AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models
Preprint, 2026.
- LLM agents
- RAG
- knowledge graphs
- economic analysis
In one sentence. An agentic framework that lets a language model write economist-style reports while every quantitative claim is produced by an explicit model computation and linked back to a retrieved source.
Why this problem
Large language models write fluent economic prose. That is precisely the difficulty: an economist's claim is only worth what the theory and the data behind it are worth, and free-form generation leaves no audit trail. A number that appears in a paragraph cannot be traced back to a series, a vintage, or a model, so a reader cannot tell an inference from a plausible-sounding sentence.
The question this paper asks is therefore not "can an LLM write like an economist?" but "can an LLM's output be made traceable in the way an economist's analysis is expected to be?"
The framework
The AI economist agent keeps economic data and economic theory in a knowledge graph, and uses LLM-based agents to plan an analysis, retrieve the relevant evidence, select an appropriate model, run it, and write the report. The division of labour is the point of the design:
- The language model plans, retrieves, and writes.
- Quantitative claims are not generated by the language model. They come from explicit model-based computations.
- The narrative is linked back to the retrieved evidence and to the computations that produced each number.
Technically the system builds on retrieval-augmented generation, graph-structured retrieval, ReAct-style reasoning, and tool calling. The contribution is the domain-specific discipline imposed on top: in economics, the model — not the decoder — is allowed to assert numbers.
Two applications
- Economist report generation on U.S. inflation persistence and Federal Reserve policy.
- Bank stress-test narrative generation for U.S. commercial real estate refinancing stress.
In both cases, grounding the generated report in retrieved evidence and explicit computation improves its economic coherence and its traceability.
Takeaway
For applied economics and finance, the useful bar for an LLM system is not fluency but auditability. Restricting the model's authority — it may reason, retrieve, and explain, but not invent numbers — turns a generative system into something a practitioner can check line by line.