AI-Native Biotechs: World Models Outperform Departments.

AI-native biotechnology companies often replicate traditional human organizational charts when designing their AI agent roles. This paper challenges that approach, proposing a "Company World Model" as a superior abstraction. This model is defined as a persistent, asset-to-value state representation, complete with transition models, explicit value functions, planning capabilities, and continuous updating across all relevant constraints—scientific, regulatory, business development, commercial, financial, and execution. To test this hypothesis, a dry-lab benchmark was developed, featuring 45 retrospective public-information decision cases with strict time cutoffs and hidden outcomes. The study compared four architectures: human-org-mimic, human-org-mimic-plus, AI-native asset-centric, and AI-native value-conversion. The value-conversion architecture, which approximates a Company World Model through a Live Asset Value Record updated by various loops (Deal, Approval, Revenue, Investment Arbiter), emerged as the most effective. Under a success function focused on external business development, regulatory approval and launch, and revenue discipline, the value-conversion architecture achieved the highest automatic score and was strongly preferred by value-specific blinded judges. While stress tests showed a strong human baseline remained competitive and a neutral judge didn't find robust dominance, mechanistic ablations suggested that components like "Revenue Room" and "Deal Room" were crucial. The central finding is that for AI-native operations, a shared, predictive asset-to-value state should be the core primitive, rather than static human organizational charts, though departments may still serve as useful governance views.