Reframing the AI alignment problem: insights from business applications

The AI alignment problem, traditionally framed as a challenge of encoding universal human values into artificial intelligence systems, faces a critical gap between philosophical aspirations and operational reality. This paper argues that business deployment of large language models (LLMs) reveals alignment as inherently contextual, requiring solutions at the point of deployment rather than solely during foundation model development. Drawing on empirical evidence from organizational AI failures and model risk management practices, a framework of contextual alignment is proposed that layers staged deployment and continuous monitoring on top of pre-deployment value encoding, recognizing that neither alone is sufficient. This reframing has significant implications for both AI ethics scholarship and regulatory approaches, suggesting that alignment in practice must be user-specific, use-case-specific, and organization-specific. This paper demonstrates how model risk management practices constitute a pragmatic instantiation of alignment mechanisms and argues that the technical alignment research community should attend more closely to the governance innovations emerging from business AI deployments. We conclude by examining the deeper challenge that true AI-human alignment will require the development of AI empathy, which faces fundamental obstacles.