By Joke Baeck and Stijn Van Ruymbeke (Ghent University)
Financial distress among companies poses a significant challenge to economic stability. Timely and effective intervention is essential. In Belgium, the Chambers for Companies in Difficulty (CCDs) within the commercial courts play a crucial role in detecting and addressing financial distress through both preventive and regulatory measures. However, the current manual selection process of companies at risk, based on so-called ‘red flags’, is resource-intensive and inconsistent across CCDs.
At Ghent University, we are exploring how artificial intelligence (AI) can assist CCDs in this task. In collaboration with the Commercial Court of Antwerp, we launched a pilot project to develop an AI-based decision-support tool aimed at improving the efficiency, consistency, and objectivity of the CCDs’ selection process. The goal is not to replace the judges of the CCD, but to support them in identifying those companies most in need of intervention.
The role of the chambers for companies in difficulty (CCD’s) in Belgian commercial courts
Each Belgian commercial court has at least one Chamber for Companies in Difficulty (CCD), tasked with identifying companies showing signs of financial distress. Once a company is identified, the CCD either encourages it to take recovery measures or, if no recovery is feasible, refers it to bankruptcy proceedings. The CCD is therefore sometimes described as an ‘economic hospital’, providing guidance and support for financially ‘ill’ companies. Where recovery is no longer possible, the CCD helps to ensure their orderly market exit.
To detect financially distressed companies, CCDs rely on a centralized digital database known as KNICLI (short for KNIpperlichten/CLIgnotants, the Dutch and French words for ‘red flags’). This database includes various red flags of potential financial distress, such as outstanding tax and social security debts, asset seizures, failure to file annual accounts and judgments terminating commercial leases.
It is not feasible for CCDs to investigate every company flagged in KNICLI. A selection is necessary, but this task is complicated by limited time and resources, as well as the sheer volume of flagged companies.
The problem: manual selection and inconsistent criteria
Despite the availability of the KNICLI database, CCDs still rely on manual processes to select companies for investigation. Selection criteria vary between CCDs and may even shift within the same CCD over time. For instance, one month a CCD may prioritize companies with high social security debts, and the next month, those with high tax debts. This ad hoc approach results in inconsistency, limits transparency, and creates the risk that companies in need of urgent intervention are overlooked or addressed too late.
The AI solution: a decision-support tool, not a replacement
Our research team is currently developing a machine learning model to assist the judges of the CCD in the selection process. The model is being trained on historical CCD decisions and associated red flags found in KNICLI. The ultimate goal is to build a model that can rank companies according to urgency and provides a prediction of the likely decision the CCD might take, such as whether the company should be further investigated, referred to bankruptcy, or asked to provide additional information.
It is important to emphasize that the model is still under development and not yet ready for deployment. Once operational, its recommendations will be non-binding. Judges will retain full discretion to accept, adjust, or disregard the model’s output. The tool is designed purely as a support mechanism, helping judges to make faster and more consistent decisions. It is I intended only for internal use within the CCD, given the sensitive nature of the data involved.
From theory to practice: the Antwerp pilot project
The development of the model began with a large-scale data collection effort. In collaboration with the CCD of the Commercial Court of Antwerp, which is the most active user of KNICLI, we began collecting data on flagged companies, those selected for investigation, and the decisions taken by the CCD for selected companies. Between March 2023 and February 2024, KNICLI recorded 9,657 unique companies in the Antwerp region. Of these, only 610 were selected for further investigation, resulting in 2,869 recorded decisions ranging from case closures to referrals for bankruptcy.
Although our dataset is not yet sufficient to train a fully robust model, it has allowed us to start developing and testing an initial prototype. This exploratory model focuses on predicting which companies are most likely to be referred to bankruptcy proceedings. By generating a ranked list of high-risk firms, it offers a preview of how the final AI tool could support CCDs in prioritizing their limited resources.
Challenges ahead: data, bias and compliance with AI Act
While our preliminary findings are promising, several important challenges must still be addressed before the model can be considered ready for practical use. One important limitation is data availability, particularly regarding financial information from annual accounts. Integrating such data will be a central focus in the next phase of development.
Another challenge lies in addressing potential selection bias. Our model is trained only on companies that were selected by the CCD in the past. This limits its capacity to identify overlooked but genuinely distressed firms (‘false negatives’). To mitigate this, we are exploring the integration of more general data on companies that went bankrupt, which could help the model learn from missed cases and improve its overall sensitivity.
Finally, the use of AI in a judicial context raises complex legal and ethical questions. As the model may qualify as a high-risk AI system under the EU’s recently adopted AI Act, it may need to comply with stringent requirements relating to transparency, human oversight, and accountability. In parallel with the technical work, we are conducting further legal research to ensure that any future implementation meets the requirements of the AI Act and contributes to trustworthy AI in the judiciary.
The broader implications: AI and insolvency law
Although this project is situated within the Belgian legal context, its relevance extends beyond national borders. The challenges faced by the Chambers for Companies in Difficulty (limited resources, heavy case load and the need for timely intervention) are shared by commercial courts in many jurisdictions. Our pilot demonstrates AI’s potential as a transformative tool in insolvency law by assisting judges in detecting financially distressed companies more effectively.
In insolvency law, early intervention is crucial. The sooner a financially distressed company is identified, the greater the chance of successful recovery, or, if recovery is not possible, an orderly market exit. An AI tool that helps prioritize urgent cases can make that process faster, more consistent and more resource-efficient.
Importantly, we do not view AI as a replacement for judges. Rather, we see it as a tool to support judges by processing large volumes of data and identifying those companies most in need of intervention. This allows judges to focus their time and expertise on more complex legal questions that still require human expertise.
As we continue to develop and refine our model, we believe its approach holds promise not only for the Belgian judiciary but also for any jurisdiction where timely intervention in cases of financial distress is a priority.
* For our full paper: see Baeck J., Arno H., Van Ruymbeke S., Audenaert A., Habils T., Mulier K. en Demeester T., ‘Developing an AI Model for the Detection of Financially Distressed Companies by Belgian Commercial Courts’, available at SSRN: https://ssrn.com/abstract=5333322 or http://dx.doi.org/10.2139/ssrn.5333322.
