What's shipping
Enterprise Java shops are testing a specific architecture: use LLMs to parse natural language business rules once, then execute those rules through standard Java objects. No proprietary DSLs. No runtime AI non-determinism.
Vishal Mysore's Tools4AI framework demonstrates the pattern across e-commerce pricing, insurance underwriting, and mortgage approval - 27 interconnected rules that compile to type-safe POJOs. The approach: business analysts write rules in plain text, OpenAI transforms them into Java classes, execution runs through annotated Action methods. Version control via Git, zero-downtime updates if rules load dynamically.
Why this matters now
Drools and similar rules engines have locked enterprise teams into proprietary syntax for two decades. The timing shift: LLMs are finally reliable enough for one-time parsing tasks, while still too unpredictable for runtime decisions. Red Hat argued this case in January 2026, pushing Java over Python for agentic apps specifically because JVM security and type safety matter in regulated industries.
The real pattern to watch: agentic AI for transformation, deterministic code for execution. Spring AI added agent routing in January 2025. LangChain4j ships with Vertex AI integration. These aren't research projects - they're pulling production workloads in finance and healthcare where audit trails and explainability are non-negotiable.
The trade-offs
You're swapping Drools complexity for LLM API dependencies and prompt engineering. Every rule change still requires careful testing - the AI parser can misinterpret ambiguous business language. Moveworks warns that custom agentic builds without pre-built governance components create consistency risks.
What you gain: business analysts who actually understand the rules can now read and modify them. What you lose: the maturity of 20-year-old tooling and IDE support.
Three things to watch
Observability gaps: Spring AI and Quarkus are adding structured output tracking, but debugging AI-transformed rules in production remains harder than debugging hand-coded logic.
Cost at scale: OpenAI API calls for rule parsing are cheap until you're managing thousands of rules across regulatory jurisdictions.
Java's resurgence in AI: If the pattern holds, enterprise AI tooling splits - Python for model training, Java for production deployment where supply chain security and type safety matter.
The code is on GitHub. The test results are documented. The question is whether your compliance team will accept LLM-parsed rules with the same confidence they gave Drools.