AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models

Research Approach

This paper proposes an 'AI economist agent' architecture designed to bridge the gap between fluent narrative generation and rigorous economic analysis. While standard Large Language Models (LLMs) can produce plausible-sounding economic reports, they often lack the ability to link claims to underlying data, theory, or formal models. The proposed framework addresses this by using AI agents to plan analyses, retrieve evidence from knowledge graphs, select appropriate mathematical models, and execute those models to generate numerical outputs. The final report is then constructed by synthesizing these model-based computations with retrieved evidence, ensuring that quantitative claims are traceable to specific model runs.

Methodology

The system utilizes a knowledge graph to separate text-based evidence from formal model specifications. The architecture follows a multi-step workflow:

1. Planning: An agent reads the economic scenario and creates an analysis plan.
2. Retrieval: The system performs both text-based RAG (Retrieval-Augmented Generation) and graph-based retrieval to identify relevant evidence and candidate model specifications.
3. Execution: The agent selects a model, which is then executed by a computational layer outside the LLM.
4. Synthesis: The system records the model run and outputs back into the graph, allowing the final report generator to reference the specific model, its assumptions, and the resulting numerical data.

Key Findings

The framework was evaluated through two applications: U.S. inflation persistence and bank stress-test narratives for commercial real estate. The results demonstrate that the model-grounded GraphRAG approach significantly improves the coherence and traceability of reports compared to LLM-only or RAG-only methods. By forcing the system to execute formal models rather than hallucinating numerical paths, the framework ensures that economic narratives are tethered to verifiable computational results.