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

A company can struggle to repay its debt without ending up in bankruptcy. That distinction makes distressed credit a harder AI problem than a yes-or-no default prediction. Harneet Kaur, founder and CEO of Axim Intelligence and a former distressed credit investor, begins with the object being analyzed: a corporate bond or loan. Coca-Cola’s bond is a credit; the difficult question is what happens when the company behind a credit comes under pressure.

Stress means an underlying business risk could leave the company without enough money or unable to repay its debt. The causes differ:

• Cyclical pressure: A construction business faces a downturn in construction or housing.
• Operational trouble: A restaurant business faces a major food recall.
• Secular disruption: A software business faces a lasting competitive change, such as disruption from AI.

Some companies recover from these challenges. Others move into distress.

Even then, bankruptcy is only one possible outcome. Kaur also names liability management exercises, out-of-court restructurings, and negotiated exchanges. These alternatives matter because understanding a troubled borrower requires more than detecting trouble: the analysis must distinguish the paths the company could take from there.

The older engineering formulation was largely binary: predict whether the borrower defaults. Kaur’s objection is that this compresses a changing, multi-path situation into too narrow a question. She singles out the Altman Z-score as an example of scoring that she considers too static for situations that can change monthly, quarterly, or yearly.

The information environment creates another difficulty. In Kaur’s account, the growth of private credit makes parts of the market more opaque, with less pricing information available. A score does not remove the need to understand what is happening inside the business and among its lenders.

The alternative is the investor’s manual investigation: analyze financial statements, understand the lender side, and read the credit documents. This can accommodate details that a simple score misses, but it is labor-intensive and hard to scale. Axim’s motivating question sits between those approaches: how much of that complex analysis can AI systematize while preserving the distinctions between possible outcomes?

Understanding a credit draws on several kinds of analysis:

• Financial condition: What liquidity does the company have, and what does its capital structure look like?
• Legal position: Which debt has higher or lower structural priority?
• Business performance: What are revenues doing, and what does fundamental analysis suggest about whether they will rise or fall?

These are different questions. Reading a financial figure correctly does not, by itself, resolve legal priority or the direction of the business.

Each question can become a benchmark task. For example, a model can be tested on whether it identifies which debt is structurally senior. Kaur wants to push beyond that kind of component test. Her proposed test of understanding is “can you predict what happens next”: produce a probability distribution over the company’s future paths, rather than stop at a collection of individually correct answers.

The test case is iHeartMedia. Kaur describes a company that filed for bankruptcy in 2018 and undertook another stressed transaction in 2024. At the time of the experiment, its next outcome remained unresolved. That makes it a prospective forecasting problem: the models must assign probabilities before anyone can check the eventual result.

What happens when two frontier models receive the same organized context? They rank the paths differently. One favors negotiated restructuring; the other favors bankruptcy. The important comparison is the distribution over outcomes, not merely whether both recognize that the company is troubled. Shared input has not produced a shared assessment of what comes next.

Kaur personally leans toward the restructuring-favoring forecast, but that preference cannot settle the test. Bankruptcy, out-of-court restructuring, and refinancing are still possible futures. An investor’s best judgment can help examine the assumptions; it cannot supply an already-known answer to an unresolved event. The disagreement therefore redirects the experiment toward something that can be inspected now: the process behind the forecast.

The next intervention separates the analysis into layers that can fail in different ways. Instead of judging only the final probability distribution, Kaur benchmarks the intermediate work and examines whether its assumptions make sense. The concrete teaching example follows financial inputs into a derived measure, then into a judgment about distress.

• Extraction: Obtain the underlying financial data, such as EBITDA. Kaur considers extraction from public-company material largely solved for her purposes, while private loans, Excel documents, and PDFs can still present challenges. The distinction matters: a model cannot reason its way around an incorrectly captured input.
• Reconciliation or derivation: Use the extracted inputs to produce another quantity. In her example, EBITDA and debt provide the inputs for leverage. This is a separate check from extraction: having the source figures is not the same as correctly deriving the measure used in analysis.
• Judgment: Interpret the derived quantity. Given the leverage, does the company qualify as a distressed credit? This step turns a financial measure into an assessment of the situation.

Prediction comes after those checks. The intended dependency is straightforward: sound data supports sound derivation; checked derivation supports a reasonable judgment; that judgment then informs probabilities over future paths. These are information dependencies, not interchangeable prompts. A plausible-looking forecast can inherit a mistake made much earlier.

The point of separating the layers is to locate what deserves scrutiny. If EBITDA was captured incorrectly, the problem begins with extraction. If the inputs are right but leverage is wrong, the problem lies in derivation. If the measure is right but its interpretation is unreasonable, the judgment needs examination. Kaur’s warning is simple: a judgment built on a bad data point is not a good judgment. Checking upstream work gives the prediction a more defensible basis without making the future certain.

How much did the forecasts change after introducing the structured reasoning process? Kaur reports the following bankruptcy estimates:

Both models then favor negotiated restructuring as the more likely path. The comparison makes the asymmetry clear: one model’s reported bankruptcy estimate stays at 25%, while the other moves from 62% to 25%.

This is a change in the analysis process, beyond the organized context both models already received. It produces agreement on the reported bankruptcy probability and on the leading path. The experiment does not identify a particular corrected input or assumption that caused the shift, so the result supports a process-level observation: the checked, layered approach brought these two forecasts together.

Does agreement establish a predictive edge? Kaur’s answer is no. With the future still unresolved, the experiment has not shown that either model predicts better than a human working from the same context. Its demonstrated result is narrower: a structured process can bring different models to a similar understanding of the business and its possible paths.

The remaining distinction is consequential: convergence is observable now; forecast correctness is not. Kaur also raises look-ahead bias as a concern with historical analysis. Looking backward can let information about what eventually happened influence an assessment that is supposed to represent what could have been known beforehand. A live case avoids having an already-settled outcome to lean on, but it leaves the forecast unverified at the time of the talk.

The ending gives convergence a different possible use. If a repeatable analytical process brings different assessments toward the same probabilistic outcomes, Kaur sees an analogy to a market price: a common assessment can emerge even while the future remains uncertain. That is a proposed interpretation of the convergence, not a demonstrated pricing model. The interesting engineering result is that the models’ disagreement changed when the work between the documents and the forecast became explicit.