Designing August : Designing an Intelligent Agent for Financial Due Diligence in Real Estate Transactions

Sammanfattning

This thesis investigates how intelligent agents can be designed to support Financial Due Diligence (FDD) in real estate transactions. FDD is a document- and knowledge-intensive process in which the transaction function must review heterogeneous financial, contractual, and operational documentation under significant time constraints. This creates challenges related to information overflow, fragmented data sources, manual review, and the need for traceable analytical conclusions. The study addresses this challenge through an applied Design Science approach, resulting in the development and evaluation of August, a multi-agent Retrieval-Augmented Generation (RAG) system designed to support information retrieval, structuring, summarization, and traceability to support FDD decision-makers. August was developed in a cloud-based developer environment, in collaboration with a Swedish real estate company, Hemsö. August was implemented using a document pipeline, semantic retrieval, agent orchestration, and several foundation models. The development process was informed by interviews, user needs analysis, iterative implementation, pilot testing, and technical benchmarking. Two versions of August were evaluated: a single-agent architecture and a multi-agent architecture. In addition, several foundation models were benchmarked across dimensions such as retrieval and sourcing, analytical accuracy, completeness, explainability, usefulness, latency, and cost. The purpose of this was to profile the foundation models' capabilities and optimize multi-agent orchestration. The results show that August can support selected FDD tasks by retrieving financial and contractual information, structuring dispersed data into summaries and tables, identifying selected deviations and ambiguities, and providing source-grounded answers. The multi-agent version performed better overall in traceability, reasoning, presentation, and efficiency, while the single-agent version showed stronger numerical correctness in handling units across ambiguous contexts. However, the results also reveal limitations related to correct unit handling, model-dependent reliability, incomplete retrieval, and the need for human validation. The study concludes that agentic RAG systems can provide valuable decision support in document-intensive Due Diligence (DD) processes. August should be understood as support for human experts, not an autonomous decision maker.