Mauro Vignati

Governing Just-in-Time Intelligence: AI, Dual-Use Capabilities, and the Limits of Guardrails

This talk traces the historical evolution of the Just-in-Time (JIT) production paradigm into contemporary Model-as-a-Service (MaaS) systems, arguing that today’s governance challenges surrounding advanced AI are rooted in a much older logic of capability distribution. Developed by Toyota after World War II, JIT transformed industrial production through lean management, waste reduction, and on-demand resource allocation. The digitalization required to support JIT contributed to the emergence of Application Service Providers and, subsequently, the broader “as-a-Service” model, including Software-, Infrastructure-, Platform-, and Everything-as-a-Service (XaaS).

The talk shows how this service-based logic migrated beyond legitimate markets. Beginning in the 2000s, cybercriminal ecosystems adopted Crimeware-as-a-Service (CaaS) and Ransomware-as-a-Service (RaaS), enabling actors with limited technical expertise to access sophisticated offensive capabilities. The emergence of frontier AI systems and centralized cognitive platforms has extended this trend into the cognitive domain through Model-as-a-Service (MaaS), where advanced reasoning and problem-solving capabilities are delivered on demand according to the same principle: users obtain only the capabilities they need and pay only for what they consume.

Unlike JIT, which optimized production, or CaaS/RaaS, which distributed explicitly offensive tools, MaaS introduces a fundamentally dual-use model. Advanced AI capabilities can support productive and socially beneficial activities while simultaneously enabling malicious applications. This development is also reshaping traditional governance structures, particularly in sectors historically regulated through ex ante controls, such as defense procurement, by replacing ownership-based control with platform-based distribution subject primarily to ex post intervention.
The talk argues that current debates focusing on guardrails and safety mechanisms overlook a deeper structural issue: the application of a seventy-year-old distribution model to advanced cognitive capabilities. The central challenge is therefore not merely how to secure AI systems, but whether the as-a-Service paradigm itself is appropriate for governing increasingly powerful cognitive technologies.