Financial officers are redesigning security protocols to address risks associated with AI agents. The primary concern is that agents capable of creating vendors, approving invoices, and executing payments can bypass traditional segregation-of-duties controls.
To mitigate these risks, companies are moving approval processes outside of AI systems. Authorization is now being bound to specific transaction details, such as the exact amount, recipient, and account, preventing agents from altering transactions post-approval.
Security teams at organizations including Google, JPMorganChase, and Weaviate have identified vulnerabilities in the model context protocol, such as prompt injection and server-side request forgery. These incidents highlight that data from agentic workflows cannot be inherently trusted.
A report from August 2025 indicated a disparity in AI adoption: 75% of technology firms reported being extremely familiar with agentic AI, compared to 33% of goods firms and 38% of services firms.
Industry experts suggest a risk-based autonomy model. Under this approach, agents handle repetitive, low-value transactions within set limits, while significant changes, such as new beneficiaries or large payments, trigger mandatory human oversight.
