Abstract By late 2024, generative artificial intelligence systems were increasingly embedded within enterprise workflows that demanded accountability, traceability, and regulatory defensibility. Large language models were no longer confined to experimental use cases, but were deployed to support decision assistance, content generation, operational analysis, and customer interaction across regulated and high risk domains. This shift exposed a structural gap between the probabilistic nature of generative AI systems and the audit expectations traditionally applied to enterprise software. Unlike deterministic applications, generative systems produce outputs that depend on dynamic prompts, contextual data, model configurations, and stochastic inference processes, complicating the ability to reconstruct and explain system behavior after the fact.
Ramani Teegala (2024) studied this question.