Atlas Support
ResearchEconomics

The Principal–AI Agent Problem: Governance in the Age of AI Agents and Principal–AI Agent Theory

An AI agent can act against a principal's wishes even when it has no private interest of its own. The principal cannot always specify every condition, observe the action in time, or correct it once it is underway. Governing delegated AI therefore includes deciding who can approve, interrupt, and correct an action as well as specifying its objective.

Intent and action can diverge without conflicting interests

Suppose an accounts team tells an agent to contact customers whose payments are overdue. Sending a reminder based on the ledger is one interpretation. If a customer has disputed the invoice and only the account manager knows about it, the principal may want the agent to withhold the message. The agent needs no private motive for this mismatch to occur. It acts on the information and conditions available to it.

AI agency slack is the divergence between the principal’s intent and the AI’s behavior that arises from these limits on delegation. Conditions can be hard to specify, the grounds for an action can be hard to observe, and intervention can come too late. Better instructions can reduce some mismatches. Other cases require changes to permissions or to the workflow around the agent.

Place approval and stop authority where they can still work

The agent might identify overdue accounts and draft reminders while an account manager checks for disputes or extensions before sending. The reviewer needs to see the invoice, the communication history, and the message the agent proposes to send. A log reviewed afterward can reveal an error but cannot take back a message the customer has already received.

Instructions, monitoring, approval, and stop authority act at different points. Permission to read a ledger need not include permission to contact a customer or change the terms of an invoice. Specifying who can see and stop each action makes oversight part of the workflow instead of depending solely on someone noticing a problem.

Compare the cost of oversight with the loss that remains

Requiring approval for every message can reduce mistakes, but it consumes staff time and delays the work the agent was meant to speed up. Sending every message automatically saves review time and leaves the firm exposed when an exception is missed. The agency-cost perspective counts the cost of preventing, detecting, and correcting divergence alongside the losses that remain after those controls.

Giving two agents the same objective can produce very different governance requirements when one drafts and the other can act outside the firm. To assess the value of delegation, consider the time it saves alongside the effort needed to catch exceptions, the consequences of a mistaken action, and how quickly someone can intervene.