Risk-based decision making makes an AI agent treat actions differently according to their potential consequences. Minor errors may proceed, while high-stakes or irreversible actions require stronger checks and human confirmation. section 9.2.2
Apply stricter controls when an action could cause substantial or irreversible harm:
- Flag transactions above a defined threshold for human approval. section 9.2.2
- Cross-check proposed actions against previous transactions to detect anomalies. section 9.2.2
- Require multistep verification before executing irreversible actions, supported by rollback mechanisms and audit logs. section 9.2.2
Finance agent
A finance-focused agent should not initiate a large fund transfer from one ambiguous instruction. It should flag the transfer, compare it with prior activity, obtain approval, and verify the action before execution. section 9.2.2
