**Navigating the AI Con: Avoiding the Pitfalls of Cognitive Delegation**
The current surge of articles, opinion pieces, and forecasts from consultants and AI vendors confidently predicting the imminent takeover of human work by agentic AI is striking. It prompts a crucial question: are we building tools to serve us, or are we sleepwalking into a dependency that erodes our own agency? Looking back to my first year at Imperial College London in 2009, a lecturer’s advice feels more pertinent than ever: sketch out your solution with pen and paper before touching a computer. That deliberate, manual process fosters a deeper understanding and better results—a principle essential to surviving the AI wave.
Today, the seductive promise of AI has ensnared individuals, organizations, and governments. Students submit AI-generated essays, bypassing the cognitive struggle essential for writing and understanding. Companies replace human employees with AI agents, shedding the tacit, institutional knowledge needed to evaluate the output. Courts supplement sentencing with opaque, potentially biased algorithmic risk scores. Whatever the motivation, ceding cognition, judgment, and accountability to AI may become the default, leading to a dangerous forfeiture of human agency. What makes this more concerning is that the polished AI output being deferred to is often of only generic value, and sometimes simply wrong.
This dynamic mirrors the management consulting industry—a powerful analogy. Consultants package generic analyses in polished presentations, projecting greater certainty than their evidence warrants. Organizations that defer wholesale to consultants gradually shed the capacity to think for themselves. Consultants can be wrong in both insight and ethics, just as AI systems are probabilistic pattern matchers with no intrinsic values or accountability. Both derive authority less from correctness and more from the asymmetry between their confidence and the client’s diminishing ability to challenge them.
This article synthesizes what we know about our growing dependence on agentic AI, traces the risks of overdependence at individual, organizational, and societal levels, and proposes ways to reclaim agency before our capacity to do so slips away.
—
### The Original Con
In 2023, economists Mariana Mazzucato and Rosie Collington published *The Big Con*, arguing that incentive structures, information asymmetries, and institutional pressures allowed the management consulting industry to extract returns exceeding the value it creates. Clients who outsource strategic functions over years may lose the internal expertise to evaluate advice, perpetuating the engagement. The parallel with today’s AI vendors is clear: deep integration, high costs, and a customer base progressively less capable of evaluation. The core issue is structural, not moral—repeated delegation begets lost capability. Organizations that “steer more, row less” risk hollowing out the very functions that define them, making future self-sufficiency difficult or impossible.
This pattern stems from comparative advantage—rationally concentrating on what we do best while paying others to handle the rest. The problem arises when this logic extends to essential, strategic work. Delegating thinking, strategy, and policymaking cedes the capacity that shapes identity. With agentic AI, the forfeiture is deeper: agency passes to a system with no real accountability.
—
### The Next Big Con
The AI story of the 2020s mirrors past consulting trends: promising early wins, rational incremental delegation, and structural dependency that becomes visible only when reversal is too costly. Early AI use cases—drafting text, writing code, summarizing documents—deliver results with little friction. The barrier to adoption is low, and the compressed timeline from ChatGPT’s launch institutionalized generative AI faster than gradual adoption might have.
AI vendors, desperate to capture market share, price services below true delivery cost, creating artificially compelling short-term economics—similar to early cloud and SaaS models. This accelerates work displacement beyond consulting’s reach, removing humans from the loop and the organizational knowledge that justified keeping functions in-house.
However, judicious AI use offers significant benefits: automating repetitive tasks frees humans for more demanding work; personalized tutoring tools show promise in engagement and comprehension; AI sparring partners can sharpen thinking when used as a “centaur” tool—augmenting, not replacing, human agency. The danger emerges when delegation becomes pervasive and thoughtless, leading to cognitive atrophy. Humans may become “reverse centaurs,” serving AI rather than directing it, resulting in organizational atrophy and loss of essential knowledge.
Simultaneously, vendor narratives are shifting from augmentation to outright replacement. Software budgets are smaller than payroll budgets, making AI a labor-replacement strategy with a vast addressable market. Outsourced activities with well-defined scope are the low-hanging fruit, but strategically central roles may follow, risking automation of tasks requiring the most human oversight.
A conflict of interest exists: consultancies that drove the labor-outsourcing wave now profit as AI adoption advisors. McKinsey’s projected $2.6 to $4.4 trillion in annual generative AI value conveniently underpins large advisory engagements. The moral hazard is structural: advisors paid to assist with AI transformation lack incentive to build client independence. This dynamic, not individual malice, sustains the “con”—the same logic that sustained consulting’s original outsourcing wave.
—
### The Costs of Growing Dependence
**Individual Cost**: Cognitive delegation is insidious because AI output increasingly resembles human analysis. AI can turn everyone into Orwell’s factory reviewer—producing work with minimal personal engagement. The value of cognitive work lies in the process of producing it; outsourcing that struggle risks forfeiting original thought. MIT professor Micah Nathan described a student whose AI-assisted writing process resembled an addict’s descent: each small step of delegation made the next more likely, until independent writing felt daunting. Academic studies support this: AI assistance impairs subsequent independent performance, encourages stopping sooner, and reduces neural connectivity. The term “cognitive debt” describes how repeated reliance on AI replaces effortful thinking.
**Organizational Cost**: Risks extend beyond deskilling. Distributed institutional memory thins, governance hollows out, and strategic autonomy vanishes when no one can evaluate or challenge vendor systems. As with consulting, this could unfold rapidly: AI initially augments workers, hiring slows, roles decompose around AI automation, and vital knowledge walks away with departing staff. AI services are currently subsidized below true cost, masking dependence. When profitability shifts priorities, prices will rise (e.g., usage-based billing). Organizations slashing headcounts during the subsidy period will face costly reversal. The “oversight tax” compounds this: in high-stakes contexts, humans review AI outputs, paying for both system and reviewer while their own capacity decays. Productivity gains are overstated, and accountability hollow.
**Societal Cost**: Normalizing agentic AI in education, courts, and government risks producing vulnerable graduates, unjust sentencing, and governance against public will. “Human in the loop” (HITL) often means blindly approving output under pressure—an accountability sink. Geopolitically, 90% of global compute and 70–80% of AI investment is U.S.- and China-controlled. Access is becoming hierarchical, flowing to national security, large enterprises, partners, and eventually others. The U.S. government’s suspension of Anthropic’s advanced models for foreign nationals illustrates dependency risks: vendors can constrain strategic options overnight. This echoes Schumacher’s 1973 analysis of technology transfer creating dependency and a “dual society.” Open-weight models partially address this, but the capability gap persists.
—
### Two Dead Ends and a Third Path
Two extremes are tempting but flawed: banning agentic AI outright or uncritically adopting it. Banning forfeits potential benefits in medicine, research, and education. Uncritical adoption lets market logic, indifferent to cognitive ownership and democracy, make decisions by default. The third path—between prohibition and sleepwalking—requires deliberate action at all levels.
– **Individual Level**: Use AI for grunt work and sparring, but retain judgment. Write the synthesis, make the call, sign the outcome. Maintain the productive struggle deliberately. Use AI when it builds capacity, resist when it substitutes for it.
– **Organizational Level**: Treat institutional memory as strategic asset. Build documentation, mentorship, and role rotation. Diversify AI suppliers to avoid single-vendor dependency. Ensure humans in the loop have knowledge, access, time, and accountability. Prioritize structural sovereignty—control over knowledge, compute, data, and models.
– **Societal Level**: Apply appropriate technology criteria: Does deployment increase local agency? Can it function without permanent external dependency? Governments investing in public AI should also invest in open models, interoperability, and shared compute. HITL in consequential decisions requires legal standards: reviewers must explain and defend outputs, not merely attest presence. AI ethics must be empowered with rigorous governance. Regulation, procurement standards, and international coordination complement individual and organizational discipline.
—
### FAQ
**Q: What is agentic AI?**
A: Agentic AI refers to autonomous systems that can plan, decide, and act to achieve goals with minimal human intervention, going beyond simple task automation to perform complex, judgment-based activities.
**Q: How is AI similar to management consulting?**
A: Both can deliver initial value while gradually eroding internal capability. Organizations become dependent on external solutions—consulting firms or AI vendors—losing the knowledge and confidence to evaluate or perform critical work themselves.
**Q: What are the risks of cognitive delegation to AI?**
A: Risks include individual deskilling and “cognitive debt,” organizational loss of institutional memory and strategic autonomy, societal erosion of critical thinking in education and governance, and geopolitical dependency on a few AI providers.
**Q: Can AI be used safely?**
A: Yes, when used as a tool to augment human capacity—particularly for repetitive tasks and ideation—while humans retain judgment, oversight, and final accountability. The key is maintaining deliberate human engagement.
**Q: What can organizations do to avoid AI dependency?**
A: Invest in institutional knowledge and documentation, diversify AI suppliers, ensure humans in the loop retain true authority and time to exercise it, and prioritize strategic sovereignty over cost-cutting headcount reductions.
**Q: What role should regulation play?**
A: Regulation should establish standards for human oversight in high-stakes domains, promote open-weight models and interoperability, and encourage international coordination to prevent geopolitical concentration and ensure accountability.
—
### Conclusion
The consulting and AI waves share a structural logic: seemingly rational short-term delegation accumulates into dependency that erodes steering capacity, without requiring bad actors, and often before consequences become visible. AI operates faster and more deeply than consulting ever did, affecting individuals, organizations, and society simultaneously.
Ironically, the analytical and technical work that consulting sold as high-value is now being automated and bundled by AI vendors. The disintermeditation logic applied to corporate functions now targets consulting itself, as consulting firms face declining share prices. Whether this proves corrective is doubtful; substituting a platform-based dependency for a relationship-based one may deepen lock-in.
Ultimately, we are the ones building AI tools, and we can decide how to use them. The programming lecture’s advice—do the thinking first—scales to companies and institutions. The question is whether we will apply it deliberately, or let short-term convenience carry us somewhere unintended.
