The Shadow Agent: Governing Unseen Delegation in the AI-Driven Economy | Carlos Galera-Zarco

The Shadow Agent: Governing Unseen Delegation in the AI-Driven Economy | Carlos Galera-Zarco

Title history
  • The Shadow Agent: Governing Unseen Delegation in the AI-Driven Economy | Carlos Galera-Zarco — observed on 20261105-1152 UTC (YouTube)

by Carlos Galera-Zarco · View the channel on PSI

Humanities, Social Sciences & Thought AI & Society UBDigital and Information technology: general topicsUBJDigital and information technologies: social and ethical aspects
🎙 Carlos Galera-Zarco 👥 878 📅 November 5, 2026 ⏱ 39 min 👁 0 📄 expert opinion 🧭 2026-11-05
Available in: English (current) Français

Keywords

AIshadow agentcognitive effortagency theorygovernance

Summary

The talk by Prof. Carlos Galera-Zarco, presented at the ADIA Lab Symposium 2026, addresses the growing phenomenon of employees delegating cognitive work to generative AI outside formal organisational oversight, creating what he terms a ‘shadow agent’. Drawing on agency theory, he explains that organisations (principals) delegate tasks to employees (agents), but with AI, a second layer of delegation emerges where employees sub-delegate reasoning to AI models that are invisible to the firm. This creates informational asymmetries, misaligned incentives, and accountability gaps. He illustrates risks with cases such as a community bank’s data leak, KPMG’s hallucination-ridden report, and a Deloitte consultancy error. Empirical evidence from Carnegie Mellon and Microsoft shows reduced critical thinking when workers over-rely on AI, and educational data indicate that AI use improves assisted tasks but harms independent competence. He proposes solutions: introducing calibrated cognitive friction in workflows, such as requiring employees to make initial judgments before consulting AI, and redesigning incentives to reward verification and quality rather than speed. He concludes that while AI boosts productivity, organisations must redesign workflows to preserve human agency, ensure accountability, and maintain cognitive skills, turning productivity into organisational wisdom.

188 words

Critical Evaluation

Value of the Information & Strength of the Argument

The presentation provides high value by synthesising well-known agency theory with contemporary AI usage problems, offering a novel framework for understanding ‘unseen delegation’. The argumentation is coherent and compelling: it establishes the problem with concrete incidents, contextualises it theoretically, and proposes practical remedies. The use of statistics (e.g., 86% of financial companies using AI) and experimental evidence (e.g., cognitive forcing functions) strengthens the credibility. However, some causal links, such as the direct erosion of cognitive abilities from AI use, are extrapolated from educational studies and may not directly transfer to workplace settings. The speaker also acknowledges limitations, such as the lack of direct workplace examples, and invites discussion, which adds balance.

Scientific Rigor, Source Quality, Title Accuracy

The talk maintains a rigorous scientific approach, referencing established theories and real-world cases without overclaiming. Sources named include Carnegie Mellon and Microsoft research, and reports from the financial sector, though no specific URLs are provided. The description only contains a link to the symposium page, which does not directly support the cited studies. The title is accurate and well-aligned with the content, as it introduces and governs the ‘shadow agent’ concept. There is no noticeable discrepancy between title and content. The author’s academic background and affiliation with UCL lend authority, but the lack of peer-reviewed references explicitly mentioned is a minor shortcoming.

228 words

Title / Content Match

The title accurately captures the core concept of an invisible AI 'shadow agent' and its governance implications, though the subtitle is more descriptive than the talk itself.

Quality & Reliability

8/10

The presentation is delivered by a UCL academic, builds on agency theory, and references real incidents (community bank, KPMG, Deloitte) and empirical studies (CMU & Microsoft). While some data points rely on unpublished or non-peer-reviewed reports, the arguments are logically structured and grounded in established organisational research.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • ADIA Lab Symposium 2026 — Symposium page where the presentation was given; contains general event information but no specific research references.

Concurring Sources

  • No specific sources provided — The speaker mentions reports from Carnegie Mellon and Microsoft, and incidents from KPMG and Deloitte, but no URLs are given.

Contribution & Novelties

The primary contribution is the conceptualisation of the ‘shadow agent’ as a second layer of delegation in the principal-agent framework, highlighting how AI can invisibly absorb cognitive effort while organisations remain accountable for errors. The talk also synthesises empirical findings on automation bias and cognitive skill erosion, and proposes practical governance mechanisms such as calibrated friction and incentive redesign. This adds a novel organisational perspective to the current discourse on AI risk management.

Pour aller plus loin :

  • Principal–agent problem — Foundational theory underlying the shadow agent concept; relevant to understanding delegation and information asymmetry.
  • Automation bias — Key psychological phenomenon explaining overreliance on AI and the ‘hazard of fluency’.
  • Cognitive load — Theoretical basis for why cognitive forcing functions reduce delegation and preserve independent thinking.

126 words

Radar Profile

The radar profile shows high scores in information quantity and global reliability, with moderate technical level and quality. This indicates a talk that is information-dense and credible, but not extremely technical or pathbreaking in methodology.

Reliability 8/10