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

Masahiro Kato. 2026. Preprint (arXiv:2606.20041). 日本語版

AI Economist Agent is an agent framework for producing economic analysis. Rather than asking an LLM to write the analysis by itself, it combines retrieval, a knowledge graph, and explicit computation based on economic models. Planning the analysis, retrieving evidence, selecting a model, computing, and drafting the report are handled as separate roles, so economic claims tie back to source documents and computed results.

The problem

An LLM can write plausible prose about the economy. Analysis by an economist has to be backed by economic theory and by real data. When a reader cannot tell which claim in the generated text rests on which document and which calculation, the text does not work as analysis. The question this study sets itself is how to keep the fluency of generation and still make the evidence traceable.

The proposed method

Economic data and economic theory are held as a knowledge graph, and an LLM agent plans the analysis, retrieves the relevant evidence, picks a suitable model, and writes the report. Numerical claims never come from the language model directly. They are produced by explicit model computation, and the statements in the report are linked to the retrieved evidence. We call this whole arrangement the AI Economist Agent.

Main results and conditions

We evaluated the framework on two applications. The first generates an economist report on inflation persistence in the United States and Federal Reserve policy; the second generates the narrative for a bank stress test of refinancing stress in US commercial real estate. Tying the generated report to its evidence improved economic consistency and traceability. That evidence comes from two cases, so we do not claim it generalizes to economic analysis at large.

Relation to earlier work

RAG, which conditions generation on retrieved documents, was proposed by Lewis and coauthors (2020, Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks), and Edge and coauthors developed the use of a knowledge graph to handle the structure of a document collection with Graph RAG (2024, From Local to Global: A Graph RAG Approach to Query-Focused Summarization). For interleaving reasoning and acting in an LLM, the standard reference is ReAct by Yao and coauthors (2022, ReAct: Synergizing Reasoning and Acting in Language Models), and for calling external tools it is Toolformer by Schick and coauthors (2023, Toolformer: Language Models Can Teach Themselves to Use Tools). What is distinctive here is the way these pieces are assembled under a requirement specific to economic analysis, that is, support from theory and data, with numerical claims confined to explicit model computation.

Related research

Causality Elicitation from Large Language Models (Kameyama, Kato, Hio, Takano, and Minakawa, 2026) draws candidate causal relations out of what an LLM knows and organizes them in a form a person can check. It complements this framework from the side of the causal structure that an analysis presupposes.

Where it applies

The situation we have in mind is a team that wants AI to write market or economic reports but has to be able to explain what the conclusions rest on. Organizing internal documents and data as a knowledge graph and pairing them with model computation carries over beyond economics, to any analytical work where the reasoning has to be explained.

Paper and materials

BibTeX

@misc{ai-economist-agent,
  author       = {Masahiro Kato},
  title        = {AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models},
  year         = {2026},
  eprint       = {2606.20041},
  archivePrefix = {arXiv},
  url          = {https://arxiv.org/abs/2606.20041},
}