AI Governance: Beyond Governable to Admissible
This article challenges the common assumption that if an AI workflow can be governed, it should be automated. It introduces the concept of "admissibility," arguing that the true question is whether the value created by AI automation justifies the extensive governance burden required to make it acceptable and safe. The author emphasizes that a workflow can be technically governable but operationally unjustifiable due to the high costs associated with ensuring trustworthiness and compliance, especially in environments with weak information.
The Overlooked Cost of AI Governance
Internal audit and assurance professionals often focus on whether an AI system can be governed, meaning if controls can be implemented, biases measured, and outputs explained. However, this article posits that this perspective is incomplete. The critical, often-missed question is not just if AI can be governed, but whether the economic value derived from automating a workflow outweighs the significant costs associated with its governance. These costs include the time, expertise, validation efforts, and ongoing maintenance required to ensure the AI system is trustworthy, compliant, and operationally sound. Without considering this "governance economics," organizations risk investing in AI initiatives where the burden of making them safe and acceptable negates any potential benefits.
The Challenge of Weak Information Environments
The article highlights that the issue of governance economics becomes particularly acute in environments characterized by fragmented and unreliable information. In such scenarios, human users historically compensated for data inconsistencies, but AI systems lack this inherent discernment. Consequently, significant additional governance mechanisms are needed to enable AI to distinguish authoritative information from mere data availability. Each control added to address issues like weak provenance, missing ownership, or conflicting sources introduces friction and increases operational costs. This escalation of governance requirements can quickly lead to a point where the organization spends more on making an AI-driven workflow safe than the workflow itself is worth, rendering it economically unviable despite being technically governable.
Introducing the Admissibility Gate for AI Initiatives
The author introduces the crucial distinction between "governable" and "admissible." A workflow is governable if enough controls can be built around it, but it is only admissible if the value generated by automation justifies the governance burden. This calls for an "admissibility gate" in AI governance frameworks, which would prompt organizations to ask: "Does the expected value of automation exceed the governance burden required to make it acceptable?" This shift in mindset from merely achieving compliance to evaluating the economic viability of governance is paramount for enterprise AI. It encourages a more holistic assessment, potentially leading to decisions where automation is deemed unsuitable if the costs of ensuring its trustworthiness and reliability outweigh its benefits, thereby preventing costly, well-intentioned mistakes.
Read more
