When AI Agents Pay: A Rollback Governance Framework for Fraud and Execution Failure in Agentic Stablecoin Payments

Stablecoin payment rails and agentic commerce are converging toward an environment in which autonomous AI agents initiate and settle payments on behalf of users without human authorisation of each transaction. This exposes a governance gap: chargeback mechanisms are too slow and too broad for machine speed execution failures, settlement finality precludes open-ended reversibility, and reversible-token standards supply mechanisms without a governance framework for their use. This paper develops such a framework for rollback, a risk-bounded recourse mechanism for high-confidence fraud and execution-failure events, specifying the first-level eligibility and routing logic of a three-layer recourse architecture (prevention, rollback, adjudication) through five governance dimensions formalised as propositions: trigger authority, rollback-eligible conditions, evidence standards, temporal boundaries, and outcome pathways. Developed through conceptual framework analysis informed by design science research, it is evaluated through scenario analysis across five payment-failure cases and a formative expert consultation. The scenarios route consistently, including a deliberately ambiguous case sent to adjudication under the framework’s own logic; the consultation provides preliminary support for all five dimensions. The framework is presented as a first-level eligibility and routing layer; the enabling institutional layer that must operate it (adjudicator accountability, appeal rights, liability allocation, and remedy funding) is delineated but deliberately not specified.