The Opaque Transform: AI Governance Through Boundary Control and Management Review

Conventional assurance over financial reporting assumes a decision can be traced from input to output. When a neural network sits inside a material process, that trace is unavailable, an audit-trail void this paper calls the opaque transform. The paper proposes governing such systems through boundary control and management review. For the two COSO generative AI capability types where opacity and stakes are highest, autonomous orchestration and judgment or forecasting, it argues that the management review control should be primary assurance rather than compensating assurance, a risk-calibrated inversion of the traditional control hierarchy. A materiality threshold limits the full regime to uses that are both of those capability types and material, tested quantitatively, qualitatively, and by reliance, with process-level aggregation and a reassessment trigger. Below the threshold, governed autonomy and automated controls apply. The framework adds deterministic secondary validation, pre-action authorization gates for agentic systems, a fourteen-item risk taxonomy, a tiered control model, and a maturity model, and maps them to COSO's 2013 Internal Control-Integrated Framework and its 2026 generative AI guidance. It also reviews the 2026 regulatory and standards landscape. The paper is a conceptual framework, not an empirical study, and has not been peer-reviewed. It is offered for review and comment. Developed with AI assistance; see the disclosure on page 2. Status. Version 6.0 is final and is the version of record. Its text will not change. Later versions will carry a new version number. This record contains the same PDF first published on LinkedIn on October 19, 2026. Request for comments. This paper is published for review and comment, not as settled doctrine. Challenges to the materiality threshold, examples, corrections with a source, and proposals for inclusion are welcome at nate.lynch@gmail.com or via https://www.linkedin.com/in/nathaniel-lynch/. Please do not include confidential or employer-proprietary information. AI assistance. Developed in collaboration with Claude (Anthropic), most recently Claude Fable 5.1, for research, drafting, and iterative refinement across six versions. The author selected, arranged, edited, and approved the content, made all conceptual decisions and final positions, and is responsible for it. Review. An earlier draft was reviewed by an external practitioner, credited by name in the PDF. This paper has not been peer-reviewed. Scope. A conceptual framework for discussion. It is not legal, audit, accounting, or regulatory advice. Regulatory and market information is current as of October 19, 2026. Licensed under CC BY 4.0: https://creativecommons.org/licenses/by/4.0/