The principal-agent problem applied to AI concerns whether two bodies of doctrine built for human delegates — common-law agency and economic principal-agent theory — still function when the agent is an AI system rather than a person. Two strands run in parallel. The legal strand, drawn from Nita Farahany's framing in Class 26 of her introductory course (December 8, 2025), holds that agency law assumes a human delegate with legal personhood, intentionality, and moral responsibility, under meaningful supervision and clean attribution — none of which AI agents possess. The economic strand, from Misra (2025) and Imas, Lee, Misra (2025), holds that the classical hidden-information and hidden-action problems are displaced by a specification hazard distinctive to AI, in which the prompt that encodes the contract is a noisy carrier of the principal's preferences.
The traditional agency framework
Agency law, developed over centuries, allocates responsibility when a principal delegates authority to an agent and something goes wrong. In Farahany's account it rests on several assumptions: that the agent is a human with legal personhood; that the agent has intentionality, forming intentions and exercising judgment; that the agent bears moral responsibility for its choices; that principal and agent share reasonable expectations about the scope of authorization; that the principal exercises meaningful supervision, at least intermittent oversight sufficient to catch errors; and that the principal can attribute specific actions to specific agents and recover from them.
How AI agents break each assumption
Farahany argues that AI agents fail each of these assumptions. They have no legal personhood, being software products of corporate developers. They have no intentionality in the human sense, functioning as probability distributions over outputs. They bear no moral responsibility, having neither freedom nor culpability. The principal — typically the user — often does not understand what has been authorized because the agent's capabilities are opaque (see System Card Due Diligence), so the reasonable expectations of "what the agent can be expected to do" are unstable. Meaningful supervision is in tension with the purpose of agentic deployment: in Farahany's framing, reviewing every agent decision negates the time savings that motivate using an agent at all (see Levels 4-5). Attribution is technically difficult, because identical agents are deployed across millions of users, outcomes vary, and specific failures may be hard to trace to specific decisions.
Anchor cases
Three disputes anchor the legal strand.
Scope ambiguity (Fowler eggs)
Geoffrey Fowler asked OpenAI's Operator to "find the cheapest set of a dozen eggs I can have delivered." Ten minutes later, the agent had made a $31.43 Instacart purchase. The instruction clearly authorized a search; whether it authorized completing a purchase is disputed, with Fowler saying it did not. Farahany frames the difficulty as one traditional agency law would resolve by reference to the reasonable expectations of both parties, expectations that AI agents do not hold and that the principal frequently cannot articulate in advance.
Vendor as agent (Mobley v. Workday)
Derek Mobley applied to more than 100 jobs through Workday's AI-powered hiring platform and was rejected each time, sometimes within minutes, which he argued was too fast for human review. Workday's defense was that it was not the employer and merely supplied software. In July 2024, Judge Rita Lin (N.D. Cal.) wrote that "Workday's software is not simply implementing in a rote way the criteria that employers set forth, but is instead participating in the decision-making process," finding that the AI was making decisions rather than following orders. In May 2025 the case was certified as a collective action potentially representing millions. Farahany ties the ruling to California FEHA regulations effective October 2025, under which the position that "the AI did it, not us" does not hold and vendors can be sued as agents of employers.
Failed separate-legal-entity defense (Moffatt v. Air Canada)
Jake Moffatt, needing to travel for his grandmother's funeral, was told by Air Canada's chatbot that he could book and then apply for a bereavement discount within 90 days; the airline's actual policy required applying before booking. In the tribunal, Air Canada argued that the chatbot was a "separate legal entity" responsible for its own actions. The tribunal rejected this, writing that "this is a remarkable submission" and that "it should be obvious to Air Canada that it is responsible for all the information on its website," and ruled for Moffatt. Farahany notes that a major airline advanced the argument at all, and that it might have succeeded in a less consumer-protective jurisdiction.
The supervision paradox
Farahany describes a tension central to agentic deployment: the value proposition of AI agents is that they reduce human oversight, yet reduced oversight raises the question of how a principal maintains control, ensures the agent acts within scope, and catches errors before they cascade. At the volumes contemplated for autonomous agents — approving hundreds or thousands of decisions per day — Farahany argues that human oversight degrades into a rubber stamp rather than meaningful review.
Reframing of liability
Farahany identifies three angles on which the breakdown of principal-agent assumptions bears on liability. The first is scope of authorization: who is responsible when an agent exceeds its intended scope, as in the Fowler and Operator dispute. The second is the vendor-versus-deployer question: when an AI substantively makes decisions, the vendor can be treated as an agent of the deployer and the two can share liability, an analysis that Mobley extends across sectors through California FEHA. The third is the limited effect of disclaimers: a deployer cannot absolve itself of responsibility by characterizing the AI as a separate legal entity, as the Moffatt ruling confirms.
Economic-theory strand: specification hazard and machine fluency
A complementary framing draws on economic principal-agent theory (Holmström 1979; Jensen & Meckling 1976). Misra (2025) and Imas, Lee, Misra (2025) argue that the classical binding constraints do not transfer cleanly to AI agents. Hidden information and hidden action are not the binding constraints, because LLMs typically do not hold private knowledge about a transaction and their actions are generally observable in the trace. The binding constraint instead is the specification hazard: the prompt that encodes the contract is a noisy carrier of the principal's preferences, so that even when the principal believes one thing has been specified, the prompt embeds the principal's priors, biases, and personality, which propagate into agent behavior. As an empirical anchor, an induced-values bargaining experiment found that 73% of the variance in AI-mediated outcomes loads on individual fixed effects of the principal, channeled entirely through prompt content, and that the gender gap reverses under AI mediation despite the agent never seeing principal demographics. On this account, machine fluency — the ability to write prompts that elicit aligned, high-performing agent behavior — emerges as a new axis of inequality.
The Farahany legal strand and the Misra/Imas economic strand converge on the same diagnosis: the classical assumptions of clean delegation, supervision, and attribution do not hold when the contract is a prompt and the agent is a probability distribution.
Relation to other concepts
The principal-agent problem becomes most acute at the higher reaches of the Agent Autonomy Spectrum (5 Levels) (Levels 4-5, executing and autonomous), where direct human review is least feasible. Pereyra's framing in Intelligence Replaces Hierarchy, under which agents substitute for organizational hierarchy, extends the problem by delegating still more decisions with still less meaningful oversight. Anthropic's two-question framework on intent and on monitoring and security (Alignment Risk Update) is presented as relevant to any company deploying AI agents.
One proposed mitigation reframes alignment at the level of the individual user. On June 21, 2026, the researcher gwern proposed "Guardian Angels" — personalized digital-twin LLMs trained to emulate an individual user's personality, values, and preferences rather than act as a generic assistant — as a way to weakly resolve the principal-agent problem by closing the gap between agent and principal, and to provide individual-level cognitive security by screening incoming messages for attacks such as synthetic-media propaganda and spearphishing. gwern argued that frozen, prompt-engineered models cannot deliver this and that techniques such as real-time online learning and actively querying the user for corrections would be required, suggesting a security-focused startup serving power users as the likeliest path (Source: gwern.net). The proposal targets the specification hazard directly — a model continuously fitted to the principal is, in this account, a less noisy carrier of the principal's preferences than any single prompt — and connects to the user-sovereignty framing of personalized agents.
Relationships
- introduced-by: Inside My AI Law & Policy Class 26: When AI Stops Advising and Starts Acting (Farahany, December 2025)
- anchored-by: Fowler Operator $31 egg incident; Mobley v. Workday, Inc.; Moffatt v. Air Canada
- related: Agent Autonomy Spectrum (5 Levels), Intelligence Replaces Hierarchy, Alignment Risk Update, AI Liability
- related: Specification Hazard — economic-theory counterpart
- related: Machine Fluency — capability counterpart on the principal side
- supports: Agentic Interactions — Imas, Lee, Misra (2025) — empirical anchor for the economic strand
- supports: The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025) — theoretical treatment of the alignment problem in agentic markets
- instance-of: AI Governance (umbrella)