Potential of LLMs in Supporting Call-taking Processes of Emergency Dispatch Centers

Abstract

In recent years, there have been major improvements in the development of large language models (LLMs). These models can process natural language and form the basis of many partially autonomous AI agents. Consequently, they have found their way into many areas of work, including medicine. However, the use of LLMs is associated with certain risks. Working in emergency dispatch and control centers presents particular challenges. Decisions must be made quickly in order to provide people with the right help in case of an emergency. LLMs could support human operators by helping them to evaluate emergency calls, triage them, and alert rescue services. However, the models must be highly reliable for this purpose. This study evaluates the ability of current LLMs (GPT 5.2 and Claude Sonnet 4.6) to classify emergency call transcripts according to their criticality using the ABCDE scheme. Different levels of detail are provided in the guidelines using different prompts.
The results show that the models perform well even without detailed guidelines. However, well-designed guidelines can significantly enhance performance. These offer a opportunity to adapt the evaluations to local conditions or continuously improve them during later real-world operation.
The study demonstrates a possibility to integrate LLMs in control center processes. Although the use of LLMs is certainly promising, it also involves numerous risks. Therefore, insecurity-critical fields such as emergency services, it is essential to carefully consider whether their use is justified.