AI-Supported Objection Management in Public Participation: Concept, Prototype and Evaluation in the Context of Infrastructure Projects

Public participation is a central component of democratic decision-making processes, particularly in planning and approval procedures. However, increasing data complexity and the growing number of submitted objections significantly raise the effort required for their review and processing. Against this background, this study developed an AI-supported objection management system that uses a large language model (LLM) to automatically pre-sort objections by topic and generate response suggestions based on historical objection texts from previous infrastructure projects. The aim is to increase efficiency in the processing workflow while maintaining consistent response quality without replacing human decision-making. The prototype development is preceded by a literature review to identify key user requirements and derive relevant use cases. Subsequently, four expert workshops with representatives from German road and rail infrastructure administrations at the state and federal level were conducted to evaluate the prototype. The results indicate significant efficiency potential, particularly through automated thematic pre-sorting of objections. However, topic structures must be adapted to the specific procedure. AI currently mainly serves as supportive pre-processing and requires human review (“human-in-the-loop”). Transparent labeling of AI use is also necessary to ensure traceability and acceptance. The findings will be incorporated into the ongoing development of the prototype within the BIM4People research project funded by the German Federal Ministry of Transport (BMV), with the aim of further improving the system’s functionality and exploring additional applications.

In 2025, the German government approved a comprehensive investment package in the form of a special fund amounting to €500 billion to promote infrastructure development and achieve climate neutrality. This special fund is expected to lead to a significant number of complex infrastructure projects being initiated, planned, and implemented in the coming years. What all these future large-scale projects have in common is that they typically have to undergo formal approval procedures that include public participation formats. In this context, so-called planning approval procedures constitute a central mechanism for the authorization of major infrastructure projects.

Within formal planning and approval procedures that include public participation, citizens are provided with opportunities to take part in the decision-making process within defined timeframes. During these phases, individuals whose interests may be affected by a proposed project can express their views and raise concerns, for example, by submitting objections or statements regarding the plans. These contributions are typically collected during officially designated consultation periods and are considered as part of the overall evaluation of the project. Given the anticipated increase in large-scale infrastructure developments, it is likely that the number of public submissions and objections within such planning and approval procedures will also grow significantly. This poses considerable challenges for hearing authorities and case workers, as even in current procedures, processing and responding to objections requires an extremely high investment of time and resources.

Recent advances in artificial intelligence (AI), particularly large language models (LLMs), have led to growing interest in their application for information processing, summarization, and decision-support tasks. In the context of public administration and knowledge-intensive workflows, LLMs are increasingly explored as tools to support structured text analysis and semi-automated reasoning. However, existing research also highlights that the deployment of such systems raises important challenges related to transparency, accountability, and institutional trust. Hernández and Rockembach (2025), for example, emphasize that the successful integration of AI systems in records and archival environments depends not only on technical performance, but also on organizational AI literacy and the ability of institutions to ensure responsible and comprehensible use of AI-generated outputs. Similarly, González-Quesada et al. (2025) demonstrated the potential of LLM-based approaches for modeling large-scale group decision-making processes, particularly through the clustering and structuring of textual input. While these findings indicate promising capabilities for supporting complex information aggregation tasks, the transfer of such approaches to formalized administrative procedures remains insufficiently explored, especially in legally constrained environments such as planning approval processes.

The aim of this study was to design, prototype, and subsequently validate an AI-supported objection management system for public participation in infrastructure projects.

The theoretical foundations are explained at the beginning. This includes the classification of key technologies such as AI, LLMs, vector databases, the embedding process and prompt engineering, which form the basis of the prototypical implementation.

Based on current challenges in handling objections, specific use cases for an AI-supported objection management system are then derived. The underlying user needs were identified through a literature review and further informed by the study of Matthei et al. (2024), which evaluated expert interviews with initiators of public participation processes.

This is followed by the presentation of the AI-supported objection management system, which focuses on thematic pre-sorting and the generation of suggested responses. Its functionality and effectiveness are subsequently validated, beginning with a description of the methodological approach, which is based on four focus group workshops involving 25 participants. The experts are representatives from German road and rail construction authorities who are confronted with the processing of objections in their work. The results of the validation are then presented.

This study concludes with a summary of the most important findings, a discussion and an outlook on future research activities.

2.1. Definition, Background, and Classification of Artificial Intelligence

Artificial intelligence (AI) is a subfield of computer science for which there is currently no universally accepted definition. Instead, the scientific literature offers a wide range of different definitions. An early conceptual approach came from Alan Turing, who formulated the so-called Turing test in 1950. According to this test, a system is considered intelligent if it is able to pass it. However, due to its limited practical applicability, this approach is not suitable as a comprehensive basis for the modern understanding of AI. In his article ‘What is Artificial Intelligence’, John McCarthy describes AI as the science and process of creating intelligent machines. Another widely held perspective defines AI using the concept of intelligent agents. These are systems that perceive their environment, process information, and on this basis, perform actions to achieve predefined goals.

In his work, Wang (2019) emphasizes that researchers are fundamentally free to define terms as long as the chosen definition is applied consistently and taken into account when interpreting scientific results. Gil de Zúñiga et al. (2024) also formulated a systematic and modern academic definition according to which AI encompasses the ability of non-human machines to perform tasks, act functionally, communicate, and interact logically.

The definitions of AI have changed continuously in the course of scientific development. Early approaches to artificial intelligence primarily focused on rule- and logic-based response systems. Over time, the focus shifted toward more complex systems, and eventually, toward the underlying concepts and learning mechanisms that enable such systems. Today’s understanding of AI is largely shaped by advances in machine learning. Modern AI systems are no longer designed solely as rule-based systems but instead learn from large datasets and dynamically adapt their behavior based on the data.

The advantages of AI are evident in numerous fields of application. In the medical context, both Aung et al. (2021) and Yeasmin (2019) emphasize the added value of AI. According to the researchers and their findings, the use of AI helps to reduce the workload of staff, enables the complete automation of specific areas of activity, and also supports improved task distribution within the organization. In the public sector, AI primarily enables more efficient processing and handling of administrative processes. However, generative chatbot systems offer the greatest benefits and have the most far-reaching impact on the general population. This category of generative chatbots includes Chat-GPT (OpenAI), Gemini (Google), and Copilot (Microsoft), among others. Their introduction between 2022 and 2024 triggered significant AI hype and contributed to the topic becoming more prominent in the wider society in recent years.

2.2. Background of Large Language Models and Retrieval-Augmented Generation

Large language models (LLMs) are a specific class of AI models that are trained on very large amounts of data to analyze language and generate natural, coherent responses to free-text inputs. Responses are generated by recognizing patterns and associations between terms and calculating the probabilities of possible response formulations. LLMs are used in numerous scientific disciplines for a variety of purposes. Since the introduction of generative pre-trained transformers, in particular ChatGPT, LLMs have gained in importance, especially in the field of generative and automated language processing. Minaee et al. (2024) considered the release of ChatGPT in 2022 to be an important milestone in the development of such language models. OpenAI’s GPT models are currently among the most powerful LLMs and achieve human-like results in individual tests. In addition to these, there are other significant models such as LLaMa from Meta and PaLm from Google.

Despite their high performance, LLMs have inherent limitations, particularly with regard to the reliability and factual accuracy of the content they generate. A well-known problem is so-called hallucination, in which LLMs generate information that appears plausible but is in fact incorrect or fictitious. This is often caused by outdated, incomplete or insufficient context-specific training data that LLMs rely on when generating responses.

A promising approach to improve the response quality of LLMs is retrieval-augmented generation (RAG). RAG refers to a process in which relevant information from external, trustworthy, and verifiable sources is specifically retrieved during text generation and integrated into the generation process. This combination of language model and information-based retrieval significantly reduces the likelihood of hallucinations, as responses are not based solely on the knowledge stored in the model, but are supplemented by current external knowledge. In this way, RAG enables the generation of well-founded, comprehensible, and contextually accurate responses and increases both the accuracy and timeliness and reliability of the content generated by LLMs. This is particularly relevant in domains where generated information must remain understandable, transparent, and traceable.

Beyond dialogue-oriented applications, LLMs are increasingly discussed in the context of administrative and information-intensive processes. Studies indicate that they can facilitate access to large and complex data collections and support the structured processing of textual information. At the same time, the use of LLMs in administrative contexts also highlights the importance of reliable information retrieval and comprehensible outputs, particularly in settings where decisions and responses require a clear factual basis.

2.3. Background on Vector Databases and the Embedding Process

Embeddings and vector databases form the central foundations of modern AI applications, especially in the context of LLMs. Embedding refers to the process of mapping complex, mostly unstructured data objects such as words, sentences, documents, or images into a high-dimensional vector space. The resulting numerical vectors have a fixed length and represent semantic and syntactic properties of the underlying data in mathematical form. Within this vector space, similarity in content is represented by geometric proximity, so that semantically related objects have smaller distances than those with different content.

Modern LLMs predominantly generate contextualized embeddings that take into account the respective linguistic context and enable identical terms to be represented differently depending on their meaning. This allows ambiguities to be resolved and complex semantic relationships to be modeled more precisely. The similarity between embeddings is typically determined using distance or similarity measures, in particular cosine similarity, which is widely used due to its robustness to different vector lengths.

Vector databases are specialized data management systems that have been developed for the efficient storage, management, and retrieval of such high-dimensional vector representations. In contrast to classic relational databases, which are based on exact matches, vector databases enable searches for semantic similarity in continuous vector space. Their basic workflow comprises the phases of indexing and querying. While indexing involves first converting raw data into embeddings and structuring it for efficient searching, the querying phase also vectorizes a query and compares it with the stored vectors.

With the increasing prevalence of LLMs in particular, vector databases have become significantly more important, as they enable the scalable processing of large, dynamic, and high-dimensional datasets that are not adequately supported by traditional database systems. To ensure efficient search times, approximate nearest neighbor algorithms are often used, which deliver fast and sufficiently accurate results even with very large datasets.

In the context of infrastructure planning procedures, the use of embeddings and vector databases appears particularly relevant due to the large volume and semantic heterogeneity of submitted objections. From the authors’ perspective, traditional keyword-based retrieval approaches may be limited in such settings, as semantically similar objections can differ substantially in wording and structure. Embedding-based retrieval approaches may therefore offer advantages for identifying contextually related content and for supporting the processing of complex textual submissions in administrative environments.

Due to the fact that AI models are based on large amounts of data, retrieving the most appropriate information often proves challenging. Prompt engineering refers to the systematic process of formulating effective instructions for AI models and the underlying LLMs. The aim of such prompts is to ensure the desired response quality without making changes to the model parameters. Since content generation is not deterministic and the same inputs can lead to different outputs, prompting the desired output with appropriate prompts can be a mixture of art and science, according to OpenAI. Nevertheless, there are various strategies for achieving consistently high-quality results. Prompt engineering has therefore become a central process for exploiting the potential of AI models, increasing the accuracy of response generation and improving the performance of LLMs.

A prompt is a predominantly text-based input whose task is to give the model a specific instruction. In their review, examined a total of 41 prompt engineering techniques and classified them into different categories. The majority of the methods fell under ‘Reasoning and Logic’, in which prompts are designed in such a way that LLMs are encouraged to engage in step-by-step, rigorous thought processes and can develop consistent chains of argumentation independently.

3. Use Cases for an AI-Supported Objection Management System

The submission of objections by citizens is an essential part of public participation in planning and approval procedures. Equally important is the proper and efficient processing of these objections and the preparation of appropriate responses. However, the process of objection management is associated with several structural challenges that promote and cause time-consuming and inefficient processing. In particular, the high time expenditure represents a central challenge in objection management. From a case management perspective, the shortcomings identified in the literature can be systematically divided into four overarching user requirements: (1) the thematic heterogeneity within objections, (2) their often considerable scope, (3) the identification of objections with similar content, and (4) the formulation of appropriate responses.

In their study on the use of AI in infrastructure projects, Kümmel and Lucka (2020) identified several areas of responsibility that can be very time-consuming when processing objections. These include, in particular, the division of individual objections into thematically coherent sections to enable subsequent forwarding and assignment to the relevant specialist staff. Against this background, AI-supported thematic pre-sorting of objections has considerable potential to reduce the processing effort. Both this potential and the practical relevance of this approach are highlighted in the study by Harding et al. (2022), among others. The German Federal Ministry of Transport and Digital Infrastructure also emphasizes that where possible, objections should be grouped by topic for reasons of efficiency and processed in appropriate thematic sections.

In the context of extensive planning and approval procedures involving public participation, it is also common for citizens to submit very detailed objections that address several topics at once. This poses considerable challenges for case workers, as the central content and concerns of an objection are not immediately apparent and must first be identified independently. Against the backdrop of the challenge above-mentioned, the use of AI-based text summarization, which presents the essential content of an objection in condensed form, appears to be a sensible approach.

In addition, identified the recognition of so-called serial letters as another key user requirement. In the context of infrastructure projects, citizens often join together to form interest groups or citizens’ initiatives in order to jointly draw attention to perceived impairments or to articulate their protest. As a result, numerous objections are submitted that are largely identical in content and differ only in formal aspects such as names or signatures. Identifying and responding individually to such serial objections is time-consuming for case workers and has been proven to be inefficient. To overcome the challenge above-mentioned, the use of automated text similarity analyses therefore appears to be useful, as they can be used to identify, group, and bundle objections with similar contents at an early stage.

Finally, the sheer number of objections, especially in the case of spatially significant construction projects, poses a considerable challenge, as each individual objection must be formally answered and responded to. Drafting appropriate responses is extremely time-consuming and inefficient for case workers. Against the backdrop of the aforementioned challenge, the use of AI-based support in formulating responses appears to be a sensible approach.

The first use case addresses the thematic pre-sorting of objections. The aim is to automatically divide objection texts, which often deal with several topics at the same time, into thematically coherent sections. This structuring facilitates both the assignment of individual text passages to the responsible specialist and the further processing of the objections in terms of content. In addition, pre-sorting enables the targeted identification and filtering out of text components with low technical relevance, such as greetings or farewell phrases, which are not of central importance for the preparation of a response.

The second use case concerns the automated summarization of objections. Case workers are regularly confronted with extensive objection texts that contain numerous irrelevant details in addition to relevant arguments. In this context, AI-supported text summarization serves as a supporting tool to give case workers a quick overview of the key content of an objection. The summary makes it easier to identify relevant aspects and helps to focus the response on the essential points. The aim is to reduce objection texts to their core statements and simplify subsequent processing.

The third use case relates to text similarity checks of individually submitted objections. Infrastructure projects regularly give rise to form letters or form objections that differ only slightly in content and usually only vary in terms of names or signatures. Processing these objections manually is a considerable burden and extra work for administrators. AI-supported text similarity analysis makes it possible to identify such objections efficiently. This allows for a standardized response to be created and applied to multiple objections with minimal adjustments, which increases efficiency in the processing workflow.

The fourth use case involves generating and suggesting text proposals for responses. Writing substantive responses is a particularly time-consuming part of objection management for case workers and needs to be optimized. AI can be used to automatically generate suggested responses based on the objections recorded. These serve as support and guidance for case workers. The generated texts can be accepted in full, customized individually, or rejected entirely, so that technical control and responsibility ultimately remain with the responsible case workers. The aim of this use case is to speed up the process of creating responses while supporting the quality and consistency of the answers.

This section begins with a presentation of the AI-supported distribution and categorization of objections, detailing the underlying system architecture and the interactions among its components. This is followed by a comprehensive description of the prototypical implementation, including the AI models employed and relevant technical specifications. Subsequently, the generation of suggested responses is addressed. The section first outlines the architecture of the response generation system, illustrating the integration of its components and the overall workflow, before providing a detailed account of the prototypical implementation. Finally, the design and functionality of the user interface are discussed, highlighting how users access and interact with the system and the generated outputs.

4.1. AI-Supported Distribution and Categorization of Objections

The architecture of the developed approach for thematic pre-sorting is based on the use of a LLM. In the first step, the LLM is tasked with dividing objections into thematic sections.

The categories used are based on a study by Matthei et al. (2025), in which 13 central topics of objections in the context of public participation were identified. The categorization itself is derived from an empirical analysis of approximately 1000 objections submitted by private individuals within German road infrastructure planning procedures. Since these objections originated from comparable administrative and procedural contexts, the identified categories represent recurring thematic patterns observed in practice and were considered suitable for the prototypical implementation developed in this study. The presented approach therefore focuses deliberately on a specific and practically relevant application scenario rather than on a universally applicable classification framework for all forms of public participation.

Additionally, two extra categories, “Greeting” and “Miscellaneous”, were included. The “Greeting” category enables the LLM to recognize common phrases such as “Dear Sir or Madam” or “Kind regards” as separate sections, while the “Miscellaneous” category captures objection topics that cannot be otherwise classified.

For the thematic division, the LLM receives the objection text along with the target categories and descriptions via a prompt template. The prompt instructs the model to divide the text into sections and output them with the corresponding categories while reproducing the original text exactly, without adding, omitting, or modifying content. This minimizes hallucinations, as LLMs can otherwise produce incomplete or altered outputs.

The architecture also limits the number of categories per section to maintain focus. Too many categories can blur classification, complicating assignments to responsible personnel, whereas too strict a limit may result in sections containing multiple relevant topics being inaccurately divided. Despite this trade-off, limiting categories per section is considered effective.

To further prevent hallucinations, a second step compares the LLM output word by word with the original text. If a complete match is confirmed, the sections are output with assigned categories. If discrepancies occur, the text is instead split according to the original line breaks and output without categories.

The technical implementation of thematic pre-sorting was based on the GPT-4.1-mini large language model (LLM) from OpenAI. At the time of the prototype development, GPT-4.1-mini was a powerful and efficient LLM, which had demonstrated convincing performance on text-based tasks in both internal tests and other studies. The prompt was designed to produce output in JSON format and to assign a maximum of two topics per section. This limitation ensures the content focus deemed necessary in the system architecture.

The implementation was evaluated by experts from both project developer and planning approval authority groups through focus group workshops. In this process, the thematic classification of objections was assessed to validate the practical applicability and accuracy of the AI-assisted system.

The architecture of the AI-assisted response suggestion system differs from the thematic pre-sorting approach in that it does not use a LLM to generate new responses. This design choice reflects the legal sensitivity of formal public participation procedures. Because responses are legally binding, AI-generated replies could introduce subtle inconsistencies across otherwise identical objections, potentially leading to unequal treatment in a legal context.

Accordingly, the system focuses on the structured retrieval of relevant, previously prepared responses without automated rephrasing. The integration of an LLM for response generation could be considered in future developments, for instance, using legal assistants or AI-based legal plugins. However, such solutions were not significant at the time of the system’s conception.

Conceptually, the system first stores general, cross-project objections and responses in a semantically searchable vector database, following a RAG process. Incoming objection sections are compared to identical entries within the project, followed by a cross-project semantic search for similar sections. If a match is found, the corresponding response section is displayed as a suggested reply.

Users can then adapt and finalize the response, which is stored along with the original objection in the vector database for future retrieval. To maintain clarity, the number of suggestions displayed is limited, and a selection interface ensures human oversight. This approach follows a human-in-the-loop principle, keeping the user actively involved in the decision-making process.

For the technical implementation of the response suggestion feature, the Chroma vector database was used. As described in the system architecture, it stores both the standardized collection of objection and response sections across projects and the project-specific collection of newly added objection and response sections.

The document containing standard questions and answers was based on a dataset of nearly 1600 documents provided by the BIM4People research project. To generate the standard questions and answers, objections and responses were reviewed to identify project-independent questions along with their corresponding answers.

It should be noted that if more than ten relevant response sections are identified (determined based on a threshold for the cosine similarity), the output of the response sections in the semantic search is limited to a maximum of ten results. This limitation ensures clarity and manageability during processing, in line with the requirements described in the system architecture.

The interface presents two central functionalities to the administrators. On the one hand, the thematic pre-sorting of objections, and on the other hand, the AI-generated text suggestions for their response. To ensure traceability, the original text of the objection to which the automated categorization refers is always displayed alongside the assigned topic category. In the area of thematic assignment, users have access to a confirmation field with which they can validate and confirm the system-generated topic suggestion. The topic assignment is displayed in an input field that allows administrators to independently reject the system-generated AI topic assignment and assign an alternative topic area to the objection text.

Adjacent to the surface elements of the thematic pre-sorting are the functional areas for responding to the individual pre-sorted topic sections. In this area, administrators can either assign the respective topic section to the persons responsible for the subject matter or formulate the response themselves. The AI-generated text suggestion for the response is displayed below the formulation field. Users can either accept and save this suggestion directly, in which case it will be adopted as the final response, or reject it and create their own response.

Focus group workshops were conducted to validate the developed AI-supported objection management system. This method represents a qualitative research approach and can be classified as a specific form of group interview in which the targeted use of communication and interaction among participants forms the basis for data generation. A key characteristic of this approach is that insights are not primarily gained through a series of interviews with individual persons, but rather through an explicit focus on the dynamics within the group.

The added value of this method lies particularly in the interaction and discussion within the group, through which participants can stimulate and inspire each other. In comparison to individual interviews, this enables a more diverse and creative examination of the topics under consideration. Furthermore, focus group workshops provide the opportunity to identify new and previously overlooked aspects, which can serve as a basis for further insights and ideas.

Henseling (2006) divides the focus group workshop process into three main phases. The first phase involves defining the subject of the study as well as organizational and methodological aspects, such as selecting suitable participants and developing a discussion guide.

Against this background, a total of 25 individuals were invited to participate in the focus group workshops. Twelve participants represented project developers who act as initiators of public participation and are therefore responsible for processing and responding to objections. Eleven participants represented planning approval authorities, which are responsible for issuing public announcements within planning and approval procedures as well as for the final assessment of objections and the corresponding responses. Two other participants could not be clearly assigned to either of these two groups and were employees of the Federal Information Technology Center (ITZBund). This was due to possible date changes, which expanded the original group of participants.

The core component of the focus group workshops is the presentation of the AI-supported objection management system, which will be made available to participants for hands-on testing. During the workshops, participants have the opportunity to submit a fictitious objection. They are then presented with a preliminary thematic sorting of their objection along with suggested responses. This is followed by a survey to evaluate the use cases, comprising both a quantitative component in the form of a Likert scale and a qualitative component in the form of a group discussion.

In addition to evaluating the general usability of the system, the workshops were designed to assess perceived usefulness, transparency, practical applicability, and potential process improvements within administrative objection-handling workflows. The combination of quantitative and qualitative evaluation elements was intended to enable both a structured assessment of user perceptions and a deeper understanding of the reasoning behind the participants’ evaluations.

A Likert scale is a well-established method for measuring personal attitudes and experiences. Its purpose is to systematically capture and quantify the intensity of participants’ attitudes toward the topic under investigation. In the present study, a seven-point Likert scale was used to provide a neutral middle option while still allowing the participants to express varying degrees of agreement or disagreement. Participants also had the option to refrain from providing an answer if they did not wish to respond.

The online collaboration platform Miro was used for implementation and documentation. This enabled collaborative work and individual coordination within the focus group workshops.

The evaluation of the focus group workshops represents the third step according to Henseling (2006) and was carried out in this study by summarizing key aspects of the discussion based on Ruddat (2012). In this context, the key findings are summarized and substantiated by key quotes from the participants.

To enable an integrated interpretation of the findings, the qualitative discussion results were analyzed in relation to the quantitative survey outcomes. This combined evaluation approach allows for quantitative tendencies observed in the Likert-scale responses to be contextualized and further explained through qualitative participant feedback and discussion dynamics.

The quantitative survey in connection with the evaluation of the individual use cases for the response assistance system shows that, according to the workshop participants, Use Case 1: Thematic Pre-Sorting was the use case that triggered the strongest improvement. The mean value of 2.30 lies between the scale values ‘strong improvement’ and ‘very strong improvement’. The mean value for Use Case 3: Text Similarity Check was 2.00, whereas the mean value for Use Case 4: Response Text Generation was 1.80 and thus lies between the scale values ‘moderate improvement’ and ‘strong improvement’. In contrast, Use Case 2: Text Summary received the lowest mean value in comparison, which was 1.50.

With regard to Use Case 1: Thematic Pre-Sorting’, one workshop participant noted a significant reduction in workload. In this context, he explained: ‘As someone who cut up 800 or 900 objections in a team last year, I can say that this application is great. That’s exactly the kind of work we used to sit here with everyone and do, like Sisyphus, because it had to be done’. A participant from another workshop mentioned having used an objection management system in the past. However, this was an ‘objection management system without AI’. This was ‘very difficult, because you had to compile a lot of information yourself and keep Excel lists ready in parallel to maintain an overview’.

One workshop participant noted that strict pre-sorting based on specified topics could pose a challenge. In this context, he noted that further differentiation within the topic areas might be necessary for some procedures. The topic of species protection was cited as an example, which requires further subdivision if necessary.

Regarding Use Case 2: Text Summarization, one participant concluded that they were ‘a little torn’. The participant added that ‘for simple sample applications sent by the whole village, it works great in individual cases. But when it comes to public authorities or other people who are professionally involved, a summary won’t get you very far’.

Another workshop participant summarized the use case as ‘good’. However, one must be careful ‘that no important questions of meaning are lost’. Another participant in a different workshop described this in similar terms, concluding that there is a risk ‘that important points will be overlooked or not taken into account’.

In connection with Use Case 3: Text Similarity Check, one workshop participant noted that this was fundamental ‘in order to respond consistently and not to react to different identical objections with different text modules’. Another workshop participant emphasized that he liked this very much, although one should ‘always check it over again yourself’.

With regard to Use Case 4: Response Text Generation, one workshop participant noted that the text suggestion should always be critically reviewed. AI training for administrative staff is necessary for this. However, this is already required by law. A participant from the same workshop agreed with this opinion and stated that ‘AI supports, but humanity must remain’. Another participant concluded: ‘I see a lot of potential in the responses, as the system is learning’.

One participant summed it up: ‘I think it’s a great approach and it could make things a lot easier’. Another participant stated: ‘All the options available are listed. It all makes sense’. This was described in similar terms by another workshop participant, who saw this approach as easily transferable to other industries. The handling of doctor’s letters was cited as an example.

One participant also noted that objections should never be answered automatically. He explained: ‘People need to deal with them and, where necessary, respond with empathy’. Another participant asked whether ‘AI can take information from other processes’.

The aim of this study was to conceptually develop, prototypically implement, and subsequently validate an AI-supported objection management system for public participation in planning and approval procedures for large infrastructure projects. The system was designed to support administrators in the efficient and user-friendly processing and response to objections within formal approval processes.

First, the theoretical foundations were presented, including AI, LLMs, and prompt engineering. The review of these concepts indicates that AI applications have considerable potential to improve administrative processes in the public sector. At the same time, their practical implementation in administrative procedures remains limited.

Based on the identified user needs derived from the literature review, four specific use cases for an AI-supported objection management system were developed. These included the thematic pre-sorting of objections, automated text summarization, the checking of text similarities, and the generation of suggested responses.

Subsequently, based on user needs that were identified through a literature review, specific use cases for an objection management system were developed. These were the thematic pre-sorting of objections, the automated creation of text summaries, the checking of text similarities, and the generation of text suggestions for responses.

The AI-supported objection management system was presented, including a detailed description of the system architecture followed by a prototypical implementation. The focus was placed on the use cases of thematic pre-sorting and response generation. The approach to thematic pre-sorting uses an LLM to divide objections into sections and assign them to thematic categories. In contrast, the generation of suggested responses does not rely on the automated creation of new texts by an LLM. Instead, relevant existing responses are retrieved from a vector database and provided as suggestions. This design choice was made due to the legal sensitivity of formal participation procedures, where the consistent treatment of similar objections is essential. Nevertheless, the integration of LLM-based rewriting or response generation functionalities represents a potential area for future development, particularly in combination with human oversight mechanisms and legally validated response templates.

Afterward, the developed prototype and the defined use cases were validated through focus group workshops. Four workshops were conducted with a total of 25 participants. These workshops provided the opportunity to present the prototype of the objection management system to an expert audience. The participants consisted of representatives responsible for processing and responding to objections, as well as representatives of approval authorities who are responsible for the final assessment of objections and the corresponding responses within planning and approval procedures.

The validation results indicate that participants perceived moderate to high optimization potential for objection management across all four use cases. In particular, the thematic pre-sorting of objections was identified as offering considerable potential for improvement and as promising a substantial reduction in administrative workload. The analysis of text similarities was also assessed positively by the participants, as it enables consistent responses to identical or similar objections and thus contributes to a noticeable reduction in processing effort.

Despite these identified potentials, several challenges and critical aspects must be considered to ensure the effective implementation of such systems. The relevance of data protection issues became particularly evident during the focus group workshops, where this aspect was repeatedly highlighted as a key challenge. Participants emphasized that data protection constitutes a decisive factor for the legally compliant use of an AI-supported objection management system. In particular, the adequate protection of personal data such as the names and addresses of individuals submitting objections must be ensured. Strict compliance with the requirements of the General Data Protection Regulation (GDPR) was identified by participants as a fundamental prerequisite for the practical use of the system.

In this context, the technical implementation of the developed prototype also raises important data protection considerations. The prototype implementation of the objection management system was based on services provided by OpenAI, a company based in the United States. In contrast to the comprehensive data protection legislation in Germany and the European Union, data protection in the United States is primarily regulated by industry-specific regulations and voluntary commitments by companies. These different regulatory approaches lead to considerable legal uncertainty regarding the use of the system and make it difficult to ensure uniform data protection. For data protection-compliant operation, the use of an independently hosted LLM appears to be advisable. In this way, dependence on commercial providers can be reduced and data sovereignty within the company’s own infrastructure can be ensured.

Beyond data protection aspects, additional legal and regulatory developments may also influence the future use of the objection management system. The European Union predicts that AI will become increasingly important in all social and economic sectors. In response, the EU AI Act was passed, which aims to establish a uniform legal framework for the use of AI applications. The AI Act provides for a risk-based classification of AI systems and defines corresponding requirements depending on the respective risk classification. The classification distinguishes between prohibited systems, high-risk systems with strict regulatory requirements, and systems with minimal requirements. Since the AI Act was only passed in 2024, its transposition into national law by EU member states is still in its early stages. Due to this delay, it is not yet possible to derive any specific requirements or binding specifications for systems such as the objection management system presented. A final assessment of whether and which regulatory, technical, or other requirements will be enacted for such systems in the future is not possible at this time and is purely speculative.

In addition to legal and regulatory considerations, behavioral aspects in the interaction with AI systems must also be taken into account. Since AI applications and systems became widely available to the general public, a critical pattern of behavior has been observed in many cases. A large number of users consider system-generated content to be correct and factually accurate without critically questioning or independently validating it. This occurs even when applications contain explicit warnings about potential inaccuracies in content or facts. In many cases, there is a lack of critical reflection on the generated content, which can lead to excessive and potentially problematic trust in AI applications. This tendency poses a key challenge for the use of the objection management system, as case workers could also be affected by such problematic dependence.

Beyond behavioral risks, the implementation of AI-supported objection management systems in formal administrative procedures also raises important questions regarding responsibility and liability attribution. Incorrect thematic classifications, misleading summaries, or inappropriate response suggestions generated by the AI system could potentially contribute to the unequal treatment of objections, administrative reconsiderations, or even legal disputes. In such situations, questions arise concerning the distribution of responsibility between the AI system provider and the responsible administrative personnel.

Several participants in the focus group workshops emphasized that objections should not be answered fully automatically by the AI application. Instead, they highlighted the necessity for case workers to retain decision-making authority over the assessment of objections and the wording of responses. As a result, the findings from the focus group workshops suggest that AI-supported objection management systems should be designed as decision-support tools rather than fully automated decision-making systems. In line with this understanding, final responsibility for all processing steps remains with the responsible administrative personnel, who must critically review and independently validate all AI-supported outputs before adoption. This applies equally to thematic classifications, text summaries, text similarity analyses, and generated response suggestions.

To support this reflective examination of AI-generated content, several control mechanisms were integrated into the prototype objection management system. For example, automatically generated thematic classifications must be explicitly confirmed by case workers before further processing is possible. If classifications are incorrect, users can independently adjust the assigned categories. A comparable mechanism was implemented for response suggestions, enabling users to reject or modify generated outputs in cases of technical or content-related inaccuracies.

In addition to these technical safeguards, participants in the focus group workshops also emphasized the importance of training measures for administrative personnel. Such training could help raise awareness of the risks associated with an uncritical reliance on AI-generated content and promote informed and responsible decision-making when handling objections within formal participation procedures.

In summary, the AI-supported objection management system developed in this study offers significant potential to increase efficiency in objection processing within public participation procedures. The four designed use cases specifically address key challenges and meet user needs that currently constrain objection management workflows and reduce processing efficiency. Empirical validation through focus group workshops demonstrated improvements across all areas compared to existing methods. These results indicate a substantial gain in efficiency through the system. At the same time, the findings should be interpreted in light of the study’s scope and participant composition. The validation was conducted exclusively with 25 participants, most of whom originated from German road and railway infrastructure administrations. However, this relatively homogeneous professional background was a deliberate methodological choice, as the study specifically focused on formal participation procedures within German transport infrastructure projects and therefore benefited from the practical expertise of stakeholders directly involved in objection processing and approval procedures. Nevertheless, the transferability of the findings to other infrastructure sectors, such as energy or water infrastructure, as well as to administrative and legal contexts in other countries, has not yet been empirically validated.

Nonetheless, further development and systematic testing are required to fully realize the identified potential. For future research, it appears sensible to evaluate cross-project findings systematically, which could inform the identification and prototypical integration of additional use cases, thereby expanding the system’s functionality and maximizing its practical benefits.