AI-adopting firms have lower distress risk and clearer predictive patterns

Traditional financial distress prediction relies heavily on backward-looking financial indicators such as leverage, liquidity, profitability and cash flow. While these variables have been the backbone of risk modelling for decades, they fail to capture intangible changes in firms undergoing digital transformation. AI adoption represents one such intangible factor, reflecting how capable a firm may be at operational optimization, process automation, innovation and long-term strategic adaptation.

A major new analysis of Chinese non-financial firms has revealed that artificial intelligence (AI) adoption inside companies can significantly improve early-warning models designed to detect financial distress, offering regulators, lenders and investors a more reliable tool for identifying firms at risk of failure. The findings arrive as global financial markets grow increasingly data-driven and machine learning becomes central to modern credit risk assessment frameworks.

The research, “Does Firm-Level AI Adoption Improve Early-Warning of Corporate Financial Distress? Evidence from Chinese Non-Financial Firms,” published as a working paper, evaluates the predictive power of firm-level AI activity using one of the most extensive datasets of Chinese listed companies to date. The study examines whether disclosures about AI usage and AI patenting can enhance distress prediction models beyond traditional financial ratios.

Across a 15-year window from 2008 to 2023, the researchers assess machine-learning models with and without AI-related variables and conclude that AI adoption consistently strengthens the sensitivity of distress forecasts. The findings challenge long-standing assumptions that accounting ratios alone provide sufficient insight into vulnerability, suggesting instead that technological readiness may be a new and increasingly important indicator of corporate resilience.

AI disclosures deliver critical signals as machine-learning models detect distress earlier and more accurately

Traditional financial distress prediction relies heavily on backward-looking financial indicators such as leverage, liquidity, profitability and cash flow. While these variables have been the backbone of risk modelling for decades, they fail to capture intangible changes in firms undergoing digital transformation. AI adoption represents one such intangible factor, reflecting how capable a firm may be at operational optimization, process automation, innovation and long-term strategic adaptation.

To evaluate this potential, the authors assemble a dataset covering thousands of firm-years and construct multiple AI indicators. These include AI-related text disclosures extracted from annual reports, AI patent activity and measures of AI density normalized across reporting length. The resulting features capture both the presence and intensity of AI adoption within firms.

The study tests six widely used machine-learning models: XGBoost, LightGBM, Random Forest, Logistic Regression, Neural Networks and Support Vector Machines. Each model is trained across several configurations, once using only traditional financial predictors and again with the addition of AI-derived variables. Because AI adoption is sparse in earlier periods and grows over time, the researchers introduce a methodological innovation known as a pruned training window. This approach shortens historical training periods while keeping the test year constant, allowing the team to examine how the importance of AI features evolves as they become more common among firms.

Across nearly all models, the introduction of AI variables results in a measurable improvement in identifying financially distressed firms. The most significant gains appear in recall and G-Mean metrics, two measures highly sensitive to rare events such as distress. The improvement indicates that AI variables allow models to capture subtle differences between distressed and healthy firms that financial variables alone may overlook.

Tree-based ensemble models show the strongest performance gains. LightGBM, XGBoost and Random Forest consistently outperform other algorithms when AI features are present, reinforcing their ability to detect non-linear relationships between technological indicators and financial health. The results highlight that AI adoption is not merely a neutral disclosure but a meaningful signal of underlying operational capability.

Despite these improvements, the inclusion of AI comes with trade-offs. While sensitivity increases, specificity declines modestly. This means models become better at catching distressed firms but also issue more false alarms. From a regulatory standpoint, this trade-off may be acceptable, particularly in market environments where missing a distressed firm is costlier than erroneously flagging one.

Healthy firms adopt AI earlier, more consistently and with greater innovation intensity

The study observed disparity between healthy and distressed firms in their AI adoption patterns. Healthy firms begin adopting AI earlier and at a steadier pace, demonstrating smoother growth in AI content within reports and more consistent patenting activity. Distressed firms not only adopt AI later but also show more volatile and irregular patterns of engagement.

The research suggests that AI adoption reflects both a firm’s technological investment capacity and its long-term strategic orientation. Healthy firms often have the financial flexibility, managerial bandwidth and strategic vision necessary to experiment with advanced technologies. Distressed firms, by contrast, typically face short-term operational pressures that constrain innovation budgets, leaving them lagging behind in adopting efficiency-enhancing systems.

AI-related disclosures also provide additional insight into the confidence and transparency of firm leadership. Companies with higher AI density tend to demonstrate clearer communication strategies and a willingness to document technological transitions. Distressed firms, on the other hand, appear less likely to highlight or invest in AI initiatives, whether because of budget constraints or limited internal capabilities.

When integrated into machine-learning models, these AI indicators improve classification stability. Models incorporating AI show less erratic behavior and a smoother distribution of predicted labels. The researchers attribute this stabilization to the additional variance AI features capture—particularly variation linked to firm innovation, adaptability and managerial sophistication. AI appears to serve as a structural complement to conventional financial distress indicators.

The study also analyzes how the performance of distress models changes under different training windows. Shorter, more recent windows may seem intuitively appealing because they mirror the contemporary relevance of AI. Yet the findings demonstrate that distress models retain higher performance when trained on longer historical datasets. The essential conclusion is that AI cannot replace the informational value of traditional financial signals; it can only enhance them. Training on longer windows balances both strengths, producing models that are more robust across economic cycles.

AI indicators strengthen distress forecasting across economic shocks and policy-driven environments

The study includes extensive robustness checks designed to evaluate whether the benefits of AI adoption persist under different macroeconomic and institutional conditions. The researchers test alternative distress years, including periods dominated by external shocks such as the COVID-19 pandemic, policy tightening cycles and broader economic uncertainty.

The results consistently show that AI-enhanced models degrade less during turbulent periods. In stress years, the historical patterns that guide financial ratio analysis become less stable, making distress harder to detect. AI variables, by contrast, remain predictive even when financial data becomes noisy or volatile. This resilience strengthens the argument that technological readiness serves as a useful stabilizing indicator for long-term firm viability.

The research additionally acknowledges China’s distinctive regulatory and industrial environment. The authors observe that China’s top-down, policy-driven AI ecosystem influences firm behavior. Companies sometimes adopt AI strategically to signal compliance with national innovation objectives or to seek government incentives. As such, AI-related disclosures reflect not only technological capability but also responsiveness to institutional expectations.

This dual nature of AI adoption presents important implications. In policy-driven contexts, AI engagement may reveal a firm’s adaptability to regulatory trends, another factor relevant to long-term solvency and survivability. The study points up that analysts should consider institutional settings when extrapolating results to other economies, as the motivations for AI adoption may differ across countries.

The authors call for future research to evaluate AI adoption in different institutional settings and across more diverse industries. They also recommend expanding the study to examine how firms’ internal AI capabilities, beyond disclosure, affect operational resilience and long-term performance.