This project explores the application of Generative AI in the tendering process for infrastructure projects. With the growing complexity of Dutch public procurement and the increasing demand for efficiency, AI-based text generation tools such as ChatGPT offer new possibilities for reviewing tender documents. The study evaluates both the capabilities and the limitations of large language models (LLMs). Stakeholder interviews and workshops with tender experts validated the practical relevance of AI tools and highlighted areas where human oversight remains essential. The project concludes with a discussion on ethical considerations, policy implications, and the potential future integration of generative AI into digital tendering platforms.

Here you can find the repository of the master thesis: Generative AI in Infrastructure Tendering

Project Information

• TItle: Generative AI in Infrastructure Tendering: A Study on Human-Generative AI Deliberation
• Author(s): Roos van Duuren
• Year: 2025
• Link: https://pure.tudelft.nl/admin/files/244439557/Thesis_Final_RoosvanDuuren.pdf
• Type: MSc Thesis, Construction Management and Engineering
• ML Tags: Large Language Models, LLMs
• Topics Tags: Generative AI, Tendering, Human–AI interaction, Infrastructure projects, Activity Theory, Digital construction
• Teachers Involved: Tong Wang, Ranjith Kuttantharappel Soman, Marian Bosch-Rekveldt

Generative AI in Infrastructure Tendering 1/1

Technical Aspectslink copied

Software & Plug-ins Used

OpenAI ChatGPT-4o, Microsoft Word (for tender templates) and qualitative analysis tools such as Atlas.ti for analyzing interview transcripts.

Design Workflow

Although not embedded in an architectural workflow, the ML approach supports text-based review decisions in tendering. It simulates document review stages commonly used by engineering and consulting firms.

ML Workflow

The project used prompt engineering and scenario testing with the GPT-4o model. A corpus of anonymized Dutch tender documents was used to construct example prompts. The workflow involved:

1. Inputting tender prompts into GPT-4o
2. Comparing with human-authored texts/judgements
3. Analyzing output quality and consistency
4. Iterating prompt strategy to reduce hallucinations and encourage deliberations
5. Validating results with domain experts through after experiment discussions and workshops