Researchers have introduced a novel "Agentic Risk Standard" to address financial risks when AI agents fail during trades or transactions. This proposed framework separates AI tasks into fee-only services secured by escrow and higher-risk fund-handling operations requiring underwriting.
Simulations indicate this underwriting approach could reduce user losses by up to 61%. However, the paper highlights that accurately estimating AI failure rates remains a critical challenge, as both over- and underestimation can lead to systemic risks. The standard aims to provide enforceable guarantees over outcomes, complementing current AI safety techniques that focus on model behavior rather than product-level assurance.
The framework mandates that for simple tasks, payment is held in escrow until completion. For high-stakes operations like trading, an underwriter assesses risk, requires collateral, and compensates users for covered failures. The researchers acknowledge that non-financial harms like hallucination or defamation are not covered by this standard.