# Can AI Understand a Distressed Credit? - Harneet Kaur, Axim Intelligence

## Executive summary

This talk explores the frontier of using AI to understand distressed credit, a complex financial domain. The speaker, Harneet Kaur, details how traditional scoring models (like the Altman Z-score) are insufficient for modern, opaque markets. The core methodology involves building a structured, multi-layered reasoning process (Extraction, Derivation, Judgment, Prediction). A case study using iHeartMedia demonstrated that while top models (Astra 6, Opus 5.5) initially disagreed on future probabilities, forcing them through this structured framework led to a convergence of their predictions, suggesting that systematic process design is critical for achieving consensus in complex financial modeling.

## Key takeaways

- Structured Reasoning Drives Model Convergence: When models are subjected to a structured reasoning process—benchmarking each layer (Extraction, Derivation, Judgment)—they converge on similar probabilistic outcomes, even when predicting non-verifiable future events. This suggests that systematic process design is key to achieving consensus in complex analysis.
- Distress is Multi-faceted, Not Binary: Distress is not simply predicting bankruptcy; outcomes can include liability management exercises or out-of-court restructuring. The analysis must consider financial, legal (e.g., structural priority of debt), and operational issues.
- The Limitation of Prediction: The speaker notes that even with structured context, the models do not have a predictive edge over human agents, as the future remains non-verifiable. However, the convergence itself is a valuable systemic outcome.

## Technical details

- Distressed Credit Definition: Credit is a bond or loan of a corporation. Distress occurs when underlying risks threaten the company's ability to pay back debt, caused by cyclical issues (e.g., construction market downturn) or secular disruption (e.g., AI disruption).
- Traditional Scoring Limitations: Traditional methods, such as the Altman Z-score, are becoming static and insufficient due to the rise of private credit and lack of pricing transparency in the market.
- AI Reasoning Layers: The proposed systematic analysis breaks down understanding into four layers: 1) Extraction (data points like EBITDA); 2) Derivation (calculating metrics like leverage); 3) Judgment (assessing if the credit is distressed); and 4) Prediction (determining future probability distributions).
- Case Study Models: The analysis used iHeartMedia (a company that filed for bankruptcy in 2018 and had a stressed transaction in 2024) and compared the outputs of Astra 6 and Opus 5.5.

## Practical implications

- The process of building a structured, multi-step reasoning pipeline (Extraction $ ightarrow$ Derivation $ ightarrow$ Judgment $ ightarrow$ Prediction) can be applied to systematize complex, manual human judgment in finance.
- For build engineers, this suggests a pattern for designing AI agents that enforce sequential, verifiable steps rather than relying on single-pass prompts, improving reliability and consensus.

## Topics

Distressed Credit, Financial Modeling, Large Language Models (LLMs), Structured Reasoning, Financial Engineering, AIE CODE summit

Source: https://www.youtube.com/watch?v=WOszaC1WWbQ