From Flow to Finished Report - Streamlining the Writing Process of Stormwater Investigations with Generative AI

Sammanfattning

Climate change and urbanization place high demands on well-planned and sustainable stormwater management, making reliable stormwater investigations increasingly important. These investigations often follow a similar structure and include several recurring elements, which makes them suitable for streamlining. The purpose of this study is to investigate how Ramboll’s AI tool, RamGPT, can be used to streamline repetitive parts of report writing in stormwater investigations, with the goal to achieve an accuracy of at least 70% - meaning that at least 70% of the AI-generated texts should be considered “good enough” according to a developed evaluation scale. To achieve the goal and answer the research questions, a literature study, interviews with stormwater investigators, and testing of prompts in RamGPT were conducted. The prompts were assessed using a criteria-based evaluation scale for objective assessment of content, structure, and language accuracy. The study focused on two parts of the report - the summary and geological conditions - as these were identified in interviews with Ramboll’s stormwater investigators as both time-consuming and suitable for standardization. With the help of theory and practical testing in RamGPT, prompt templates could be developed for both parts. The tests showed that RamGPT was able to generate text of sufficiently high quality to be used as a first draft for further development, achieving an accuracy of 73%, which exceeds the goal. Generative AI can be a valuable support in report writing as it increases efficiency by quickly providing a first draft and can process large amounts of data. Well-thought-out and clear prompts reduce the risk of misinterpretations and misunderstandings, but it is also important to be aware of the tool’s limitations, such as the risk of incorrect information. Therefore, manual review is always required to ensure the quality and accuracy of the content. In conclusion, the study shows that generative AI has the potential to streamline parts of the stormwater investigation process, especially when clear prompt templates are combined with manual review. There are also opportunities to further develop the method for additional parts of the reporting process.