A Conceptual Study for Cognitive Bias Amplification in Agentic AI-Driven Business Processes, Management, and Intelligence

Agentic artificial intelligence (AAI) and retrieval-augmented generation (RAG) are increasingly embedded in organisational business intelligence (BI) and business process management (BPM). Unlike conventional AI, these systems set sub-goals, plan multi-step workflows, and retrieve and store information under limited human supervision. This article argues that such autonomy not only transmits human cognitive bias into organisational decisions but also amplifies that bias under identifiable conditions. We develop the Bias Amplification Model (BAM), a three-layer account of how bias enters and escalates. In the injection layer, human biases enter through goal framing, prompt design, and data scoping. In the propagation layer, the agent compounds these biases across autonomous execution steps. In the crystallisation layer, RAG memory encodes biassed outputs as retrievable organisational knowledge that later cycles treat as evidence. A feedback loop links the three layers, converting episodic human bias into structural organisational bias. We state seven propositions, specify amplifying and mitigating conditions, and ground the model in a PRISMA-guided synthesis of 47 studies. A controlled, replicated experiment with an autonomous agent in a supplier-selection pipeline provides initial empirical support for injection, propagation, the mitigating effect of a single governance checkpoint, and bias transmission across a task boundary via retrieved memory. The framework reframes AI-driven process optimisation as a possible source of silent decline in decision quality.