained attention." When the struggle of creation is removed, the capacity for critical judgment often follows.
Geopolitical Risks and Official Responses
The societal cost of AI dependence has taken on a geopolitical dimension. With the United States and China controlling nearly 90% of global compute capacity, other nations find themselves in a position of "digital subservience."
The June 2026 U.S. government directive to Anthropic served as a wake-up call for international partners. By ordering the suspension of model access for foreign nationals based on national security concerns, the U.S. demonstrated that frontier AI is not a neutral utility but a strategic asset that can be withdrawn.
Singaporean parliamentarian Kenneth Tiong reacted to these developments by stating that building a national AI strategy on foreign assumptions is a fundamental risk to sovereignty. In response, several European and Asian nations have begun pivoting toward "appropriate technology" frameworks, emphasizing open-weight models and localized compute infrastructure that can function independently of a single vendor’s cloud.
The Oversight Tax and Accountability Sinks
Organizations are also grappling with the "oversight tax." To satisfy legal and liability requirements, companies often keep a "human in the loop" (HITL). However, research into medical AI screening suggests a "safety-net effect," where human reviewers become less diligent because they assume the AI has already caught major errors.
This creates what analyst Dan Davies calls an "accountability sink." The human reviewer is often not there to improve the output but to absorb the blame if the system fails. Because the human reviewer has often lost the deep technical understanding of the task (due to lack of practice), their "sign-off" becomes a hollow formality rather than a meaningful check.
Reclaiming Agency: The Third Path
To counter the risks of overdependence, experts suggest a "Third Path" that avoids both outright bans and uncritical adoption. This framework requires action at three levels:
Individual Level: Deliberate Resistance
Users are encouraged to maintain the "productive struggle." Research suggests that workers who frequently modify and challenge AI outputs retain higher levels of independent reasoning. The goal is to remain a "centaur"—someone who uses the tool to enhance their own agency—rather than a "reverse centaur" who merely serves the AI’s prompts.
Organizational Level: Structural Sovereignty
Companies are advised to treat institutional memory as a strategic asset. This includes:
- Vendor Diversification: Avoiding single-vendor lock-in for critical functions.
- Mandatory Rotations: Ensuring staff spend time performing AI-assisted tasks manually to preserve internal expertise.
- Small Model Integration: Using locally-hosted, open-weight models for routine tasks to maintain operational continuity if frontier model access is interrupted.
Societal Level: Regulatory Explainability
Governments are moving toward legal standards for AI in public service. New proposals suggest that any human "in the loop" for judicial or welfare decisions must be able to explain the AI’s reasoning in plain language. If the human cannot explain the output, they cannot legally ratify it.
Future Implications
The current wave of agentic AI is poised to disintermediate the very consulting firms that promoted it. As AI agents begin to handle the generic analysis and polished presentations that were once the hallmark of management consultancies, firms like Accenture and Capgemini have seen significant market valuation corrections.
However, substituting a relationship-based dependency (consultants) for a platform-based dependency (AI vendors) may lead to a deeper, more structural lock-in. The "con" remains the same; only the medium has changed. The challenge for the coming decade will be to integrate these powerful tools without hollowing out the human capacity to steer them. As the first-year programming students at Imperial College were once told, the most important part of the process happens before the computer is even turned on: it is the act of thinking for oneself.
