Design and Development of an Intelligent Governance and Compliance Framework for Prompt Engineering and AI Output Validation in Regulated ERP Environments Using Generative AI

Abstract

In enterprise resource planning (ERP) environments, generative AI is becoming more and more part and parcel of the work, and while AI prompts are useful, the risks of being triggered, sensitive to transactions, and plagued by unreliable outputs can compromise accountability and operational control. In this study, this was a proof-of-concept, script-based governance framework that encompasses the operationalization of AI-output validation (hallucination-risk assessment), prompt-injection detection, and ERP transaction-risk classification. Three secondary datasets were combined into 5,000 synthetic ERP–GenAI scenarios, which were then fed into specific machine-learning modules and a deterministic decision engine that yielded Approved, Human Review, or Rejected results with structured audit reasons. Source label and pre-defined governance rules were used to generate reference decisions. It attained the following results: exact agreement rate of 76.96% with these decisions, unsafe detection rate of 90.89% and unsafe approval rate of 9.11%. This produced 1,377 Approved, 1,386 Human Review and 2,237 Rejected outcomes. Error analysis revealed that the largest number of unsafe approvals were due to missed prompt attacks (76.72%), followed by missed hallucinations and ERP-risk errors. The results prove multi-layer governance to be achievable, incorporating transaction context, prompt integrity, output risk, escalation logic and auditability. The results, however, show a synthetic integration benchmark, and are NOT a confirmation of regulatory conformity or production readiness. Better detection, ERPS based verification and independent risk labelling is needed before operational deployment.