AI in local government: lessons from a live municipal system

When AI comes up in local government, the question is usually framed as "should we use AI?" I think the better question is this: which piece of work gets done better, and more safely, once it is handed to AI? Projects that start without answering it tend to end up in one of two places: a flashy chatbot nobody uses, or a system that puts the institution at risk over personal data.

In this article I describe how we added AI to a management system that runs live in a municipality, and the principles I took away from it: which task to start with, how to build the integration technically, and what to watch out for.

Three layers of AI in local government

It helps to think about where a municipality can use AI in three layers. The further up you go, the more both the benefit and the risk grow.

1. Internal efficiency. Summarizing reports, simplifying long documents, searching internal files. The output is read and used by staff.
2. Support for staff. An assistant that answers questions about regulations and service guides, drafts of letters and replies. The output feeds into a decision, but a person is still in between.
3. Services for residents. Classifying applications, answering frequently asked questions, routing requests to the right department. Here, a mistake by the AI reaches the resident directly.

My advice is always to start with the first layer. The reason is simple: the organization learns both the technology and the real state of its own data at low risk.

How to choose the first use case

I assess whether a task is a good first candidate for AI with four questions:

• Is the data ready? Is the information the task needs already digital and well structured, or does it first have to be transferred from paper?
• Is the cost of an error low? What happens if the AI produces something wrong? An incomplete summary is not the same as giving a resident wrong information.
• Does a person see the output? If a member of staff looks at the result before it is used, the risk drops considerably.
• Does it need personal data? If not, questions about data protection law and transfers abroad disappear from the start.

A task that answers "yes, low, yes, no" to these four questions is a good place to begin.

Our first step: the executive summary

In the system we built, departments enter their weekly and monthly activity reports into the system together with their KPI values. The mayor's office cannot realistically read every department's report one by one. AI turns these reports into a one-page executive summary: in a few sentences you can see which reports came in on time, where approvals are piling up and what stands out.

This task passes all four questions above:

• The data is already in the system and structured.
• If the summary misses something, the damage is small; the original reports are still there.
• Senior management reads the summary and drills down into a report when needed.
• An activity summary does not need any personal data.

On the technical side, this feature runs on Google Gemini, and only users with senior management permissions can see the executive summary.

Next steps: knowledge base, daily briefing, resident requests

After the first step, these are the logical ways to continue. They are on the roadmap of our system as well; none of them is live yet.

• Internal knowledge base. An assistant that answers natural-language questions about regulations, by-laws and service guides. The method behind it is called RAG (retrieval-augmented generation): the model does not answer from its own memory, but from the relevant passages it finds in the organization's documents, and it shows its source. For the public sector this is the most important property: you can see which clause an answer is based on.
• Morning briefing. A short summary prepared from the day's data and delivered to the mayor's office every morning: pending approvals, delayed tasks, topics on the agenda.
• Resident requests. Collecting requests, suggestions and complaints and routing them to the right department. Because it belongs to the third layer, this is the step that should come last.

How to build the integration technically

"AI integration" is often thought of as getting an API key. A sustainable LLM integration in the public sector needs at least these six parts:

1. Data layer. The data the model sees must be pulled regularly from existing systems and cleaned. AI working on scattered data only produces the mess faster. That is why the data has to be brought together in one place first.
2. Model choice. Large cloud models are powerful and easy to access, but data may leave the country. Open models running on your own servers keep the data in-house, but they need hardware and maintenance. The right choice depends on the data the task requires; for a summary that needs no personal data, a cloud model can be reasonable.
3. Permission boundary. AI output must follow the same role and permission rules as every other screen in the system. An assistant should never build an answer from a document the user is not allowed to see.
4. Logging and audit. Record which request was made, when, by whom, and what was sent to the model.
5. Evaluation. Before going live, prepare a test set of real examples and have an expert check the outputs. Whenever the model or the prompt changes, run the same test again.
6. Cost tracking. With pay-per-use services, monitor usage and cost, and shorten unnecessarily long requests.

Data protection and transfers abroad

The municipality in this project is in Turkey, where personal data is governed by KVKK, the Turkish Personal Data Protection Law. In November 2025 the Turkish Personal Data Protection Authority published a guide on generative AI and the protection of personal data. It covers legal basis, transfers abroad, transparency and security measures, and it is worth reading before any public-sector AI project in Turkey starts. Wherever you are, your own data protection framework (in Europe, the GDPR) raises the same questions.

From an architecture point of view, the most effective safeguard is that personal data never reaches the model at all. To get there:

• Ask, separately for every feature, whether the task actually needs personal data.
• If it does, anonymize or pseudonymize the data wherever possible.
• For tasks that genuinely need personal data, consider a model that runs inside the organization.

This article is not legal advice; before starting a project, the assessment should be made together with the organization's legal department and its data controller.

Bringing staff along

The most underestimated topic in AI projects is how employees see them. If the question "is this going to take my job?" goes unanswered, even the best technical solution will not be used. That is why it matters to pick a first feature that makes the employee's work easier and leaves the decision with them. A feature like the executive summary replaces nobody; it makes sure everyone is looking at the same picture.

Frequently asked questions

Do we need to replace our existing systems to integrate AI?

No. AI is usually added as a layer on top of existing systems. What you really need is data that is well organized and accessible.

Do we need to train our own model?

For most public-sector uses, no. Off-the-shelf models, good prompts and a RAG approach grounded in the organization's own documents deliver enough. Training a model only becomes relevant for very specialized, high-volume tasks.

How long does a first project take?

If the data is ready, a first-layer feature such as report summarization can be piloted within a few weeks. The part that usually takes the most time is preparing the data.

Conclusion

AI in local government is not a technology choice; it is a question of sequence. Data first, then low-risk work under human control, and services for residents last. A municipality that moves in this order earns the trust of its own staff and protects itself on personal data.

I described the order in which to approach digital transformation in my article on digital transformation in municipalities. If you would like to talk about where AI could start in your own organization, take a look at the AI and automation service or get in touch.