Evgeny Bobkov, Managing Director for Digital Solutions at MegaFon, stated that AI agents must not be granted the ability to independently expand their own authority. Specifically, agents should be restricted from increasing financial limits or disabling control mechanisms.
According to Bobkov, while diagnostics and execution can be delegated to AI, the authority to modify operational boundaries must remain outside the agent's control. The company emphasizes that the right to change these limits should not be accessible to the AI itself.
There is no universal acceptable error rate for automation, as risks vary by operation type. Bobkov noted that while minor errors in request categorization are manageable, unauthorized debits or exceeding authority are critical events that require architectural blocking.
The evaluation of an AI agent's performance should be conducted by a target system rather than the agent itself. Key metrics for assessment include the share of tasks completed without intervention, the frequency and severity of errors, access rights compliance, as well as cost and execution time.
