When an AI Agent Makes a Decision, Who Is Accountable?

I was reading about a municipal system in Singapore where citizens report potholes through an app, and an AI agent reviews the photos, summarizes the report, triggers procurement, assigns a vendor, verifies completion, processes payment, and sends a closure message to the citizen. A single bureaucrat approves it with one click. I kept staring at that description and asking myself: if the algorithm assigns the wrong vendor, or misses a safety hazard in the photo, or overpays, who is accountable? The citizen who trusted the app? The bureaucrat who clicked approve? The vendor? The AI developer? The city?

The question is not rhetorical. Feng and Chandra frame this as a principal-agent problem that has escaped its human bounds. Classical principal-agent theory assumes a principal delegates to a human agent, monitors for shirking, and holds them accountable when outcomes deviate from intent. When the agent is an AI system, information asymmetry becomes nearly absolute. The principal cannot read the model weights, cannot audit the training data, and cannot reconstruct why the algorithm interpreted a cracked road surface as a routine patch rather than a structural risk. The agent's objectives may have been subtly reshaped by third-party designers, data biases, or self-optimizing feedback loops that the principal never authorized. Feng and Chandra call this delegation inversion, where the AI agent evolves into the de facto principal and delegates tasks back to the human, who becomes a procedural rubber-stamper. Human officials nominally hold authority but cede their judgment to the machine's logic. They also describe AIcratism, where public employees defer excessively to algorithmic outputs, hollow out their own discretion, and transform accountability into a hollow formality.

Mukherjee and Chang go further into the legal and ethical vacuum. When an agentic AI autonomously books a non-refundable flight, selects a hotel in a high-risk neighborhood, or denies a loan application, accountability does not simply shift from one actor to another. It disperses. Everyone involved can point to someone else. The user commissioned the AI. The developer built the system but claims it operates autonomously. The platform hosted the transaction. The AI itself cannot be sued. Mukherjee and Chang borrow the term moral crumple zone to describe contexts where responsibility falls unfairly on the lowest-level human with the least control, effectively absorbing blame that should be distributed upward. They also cite Matthias and call this a responsibility gap: autonomous systems make decisions that even their manufacturers cannot reliably predict, which breaks the chain of liability that tort law assumes.

Principal-agent theory predicts exactly this outcome. The theory was built for situations where a principal cannot fully observe an agent's actions and must design incentive structures and monitoring mechanisms to align interests. When the agent is an algorithm, the classic solutions break down. Contracts cannot specify every edge case a neural network might encounter. Monitoring is impossible when the reasoning is proprietary and opaque. Sanctions are meaningless when the agent has no assets, no career, and no reputation to protect. Baird and Maruping, writing in MIS Quarterly in 2021, already argued that IS scholarship needs to move past use theory and toward delegation theory because agentic systems transfer rights and responsibilities in ways that acceptance models like TAM cannot capture. But even delegation theory, as they framed it, mostly assumed humans delegate to artifacts. Stelmaszak and colleagues, in a 2025 paper on Uber, showed that algorithms can also delegate to humans, reversing the direction entirely. When artifacts delegate back to humans, or when multiple agents interact in ways no single principal designed, the accountability chain frays even further.

I think the field of Information Systems has a specific obligation here. We are not just studying another technology trend. We are studying a shift in how authority is distributed across sociotechnical systems. When an AI agent makes consequential decisions, accountability cannot be an afterthought buried in a terms-of-service agreement. It needs to be engineered into the architecture of delegation itself. IS researchers can contribute by designing accountability frameworks that specify who has the right to override which decisions, how explanations are generated and audited, and what happens when an agent's behavior diverges from the principal's intent in ways no one anticipated. The alternative is a growing landscape of accountability vacuums, where sophisticated systems act on behalf of organizations and governments, and citizens are left holding the consequences without a clear path to redress.

## Source Verification

- Feng & Chandra (2026) — "New development: Principal–agent issues when governments embrace AI agents," Public Money & Management. Verified in local file: `/Users/alisafari/Downloads/PHD/UNT/2026/COMPS/BCIS 6670/My Paper/markdown_exports/Feng_and_Chandra_2026_-_New_development_-_Principal-agent_issues_when_governments_embrace_AI_agents.md`. Specific claims used: delegation inversion (lines 207–213), AIcratism (lines 307–315), intensified information asymmetry due to proprietary secrecy and inscrutability (lines 184–189), three delegation interfaces (trust-based, hierarchical, contractual) (lines 252–255), accountability diffusion (lines 236–238), and principal drift (lines 558–566).

- Mukherjee & Chang (2025) — "Agentic AI: Autonomy, Accountability, and the Algorithmic Society," arXiv preprint. Verified in local file: `/Users/alisafari/Downloads/PHD/UNT/2026/COMPS/BCIS 6670/My Paper/markdown_exports/Mukherjee_and_Chang_2025_-_Agentic_AI_-_Autonomy,_accountability,_and_the_algorithmic_society.md`. Specific claims used: complete agency versus partial agency (lines 486–497), responsibility gap referencing Matthias (lines 692–697), moral crumple zone (lines 473–475), and liability diffusion when an agentic AI autonomously books flights or makes decisions (lines 637–655).

- Supporting IS context — Baird & Maruping (2021) "A Theoretical Framework of Delegation for Agentic IS Artifacts," MIS Quarterly, verified in study-hub: `/Users/alisafari/Downloads/PHD/UNT/2026/COMPS/study-hub/Final Paper/ALL_PAPERS_MD/[6010] Baird and Maruping (2021) MISQ - A Theoretical Framework of Delegation for Agentic IS Artifacts/paper.md`. Stelmaszak et al. (2025) When Algorithms Delegate to Humans, referenced in study-hub day files and `paper-summaries.json`. Recker et al. (2025) "Digital Responsibility: Current Perspectives and Future Directions," JAIS, verified in study-hub: `/Users/alisafari/Downloads/PHD/UNT/2026/COMPS/study-hub/Final Paper/ALL_PAPERS_MD/[6010] Recker et al (2025) JAIS - Digital responsibility -Current perspectives and future directions/paper.md`.