Enterprise AI Strategy Demands Weighing Agent Risks Against Deterministic Software, Analysis Argues

As artificial intelligence agents see broader adoption across software development, customer communications, and enterprise information technology management, organizations face operational questions regarding their reliability. In an analysis published by SiliconANGLE (https://siliconangle.com/2026/10/11/why-ai-agents-might-not-be-right-for-you/), Jason Bloomberg, founder and managing director of digital transformation advisory firm Intellyx, outlined the technical and economic trade-offs leaders must weigh before substituting deterministic software with autonomous agents.

At the core of agentic systems is nondeterministic behavior stemming from large language models. Unlike traditional software, LLMs process natural language inputs by predicting sequential words, meaning identical inputs can generate divergent outputs. Bloomberg notes that this variability gives agents their agency—the capacity to make decisions and take distinct actions to achieve assigned goals. Systems without this agency remain deterministic, despite marketing efforts to rebrand conventional software as agentic. Prior to LLMs, nondeterministic systems were largely limited to stochastic software relying on randomization, such as Monte Carlo simulations and reinforcement learning.

In contrast, traditional enterprise software is deterministic: given specific inputs, it reliably produces identical results. Deterministic consistency is required in domains where absolute accuracy is paramount, such as executing payroll, processing financial transactions, or allocating airline seating. In these scenarios, software is engineered to find and return a single correct answer.

Conversely, LLM-driven nondeterministic reasoning is suited for environments with variable inputs where no single right answer exists, including weather-dependent operational planning, IT incident mitigation, and multi-variable business forecasting. However, because these systems rely on probability, agents carry a perpetual risk of generating suboptimal responses or initiating unintended actions.

To evaluate whether deploying an agent is justified, Bloomberg recommends adopting the concept of an error budget from site reliability engineering. An error budget quantifies the acceptable rate of agent misbehavior as a percentage based on risk tolerance. If an organization determines that the acceptable error rate for a specific workflow is zero percent, deploying a nondeterministic agent is inappropriate.

Organizations can then apply decision theory to assess deployment viability. Under this framework, an agent should not be deployed if the probability of failure multiplied by the total cost of failure exceeds the probability of success multiplied by the quantified business benefit. Because these calculations rely on estimates, Bloomberg advises that expected benefits should exceed potential failure costs by a substantial margin.

The analysis also identifies secondary risks, including the anthropomorphism of AI systems, where attributing human cognitive qualities to bots can lead to overdependence and counterproductive behavior. Additionally, Bloomberg cautions against governance tools that overly constrain agent autonomy. Restricting nondeterministic behavior to eliminate risk can strip agents of the adaptability that makes them useful. Bloomberg predicts that after market hype moderates, enterprises will find agents are best deployed sparingly alongside established deterministic applications.