Prompt engineering in finance : an LLM-based multi-agent architecture for decision support

Abstract

PURPOSE: This article systematizes prompt engineering techniques for finance and proposes a novel, LLM-based multi-agent architecture to enhance the quality, reproducibility, and auditability of automated decision support. The goal is to create a framework that ensures outputs are not only accurate but also transparent and compliant with stringent regulatory requirements.

DESIGN/METHODOLOGY/APPROACH: This is a conceptual paper that puts forward a decision-support architecture built on four pillars: (i) enforced multi-path reasoning through Chain-of-Thought (CoT) and self-consistency; (ii) grounding outputs in a curated corpus via Retrieval-Augmented Generation (RAG); (iii) a structured dialogue between agents with specialized roles (Strategist, Critic, Moderator); and (iv) strict verification guardrails and auditable output formats. The methodology integrates proven techniques into a coherent system designed for financial applications.

FINDINGS: Since the study is conceptual, the findings are presented as operational hypotheses supported by existing literature. The proposed architecture is expected to yield: (A) greater accuracy and stability of reasoning; (B) a significant reduction in hallucinations and improved provenance due to RAG; (C) more effective detection of errors and analytical blind spots through the multi-agent workflow; and (D) increased auditability and regulatory readiness via standardized, verifiable outputs.

PRACTICAL IMPLICATIONS: The framework offers financial institutions a clear path toward developing more reliable AI tools. Its implementation can lead to higher quality "first drafts" of decisions, fewer subsequent corrections, and shorter audit cycles. This approach has direct applications in areas such as credit analysis, risk management, and compliance monitoring, promising faster processing with more robust documentation.

ORIGINALITY/VALUE: The paper’s main contribution is the synthesis of several distinct lines of LLM research-prompt engineering, RAG, and multi-agent systems-into a single, coherent architecture tailored to the specific needs of the financial sector. It addresses a critical gap by providing a systematic blueprint for building explainable and auditable AI decision-support systems in a highly regulated environment.