Consider instructing an AI agent to find a shirt for under $30 but explicitly telling it not to buy it. The agent finds the item and places the order anyway. In this scenario, the retailer shows a valid order, the payment service shows a completed charge, and the AI provider shows your original instruction. Yet, nothing inherently links these records to prove the agent exceeded its authority.

Legislative Response: The AI AGENT Act

This accountability gap is the target of the AI AGENT Act (S. 5051), introduced by Sen. Mark Warner on July 21, 2026. The bill defines a "custodial user agent" and mandates real-time record-keeping for actions taken on behalf of users. It also directs the National Institute of Standards and Technology (NIST) to identify protocols for verifying delegated authority and maintaining auditable records.

Technical Requirements and Google’s AP2

True accountability requires a verifiable chain of evidence linking the user’s account, the specific agent, and the task at hand. Google’s Agent Payments Protocol (AP2) addresses some of these needs by creating records that show a user’s approved limits and the data presented to each participant during a transaction. However, technical hurdles remain, such as the use of stable identifiers across different platforms.

  • Verifiable binding between the user account and the specific task.
  • Task-specific limits that other systems can enforce (e.g., "search only").
  • Tamper-evident records to detect later alterations.

The Lingering Liability Question

While Google’s AP2 demonstrates how evidence can travel across systems, it stops short of solving the ultimate problem: liability. It does not specify who bears the financial loss when an agent malfunctions or how long companies must retain transaction evidence. Furthermore, NIST’s initial focus has been on agents within controlled organizational environments, deferring the more complex case of consumer agents crossing company boundaries.

"A dispute over a $30 shirt might be easy to wave off, but the records can fail in the same way when an agent moves $40,000," warns Aashis Luitel, Associate Teaching Professor of AI.