The Berufsgenossenschaft der Bauwirtschaft (BG BAU), Germany's statutory accident insurance institution for the construction sector, has developed and deployed a machine learning application to optimise the targeting of occupational safety inspections across the German construction industry. The project, formally titled 'KI-basierte Unterstuetzung fuer zielgenaue Unfallpraevention' (AI-based Support for Targeted Accident Prevention), was funded by the German Federal Ministry of Labour and Social Affairs (Bundesministerium fuer Arbeit und Soziales, BMAS) as a lighthouse project under the ministry's 'Digital Working Society' initiative. The system was developed using open-source technologies and trained on approximately 10 million data points drawn from inspector field reports, member company submissions, and internal departmental records. It uses a Random Forest algorithm, selected after the project team tested 60 different models during a nine-month development period from 2022 to 2023, to generate risk scores for BG BAU member construction companies nationwide.
The core challenge the system addresses is one of resource allocation under significant operational constraint. BG BAU employs approximately 500 supervisory personnel (Aufsichtspersonen) who are responsible for advising companies on occupational health and safety and conducting workplace inspections. These inspectors conduct approximately 25,000 company consultations per year, but can reach only a limited share of firms needing advisory support, and can provide on-site personal visits to only a small portion of the total membership. Prior to the AI system, inspectors relied on manual, sporadic review of data tables and spreadsheets to select which companies to visit, a process that was labour-intensive and inconsistent in its ability to identify the highest-risk firms.
The AI system generates a risk score between 0 and 1 for each company, where a score closer to 1 indicates a high need for advisory consultation. These scores are presented to inspectors through a traffic-light display system (red, yellow, green), enabling rapid visual prioritisation. The system is updated weekly, incorporating new data such as recently reported accidents, training and seminar participation by company employees, and updated field inspection reports. This weekly refresh cycle allows inspectors to detect emerging trends in accident patterns more quickly than was previously possible. The AI application identifies a high-priority group of companies each year that aligns with the annual consultation capacity of BG BAU's inspection workforce.
The system operates under a clear principle of human-in-the-loop oversight, summarised in the project's guiding phrase 'Die KI empfiehlt, der Mensch entscheidet' (the AI recommends, the human decides). Inspectors retain full decision-making authority over which companies to visit and what advice to provide. The use of the AI tool is voluntary, and the system is designed to support preparation for consultations rather than replace professional judgement. The project team conducted bias analyses to identify and minimise distortions in the training data. For example, accident counts and deficiency records are normalised by employee count within each company to prevent systematic over-flagging of larger firms. The system continuously monitors outputs for potential discriminatory patterns. The project was developed within the framework of self-binding AI guidelines (KI-Leitlinien) created by the Network AI in Labour and Social Administration (Netzwerk KI in der Arbeits- und Sozialverwaltung), a collaborative body of over 40 experts from 20 government agencies established under the aegis of the BMAS. These guidelines require that AI applications in government administration be reliable, non-discriminatory, transparent, and auditable.
The reported results show a substantial improvement in targeting efficiency. The AI system increased the hit rate for identifying companies with high consultation needs to 64 percent, representing a 29 percentage point improvement over previous manual methods. According to Michael Kirsch, Chief Executive of BG BAU, this improvement represents a major efficiency gain for inspectors' work. The project was recognised internationally, receiving the ISSA Good Practice Award for Europe 2024 at the International Social Security Association conference in Porto under the Vision Zero campaign for workplace accident prevention. It subsequently won the Public Leadership Award 2024 in the Leadership and Digital Transformation category at the 10th Future Congress for State and Administration in Berlin in June 2024, and was nominated for the Working on Safety (WOS) Award and the Young Scientist Award, both administered by the German Social Accident Insurance (DGUV).
The project team comprised staff from BG BAU, BG-Phoenics GmbH (a specialist IT subsidiary), and Accenture SE, with academic input from the Hertie School. The project leader was Ellen Hellmann, Head of Digital Transformation and Corporate Development at BG BAU. The BMAS provided approximately 3.5 million euros in funding. The project officially ran from February 2023 to June 2024, with operational rollout underway in 2024. BG BAU continues to operate and refine the system beyond the formal project conclusion, collecting ongoing feedback from supervisory personnel to improve the model. The institution has also presented the project internationally at the World Congress on Safety and Health at Work in Sydney in November 2023, establishing connections with counterparts in Australia, India, and China. BG BAU is now developing a follow-up project, the KI-gestuetzte Besichtigungsassistenz (BeA), an AI-supported inspection assistant developed in collaboration with six other Berufsgenossenschaften, which incorporates speech recognition and AI-based defect identification to further support inspectors during on-site visits.