The AI System of Record: Governance for Agentic AI

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Agentic AI is being increasingly implemented in organizations and is becoming more autonomous. However, enterprise data remains fragmented, inconsistent, and potentially stale, while large language models (LLMs) are not completely deterministic. Moreover, governance, control, and traceability are often lacking when AI agents are directly connected to enterprise data sources. As a result, autonomous delegation with consistent and high-quality decisions remains difficult to implement and difficult to trust. This paper proposes the AI System of Record (AI SoR) as an architectural concept between AI agents and heterogeneous enterprise data sources. It provides a canonical state for decision-relevant data, including freshness and eligibility checks. The AI SoR further provides controlled tool interfaces and externalizes organization-specific rules and requirements into deterministic code. Its Policy & Approval Layer evaluates agent-proposed actions as permit, deny, or escalate. An immutable audit trail records relevant states, interactions, rules, and actions. The LLM itself does not become deterministic; instead, the environment in which the agent makes decisions becomes more controllable. The evaluation of the AI SoR is conceptual. The paper argues that the architecture can enhance consistency, traceability, and policy enforcement while introducing additional latency, implementation complexity, and cost. It does not eliminate hallucinations or incorrect underlying enterprise data. The AI SoR is therefore proposed as an architectural foundation for enterprise agentic AI systems that can be extended with organization-specific rule sets, workflows, and multi-agent mechanisms by translating organizational context, requirements, and implicit rules into explicit and enforceable system logic.