How to wrangle non-deterministic AI outputs into conventional software? (2025) (www.domainlanguage.com)

In a recent exploration of integrating AI into traditional software systems, Eric Evans addresses the challenge of managing non-deterministic outputs from AI models, particularly large language models (LLMs). He emphasizes the difficulty of combining the flexibility of AI-generated results with the structured nature of conventional code. For instance, while an LLM can intelligently assess the domains a software module addresses, it often produces varied responses each time it is queried, making standardization and further processing challenging. Evans proposes a method to create reliable outputs by clearly separating the tasks of domain modeling and classification, suggesting the formation of a canonical model that allows structured outputs to be integrated back into software. This approach is significant for the AI/ML community as it highlights a pragmatic method for dealing with stochastic outputs in a deterministic environment, promoting more systematic integration of AI into software development. By using established classification standards like NAICS, Evans demonstrates how to enhance consistency in AI outputs. The implications are far-reaching, suggesting that careful task separation and iterative modeling processes can significantly improve software systems' ability to leverage AI without sacrificing reliability or usability.