In this LinkedIn post, I compared the 2003 and 2026 results of using LLMs to generate decision models from plain English descriptions. Here is the conclusion:
Three years ago, I was cautious about judging newly introduced generative AI capabilities. Today, even these two examples demonstrate that LLMs can be really good at transforming relatively simple business logic, while they struggle with complex, intersecting conditions. Still, the progress is undeniable. Modern LLMs can take a plain English description of a well-specified business function and generate a reasonably good initial decision model complete with glossaries, business rules, and test cases. These models can then be refined using business-friendly graphical interfaces and executed via decision engines to determine optimal or good decisions. We will undoubtedly see such capabilities, with significantly improved quality, integrated into Decision Intelligence platforms sooner than one might expect.
OpenRules - Decision Intelligence Platform 
Initially, I was impressed by how the LLM built a ‘complete’ decision model from a short plain English problem description. My enthusiasm for the nicely-formatted decision tables and test cases faded when I began validating their correctness. Colleagues reviewing the model discovered even more serious errors—not only in the business rules but in the test cases themselves. This reveals a troubling pattern: the more information an LLM generates, the more validation it requires, and finding logical errors is far from straightforward. Real-world problems, of course, involve far more complex business logic.
Everyone agrees LLM outputs need validation, but the reality is challenging—even for this simple case, we needed a manually created decision model to verify the correctness of the automatically generated model. As one colleague asked: Doesn’t that defeat the purpose? It would be interesting to conduct a similar analysis in January of 2027.