IBM Research argues that raw LLM capabilities are insufficient for scalable enterprise AI adoption, and that 'agent logic' — software primitives like knowledge graphs, program analysis libraries, and algorithms operating at the agentic layer — is the key differentiator. Four enterprise use cases are presented with concrete results: (1) Legacy COBOL/PL1 app understanding with ~30× lower token consumption vs. frontier LLM-only; (2) Test generation with Aster achieving 20–45% coverage improvement and up to 15× fewer tokens; (3) Incident root cause analysis using knowledge graphs achieving 4× improvement over ReAct agents; (4) IT compliance automation boosting success rates from single digits to 80%+. Two domain case studies in healthcare and physical asset maintenance further illustrate the approach, with the asset maintenance agent reducing analysis time by 97% and token usage by 77%.
