AI in Public Administration: Generative AI Instead of Rule-Based Bots for Citizen Services

When a citizen inquiry fails today is well documented: A father asks the municipal chatbot whether registering a second car can be done online or if he needs to come by in person. The bot does not know the term "second car"; its decision tree ends at "passenger car". The citizen calls, the clerk is busy, and the inquiry ends up on hold. Rule-based bots work as long as the question matches the pattern known to their question-and-answer tree. Citizens rarely ask in patterns.

Why citizen bots mostly fail today

Most "administrative bots" are rule-based. A rule-based system translates user input into keywords, assigns them to a category, and plays the associated answer. The fundamental problem of all rule-based bots remains the same: Any question that does not match a stored pattern breaks the dialogue.

On a municipal website with several hundred services, the number of possible questions is practically unlimited. A rule-based bot that tries to maintain every variant individually is never finished. And maintenance never stops: new services, changed responsibilities, new forms. Every update means manual work on the question-and-answer tree.

Added to this is fragmentation across individual channels. A FAQ chatbot on one subpage, a contact form on another, a phone number in the legal notice. Anyone switching channels loses the context of their inquiry and has to start from scratch.

What generative AI with RAG specifically changes in public authorities

Generative AI does not answer a question from a list of predefined answers, but formulates the answer from the content it finds. The technology behind this is Retrieval-Augmented Generation, or RAG. A RAG system searches the authority's existing content—such as bylaws, FAQs, and service descriptions—and answers the question based on that information. It can therefore correctly resolve phrasing that has never appeared in a Q&A before, as long as the underlying information is available on the website.

The difference is tangible for citizens. The same question about a second car is answered by a generative system with a sentence about online registration and a link to the exact form pathway, instead of aborting the conversation. Answers remain transparent and traceable because they cite the sources they originate from.

Agentic RAG goes one step further. Instead of just delivering a retriever-based answer, the system independently plans multiple steps, such as: create an appointment, check which documents are missing, and guide the citizen to the next administrative step. For multi-step administrative processes, this is precisely the relevant capability.

The limitation is equally clear. Generative AI only answers what is in the data repository. If information is missing from the linked sources, it cannot invent it. Quality therefore depends at least as much on content maintenance as on the underlying technology. That is why closing content gaps is part of every project, not just building the bot.

What is already running in regular operations in municipalities

Four examples show where municipalities stand today.

FITKO's 115 chatbot has been in regular operation since 2024 and answers questions from a fixed administrative service catalog. It works for municipalities in the 115 network, but it is a network solution: If you are not a participant, you cannot use it. It is based on a compact language model, and answers remain closely tied to the service catalog.

Munich's "Muckl" is a municipal FAQ chatbot based on the city administration's content. It answers inquiries across a single channel, traditionally as a website widget. It is worth mentioning as a publicly visible example of municipal bodies taking the step toward AI, not as proof of a multi-channel approach.

Heidelberg's "Lumi" has been running since 2022 on the basis of Aleph Alpha and has conducted around 4,500 conversations to date, averaging about 39 per day. The volume is modest, and the project serves as an example of an early, isolated implementation.

And the Fraunhofer IESE/PD whitepaper "AI Agents in Municipalities" (October 2026) officially recognized the trend: Agents are described as a forward-looking topic for municipalities, no longer merely an experiment.

The common gap across these examples is structural. Most solutions are single-channel FAQ bots or tailored to a specific network. What is missing outside the 115 network is a generative, multi-channel AI layer that serves chat, search, form, and phone from a single shared knowledge base.

For the public sector, legal compliance is not an optional extra, but a prerequisite for procurement. Three regulations intersect here.

The GDPR requires personal data to be processed within the EU. An AI layer for German authorities should therefore run on European hosting and not use any data for training purposes. In practice, this specifically means: Azure data centers in the EU for the models, user data kept separate, and both without data outflow to third countries. Learn more in the glossary entry on GDPR-compliant AI.

The EU AI Act has been in force since February 2025 and classifies systems by risk. For website AI, the most relevant aspects are transparency obligations, logging, and the possibility of human oversight. What the regulation specifically requires and where deadlines apply can be found in the entry on the EU AI Act.

And since June 28, 2025, accessibility under the Accessibility Strengthening Act (BFSG) has been mandatory for public bodies. WCAG 2.1 AA serves as the technical foundation. A chatbot that excludes screen reader users or can only be operated with a mouse fails the audit process.

Getting all three requirements from a single provider, rather than piecing them together from three separate contracts, saves more than just procurement time. It reduces the risk of a solution failing due to a single compliance requirement.

Citizens in every language, around the clock

A public authority website is usually only available in German. Citizens with a different native language, refugees, or international visitors hit this language barrier. A generative AI layer answers inquiries in over 101 languages from the same knowledge base, without requiring each language to be maintained separately. The guidance on registering a second car thus also works in Turkish or Ukrainian—as text during the day and as a phone response outside office hours.

At the same time, this relieves citizen services. Routine questions that take up the same amount of staff time every day are answered independently by the layer. The team remains available for cases that require real human decision-making, rather than facing the same routine questions day in, day out.

The branchly perspective: The AI layer instead of point-solution chatbots

branchly positions itself here as a generative AI layer for the website, not as a standalone rule-based bot. Instead of a question-and-answer tree, branchly works with Agentic RAG live on the existing content of bylaws, FAQs, and services. Content does not need to be rewritten or re-indexed; an update is reflected immediately in the responses. The compliance trifecta of GDPR-compliant EU hosting, EU AI Act compliance, and WCAG/BFSG accessibility is built-in—standard rather than an afterthought.

*If you want to find out for your municipality where a generative AI layer makes a real difference compared to today's rule-based bots, talk to us at https://branchly.io/schedule-appointment.*

Frequently Asked Questions

Is AI in public administration compliant with data protection laws?

AI in public administration is data protection compliant if processing takes place within the EU, no data is used for training purposes, and a data processing agreement under Art. 28 GDPR is in place. A provider that relies exclusively on EU hosting and does not use data for model training generally fulfills these core requirements.

Do municipalities need their own AI strategy?

An AI strategy makes sense for municipalities as soon as AI is no longer just an isolated project, but affects multiple services. The strategy clarifies responsibilities, data sources, and the rollout sequence before individual bots are built.

How much does a citizen bot cost?

Costs depend on scope and the number of channels. A simple FAQ chatbot starts in the low three-digit range per month. A generative, multi-channel AI layer for multiple services starts in the range of a few thousand euros per month with established providers, depending on the number of active user sessions.

Rule-based bot or agent: What is the difference?

A rule-based bot matches user input to a stored response pattern and fails when there are discrepancies. An AI agent understands the question, searches for the answer within linked content, and can plan multiple steps. Agents are therefore able to answer questions that were never pre-formulated.

AI in Public Administration: Why Generative AI Instead of Rule-Based Bots Is Decisive for Citizen Services