The promise of AI agents is that they do the work for you. The reality is that they do the risk for you. When something goes wrong, the industry has built a system where nobody signs for the loss. The liability perimeter has moved. It is now sitting squarely on your desk.
Look at the foundation of the stack. StepFun Open Platform Terms Section 8 sets a clear tone. They cap their liability at 100 dollars. They also demand broad customer indemnity and provide their services on an AS IS basis. If the model fails, you are the one holding the bill.
This shift moves up to the institutions that hold your money. The Wells Fargo Online Access Agreement update effective Nov 24, 2026, is the first major US bank to explicitly codify agentic payment liability. They state that third-party AI tools are not operated by the bank. If an agent initiates a transaction, the bank treats that request as your own. You are responsible for the outcome.
The law is catching up to the mess, as we discussed in our analysis of the agent governance choice. Colorado Rule 8.7, part of the Oct 6 2026 draft redline of the Artificial Intelligence Act, requires “meaningful human review.” This means the reviewer must have actual authority to approve, modify, or override decisions made by an automated tool.
The state wants records documenting the identity and training of that human. If the review is not meaningful, you lose your compliance defense. This creates a trap. You are forced to provide human oversight for a machine that is supposed to be autonomous.
This adds a hidden labor line to the deployment bill, reinforcing the problem of the unpriced reviewer. While enterprise harnesses move toward 24-hour durable execution, vendor benchmarks price the reviewer at zero. Companies build agents to save time, but they offload the legal burden of failure onto the person clicking the button.
The system is designed to make the agent look like a tool while making the user look like a principal. When the agent makes a mistake, the bank points to the user. The model provider points to the user. The law points to the user.
Users must document the training and maintain the authority to override, all while paying for the loss when the agent goes rogue. The practical caveat here is simple. If you cannot afford to lose the money or the reputation, do not let an agent touch the transaction. The liability is not shared. It is transferred.