AI agents, which rely on large language models, are increasingly used for tasks like coding, customer communication, and IT environment management. These systems are inherently nondeterministic, meaning they can produce different outputs for the same input.
While deterministic software is preferred for tasks requiring a single correct answer, such as payroll, AI agents are better suited for scenarios where no single right answer exists, including business forecasting and incident mitigation.
Organizations can evaluate the risks of deploying agents by calculating an error budget. This involves determining the acceptable percentage of agent misbehavior and comparing the potential business benefits against the total costs of failure.
Current efforts to govern and secure AI agents often aim to restrict their unpredictable nature. However, these limitations may undermine the very agency that makes LLM-based tools valuable for complex tasks.
