AI, LLMs, and Economic Decision Making

• LLM agents
• RAG
• knowledge graphs
• AI-ready data

The question

A language model can produce an economic narrative in seconds. Whether that narrative can be used — cited in a report, defended in a committee, audited afterwards — is a different question, and it is the one this project works on.

An economist's claim carries an implicit chain: this number came from this series, transformed this way, fed into this model, under these assumptions. Free-form generation breaks that chain. The aim here is to rebuild it inside an agentic system, so that fluency and accountability are not traded off against each other.

Model-grounded analysis

The AI economist agent keeps economic data and theory in a knowledge graph and uses LLM agents to plan an analysis, retrieve evidence, choose a model, compute, and write. The design constraint that does the work is simple: quantitative claims are never produced by the decoder. They come from explicit model-based computation, and the report text is linked back to the evidence and the computation behind it.

The framework has been evaluated on economist report generation for U.S. inflation persistence and Federal Reserve policy, and on bank stress-test narrative generation for commercial real estate refinancing stress.

Knowledge infrastructure for AI

A second strand asks what it takes for documents, data, and algorithms to be AI-ready: organized so that a model can retrieve, cite, and reuse them reliably rather than paraphrase them. Ongoing work measures how the way an algorithm is written down affects the accuracy with which a language model implements it — a question that matters for any organization planning to put its internal methodology in front of an LLM.

Open directions

• Evaluation criteria for grounded economic narratives beyond text quality: coverage, traceability, and robustness to data revisions.
• How model selection should be delegated to an agent, and when it should not be.
• Representations of economic theory that are simultaneously machine-retrievable and faithful.