Projects such as F13, LLMoin and MUCGPT show that large language models have found their way into many German municipalities, currently mostly as chatbots with research functions (RAG). AI agents go a step further. They enable not only the generation of content but also automated sequences of actions within existing processes—from handling internal administrative tasks to controlling smart city applications. Under the umbrella term “Agentic AI,” new questions are coming into focus: Which tasks can AI agents meaningfully take on in public administration, where are their limits, and what is needed for their safe and productive deployment?
What are AI agents?
AI agents, often referred to as Agentic AI , are autonomous systemsthat independently pursue goals, break tasks down into sub-steps, and interact flexibly with their environment. Unlike traditional software, AI agents dynamically adapt their strategies.
The technical foundation is usually large language models (LLMs). These decide via tool calling , which tools they use and when: The language model creates the plan, a client executes it, and returns the results [1].
Unlike chatbots, AI agents do not merely answer queries but execute multi-step actions within existing processes. In public administration, this includes, for example, creating and prioritizing tickets in specialized procedures or dynamically optimizing traffic flows in a smart city. Depending on the deployment scenario, they operate with varying degrees of autonomy and are generally still embedded in human approval and control processes today.
Der Tagesspiegel Background describes them as the “latest[n] beacons of hope for the digitization of German public administration.” One example of this is the so-called Germany App: The BMDS apparently isn’t planning a single application, but rather a new AI-supported administrative platform—with agent-based AI systems as a central component [2].
At Fraunhofer IESE, we critically examine AI agents in public administration and smart cities, analyze their potential, test their deployment, and evaluate in which areas they deliver real added value. Public administrations are characterized by recurring, data-intensive tasks. This involves converting text-based information, applications, reports, and requirement documents into structured formats, entering them into specialized systems, or preparing them for further processing. This is often followed by an approval process, which in turn triggers further steps. In these multi-stage, rule-based workflows, agent-based systems can be efficiently deployed for process automation.
Application Scenarios: AI Agents in Public Administration and Smart Cities in Three Levels
Early real-world examples demonstrate that AI agents are already delivering benefits in public administration and urban planning, even though the systems are still far from operating completely autonomously. To structure the range of possible application scenarios, we have divided the examples into three levels based on the complexity of the respective task.
Level 1: Automation of structured administrative processes
A concrete example comes from the Federal Employment Agency (BA), which is currently testing AI agents based on large language models within its administrative organization. The goal is to automate the manual transfer of change requests and user requirements into Jira tickets—a time-consuming and error-prone routine task. To achieve this, several specialized agents work together in a division of labor. A Reader agent extracts relevant content, a Planning agent breaks the task down into steps, a Creator agent creates the ticket including title, description, priority, and category, and a Review agent finally checks the results for consistency and filters out duplicates. Initial results are positive; automated ticket creation works smoothly in most cases, and manual effort is largely eliminated. For security reasons, the BA relies exclusively on locally operated models from Aleph Alpha, Meta, Mistral, and Alibaba. The pilot environment has been fully integrated into the organization’s own on-premises infrastructure [1].
Stage 2: Dynamic Optimization of Urban Processes
An example of the next level of complexity can be found in Lagos, Nigeria, where AI systems autonomously design public transit schedules based on passenger movements and real-time traffic data. The system does not respond to predefined triggers but continuously and dynamically optimizes operations. This example demonstrates how agent-based systems can make urban mobility more efficient and adaptable [3]. In Germany, comparable approaches are still under development, for example in the field of AI-supported traffic control.
Level 3: Cross-system control of complex urban infrastructures in China
The City Brain project, developed by Alibaba Cloud, exemplifies AI-supported urban management that goes beyond traditional automation. Originally deployed for traffic optimization, the system now coordinates emergency response, resource allocation in healthcare, and aspects of urban planning. In emergencies, City Brain autonomously reconfigures traffic control to clear the way for emergency vehicles, often without real-time human supervision. The system does not prioritize based on fixed rules, but rather on dynamic assessments of multiple urban variables simultaneously [3].
Not every AI agent that takes on administrative tasks needs to be as complex as the examples from Nigeria and China. For rule-based Level 1 processes, such as automated ticket generation, a language model is often sufficient—one that is guided along clearly defined paths by a well-designed system architecture and targeted system prompts. The current process status determines which options are available to the agent at any given time. At the same time, a working prototype is not the same as a production-ready system. The real challenge lies in designing the path to get there—one that is reliable, maintainable, and scalable. We have the expertise to follow this path: start small, measure, improve, repeat.
At Fraunhofer IESE, we are exploring the potential of this technology in concrete projects. One example is an AI-supported platform for digital citizen participation. The platform collaboratively supports administrative staff in creating participation formats. The underlying language model helps structure, formulate, and adapt content. The finished participation content is then automatically entered into the open-source platform Consul via API. The system thus handles recurring, rule-based steps, but still operates clearly within the framework of human control and approval. It is less of an autonomous agent and more of an AI assistant that relieves administrative staff of some of their workload. However, with an expanded system architecture—such as the ability to independently monitor participation processes, evaluate feedback, and initiate follow-up processes—a transition to an AI agent would be conceivable.
Challenges and Security: What Matters When Deploying AI Agents
Relevant Risks of AI Agents in Public Administration
Despite the potential outlined above, the productive deployment of AI agents in public administration also necessitates new requirements for the security and quality assurance of these systems.
Unlike traditional software, agents based on language models can exhibit various new error patterns , such as:
- hallucinations
- undesirable content
- unforeseen reactions to adversarial inputs (deliberately manipulated inputs designed to induce errors in the system).
Which risks are relevant here depends heavily on the specific use case.
Decisive factors include:
- the system’s degree of autonomy
- what data it processes
- whether it has write access to external systems
- whether actions can be undone.
Example: If the BA’s Creator Agent creates a Jira ticket based on a fabricated requirement, this can lead to incorrect development priorities. An error that persists through several process steps before it is noticed. It is precisely these types of error classes that must be defined in advance and covered in testing.
Risk-based testing approach for AI agents
A risk-based approach is recommended for testing.
- Define critical error classes
First, the critical error classes for the respective use case should be defined, such as hallucinations, discriminatory language, data leaks, or susceptibility to manipulative inputs. - Build targeted test sets
For each of these areas, targeted test sets can be built and evaluated using a combination of human assessment and automated methods such as LLM-as-Judge. - Set strict limits for error tolerance
In doing so, it makes sense to set strict limits. A zero-tolerance policy applies to clearly illegal content or the leakage of personal data, while stylistic weaknesses can tolerate a higher error rate. - Detecting regressions early
It is also important to run a manageable re-test set after every model or prompt update to detect regressions early. [4; 5].
Outlook: AI Agents in Public Administration on the Path to Practical Application
AI agents in public administration are still in their infancy, but the direction is clear. The technology is evolving rapidly, initial projects demonstrate added value, and the pressure for modernizing public administration is growing. At Fraunhofer IESE, we are working to translate theoretical concepts into operational applications by developing suitable architectures and testing methods, as well as through close collaboration with municipalities and government agencies.
References:
[1] eGovernment Computing. »AI Agents in Public Administration«, 2025.
[2] Tagesspiegel Background. »What the BMDS Plans with Agent Systems«, 2025.
[3] Tiwari, A. »Conceptualising the emergence of Agentic Urban AI: from automation to agency“. In: Urban Informatics 4 (13), 2025
[4] GovTech Singapore. »Agentic Risk & Capability Framework“, 2025.
[5] Google. »SAIF: Secure AI Framework for Agents«, 2024.
