Architecting AI for Industrial Determinism and Reliability
An expert analysis of the architectural challenges when integrating probabilistic AI models into deterministic industrial environments. The focus is on mitigating hallucinations through Simplex and Hexagonal architectures and ensuring regulatory compliance.
The Determinism Dilemma in Industrial AI
For leadership and investors in the industrial sector, the allure of Large Language Models (LLMs) is significant. However, a critical architectural gap exists: industrial automation relies on determinism—the guarantee that a specific input always produces the same, correct output. LLMs, by nature, are probabilistic engines. In a factory setting, a system cannot suggest that a production line "probably" needs to stop; it requires a binary, reliable decision.
Strategic Mitigation Patterns
To bridge the gap between probabilistic AI and deterministic requirements, high-reliability systems must employ specific architectural patterns:
- Simplex Architecture: This approach utilizes three distinct blocks: the AI model, a deterministic fallback logic, and a monitor. The monitor evaluates the AI's confidence and input distribution, automatically switching to the deterministic path or a human operator if the AI behaves unpredictably.
- Hexagonal Architecture: By isolating AI models within their own "hexagons" and interacting via strict ports and adapters, organizations can replace outdated or drifting models without disrupting the rest of the system's business logic.
- AI Gateways: Implementing a gateway to filter metadata and confidence scores ensures that only high-probability outputs reach the execution layer.
Governance and the Path to Compliance
With the upcoming AI Act (2026), transparency is no longer optional. The extension of the ARC42 documentation framework for AI is crucial. This framework mandates detailed documentation across four views: data sources/transformation, model behavior reasoning, runtime deployment/retraining, and explicit risk ownership and mitigation.
Conclusion
AI should be viewed as a modular block with specific peculiarities, not the foundation of a decision engine. By applying classical resilience patterns—such as Circuit Breakers and Dependency Injection—industrial leaders can leverage the power of AI while maintaining the safety and stability required for physical production.
Key insights
-
LLMs are probabilistic engines rather than knowledge generators, meaning they interpolate patterns rather than computing facts. This makes them fundamentally incompatible as primary decision engines in industrial automation.
Impact: Prevents catastrophic failures in industrial settings where deterministic outcomes are mandatory for safety and efficiency.
-
Classical AI models, such as decision trees and random forests, remain superior for regulated industries because they are interpretable and transparent.
Impact: Ensures compliance with regulations like GDPR Article 22, which requires clear explanations for automated decisions affecting humans.
-
The "Confidence Illusion" occurs when a model provides a high probability score for an answer that is factually incorrect due to pattern matching rather than reasoning.
Impact: Highlights the danger of relying solely on AI-provided confidence metrics without independent validation.
-
Local LLMs are preferable to vendor-hosted APIs for industrial use cases to avoid costs, enhance security, and prevent system breakage caused by frequent vendor model updates.
Impact: Increases system stability and data sovereignty for enterprise-level deployments.
-
Synthetic data generated by Generative AI can be used to simulate rare failure modes (edge cases) that are not present in real-world training sets.
Impact: Improves model robustness by training systems on high-risk scenarios without needing to cause real-world accidents.
Action items
-
Implement a Simplex Architecture consisting of an AI model, a deterministic fallback, and a monitor to handle AI uncertainty.
Impact: Eliminates the risk of probabilistic errors affecting physical production or human safety.
-
Adopt the ARC42 extension for AI documentation, focusing on data provenance, model behavior, runtime, and risk mitigation.
Impact: Facilitates compliance with the 2026 AI Act and improves architectural transparency.
-
Encapsulate AI components using Hexagonal Architecture to ensure models can be swapped or updated without altering core business logic.
Impact: Reduces vendor lock-in and mitigates the impact of "silent drift" or model deprecation.
-
Deploy an AI Gateway to monitor metadata and confidence scores, triggering a Circuit Breaker pattern when model performance degrades.
Impact: Enhances system resilience by preventing erroneous AI outputs from propagating through the architecture.
-
Maintain a "Human-in-the-loop" requirement for all final diagnoses or high-impact decisions, regardless of AI confidence levels.
Impact: Provides a critical layer of verification that AI, limited by training patterns, cannot provide.
Quotes
“Industrial applications, based on the main assumption that everything should be deterministic.”
“AI is not even LLMs, not only LLMs and not only generative AI, there are different types and how these types can work as a part of architecture.”
“the most probabilistic part of your system should be isolated from the control group where you're affecting the physical reality around human beings or your production.”