AI for municipalities and government: an assistant, not a decision-maker

Q: Can we use cloud LLMs such as OpenAI?

This can be considered for data without personal data or business-confidential information. For processes involving BSN numbers, BRP data or case records, we recommend open-source models that run within EU regions or on-premises. Schrems II and the requirement for data residency weigh heavily.

Q: How do we prevent the model from reproducing bias?

By building bias testing into model validation, not as an after-the-fact check. We measure disparate impact across relevant demographic dimensions, check whether the model uses proxy variables for protected characteristics, and set a threshold below which the model may not go into production. When in doubt, the system falls back to rule-based logic with human review.

Municipalities are under pressure: fewer staff, more laws and regulations, and a citizen who expects an immediate answer, with a supervisor checking every automated decision against the Awb and the AI Act. We build AI applications that support civil servants: chatbots for citizen questions, NLP for Woo requests, automation of WMO intake — without the algorithm making the final decision. Human-in-the-loop is not an add-on; it is the starting point.

Citizen chatbots Woo/Wob request processing WMO automation Permit AI Council report summarisation

Discuss your case View applications

Why AI for government is a different conversation

A municipality is not a webshop. Decisions affect citizens' lives: benefits, care assessments, permits, enforcement. That changes how AI may be deployed, not whether it can be deployed.

The childcare benefits scandal has permanently damaged Dutch trust in automated government decision-making. For digitalisation managers and CIOs at municipalities, that means every AI application is tested for explainability, for bias, for compliance with the General Administrative Law Act (Awb art. 4:8, the duty to state reasons), for the AI Act high-risk classification, for GDPR, for the Information Security Plan (IBP) and for the Baseline Information Security for Government (BIO). That may sound daunting, but it is good news: organisations that deploy AI responsibly achieve measurable gains without the risks of a Syri or childcare benefits scenario.

In practice, many municipal processes are not about decisions but about providing information, completing forms, summarising case files and looking up documents. A chatbot that explains at 11 p.m. how to renew a passport makes no decision. An NLP pipeline that searches a Woo (Open Government) request and proposes draft redactions makes no decision; an official still presses publish. A council report summariser that condenses 200 pages into ten bullet points for councillors makes no decision. That is exactly where the first, safe layer of AI impact lies: support, not replacement.

We build that layer. No experimental algorithms that make opaque decisions about who does or does not receive a WMO (social support act) assessment, but concrete tooling that halves the processing time of a Woo request, relieves the contact centre and lets councillors come to the meeting better prepared. Our municipal software practice and our government apps portfolio form the foundation; AI is a component, not a standalone product.

Where AI works in the municipal organisation

Six areas of application where impact, feasibility and legal room are in order. Each scenario is built around support, not replacement.

💬

Citizen chatbot for 24/7 questions

An AI chatbot on the municipal website or via WhatsApp answers questions about passports, driving licences, moving house, waste, parking and events. The model draws on your own knowledge base and directs people to the right desk or form. Multilingual support (Dutch, English, Turkish, Arabic, Polish) makes information accessible to the whole population. When in doubt, it escalates directly to an official.

📑

Speeding up Woo/Wob requests

The Open Government Act has multiplied the number of requests. NLP models search thousands of emails, memos and attachments for relevance to the request, propose redactions based on the Woo exemption grounds (article 5.1) and keep track of which documents have already been reviewed. A civil servant either accepts each proposed redaction or rejects it: the tool accelerates, it does not decide.

📋

WMO/Youth Act intake automation

The first contact interview for a WMO application generates a file recording provisions, living situation and the request for support. AI speech-to-text plus structuring completes a standardised intake form in line with your ordinance. The case worker corrects and validates it. The decision on the assessment remains with a human: Awb article 4:8 requires a reasoned decision that the applicant can test.

🏛️

Permit AI for local bylaws and the Environment and Planning Act

The Environment and Planning Act requires an integrated assessment. AI models check applications against the environmental plan, flag conflicts and suggest conditions drawn from earlier comparable decisions. For simple applications (dormer window, driveway), this speeds up processing; for complex applications, the permit officer receives a substantiated starting document instead of a blank template.

📰

Council-paper summarisation and searchability

Councillors receive hundreds of pages of documents every week. An LLM pipeline generates a summary of fixed length for each document, recognises motions and amendments, and makes the entire archive searchable through natural-language queries. For council clerks, this saves manual reading; for councillors, it improves the quality of deliberation.

🔎

Anomaly detection without bias risk

Where Syri foundered, anomaly detection can work: not by scoring citizens for suspicious behaviour, but by monitoring internal processes. Deviating invoice patterns, irregularities in subsidy applications from the same organisation, duplicate claims with suppliers. The model looks at transactions, not citizens, which is a fundamentally different risk profile under the AI Act.

AI chatbot implementation AI document processing Permit software Software for municipalities

How Appfront builds AI for public organisations

We don't start with a model, but with a process analysis. Which task takes civil servants the most time? Where do citizen complaints about unreachability arise? Which workflow is under strain from the Open Government Act (Wet open overheid) or the Environment and Planning Act (Omgevingswet)? Only once that is clear do we assess whether AI makes sense — and if so, whether it becomes a generic LLM application, a specifically trained model, or no AI at all but smart automation.

Our projects fit within your existing infrastructure. For authentication we integrate with DigiD and eHerkenning, and for identity data with the Municipal Personal Records Database (BRP, formerly GBA) and the Register of Non-Residents (RNI). Suwinet integrations are only used where the Suwi chain permits it and with explicit authorisation. We align with the VNG's Common Ground principles: data at the source, reuse through standard APIs, no shadow administrations.

Pseudonymisation of citizen service numbers (BSN) and other personal data takes place before data reaches the AI model. Training data is anonymised or synthetically generated. Models run by default within Dutch or EU cloud regions, or on-premise within the municipal data centre. We follow the BIO and the IBP, not as a tick-box exercise but as an architectural starting point.

AI development Building an AI app Getting a municipal website built Municipal app

Four phases: from problem definition to production AI

An AI project for a municipality requires more preparation than a commercial project. Not because of the technology, but because of governance. Our phasing reflects that.

Problem exploration and DPIA

We discuss the use case with the process owner, the DPO (functionaris gegevensbescherming), the CISO and legal affairs. A DPIA (Data Protection Impact Assessment) determines whether the application is high-risk under the AI Act and which additional safeguards are needed.

Proof of concept with dummy data

Initial validation takes place on pseudonymised or synthetic data. We demonstrate that the model works without production data being put at risk. The PoC result is concrete: a working demo that process owners can test.

Pilot with human-in-the-loop

A defined production pilot in which the model makes proposals and civil servants carry out the final action. We measure processing time, quality, user acceptance and false positives. The DPIA is updated on the basis of the pilot.

Production and governance

Rollout with monitoring, audit logging and periodic model retraining. We document in line with AI Act requirements: datasets, evaluations, limitations, escalation procedures. Municipalities can demonstrate what the model does and does not do.

Technologies and standards we use

Public sector organisations have different needs from a typical enterprise stack. Open source is preferred wherever possible; vendor lock-in on a US cloud LLM is unacceptable for many municipalities because of Schrems II implications and the requirement for EU data residency. We therefore work by default with models that can run within EU regions or that can be deployed fully on-premise.

For chatbots and document AI, we deploy Llama, Mistral and Phi models on hosted GPUs within the Netherlands or the EU, using retrieval-augmented generation (RAG) so the model answers from your own knowledge base. For classical ML applications (anomaly detection, request classification), we use scikit-learn and XGBoost: explainable, auditable, no black box. We build frontend components with React and the NL Design System component library, so the application fits the NLDS standard and the municipal visual identity.

Llama 3 Mistral Phi-3 LangChain Haystack RAG scikit-learn XGBoost PostgreSQL DigiD eHerkenning Common Ground NL Design System Docker Kubernetes

AI Act, GDPR and the General Administrative Law Act (Awb): what this means for you

Compliance is not an afterthought in a municipal AI project. It is the starting point. Four frameworks we factor into every architecture decision.

EU AI Act high-risk classification

Applications that affect access to public services, social security or essential provisions fall under Annex III and are classed as high-risk. This brings requirements around risk management, data quality, technical documentation, transparency, human oversight and accuracy. We design the system so that documentation and oversight are built in from the outset, not reconstructed afterwards.

General Administrative Law Act (Awb) art. 4:8 and the duty to give reasons

A government decision must be reasoned in a way that the citizen can scrutinise. "The algorithm said no" is not enough. Our AI applications produce written justification that can serve as the statement of reasons, or deliver a substantiated proposal that the civil servant confirms in their own words. The legal decision remains with a human.

GDPR, BIO and IBP

Personal data is pseudonymised before model input wherever possible. Logging is set up in line with BIO: no more than necessary, and no less than required. The municipality's IBP forms the baseline; we integrate within the existing security architecture, not a separate silo.

Bias and non-discrimination

Models trained on historical data can reproduce historical inequalities. We test for disparate impact by nationality, neighbourhood, age and gender. Models that use sensitive attributes as a proxy for decision-making are not deployed, full stop. When in doubt: a rule-based fallback with a human decision-maker.

AI document processing Interim AI tech lead AI literacy training

Example projects from municipal practice

Three scenarios showing what a municipal AI application looks like in practice, from problem, through DPIA, to production.

A Woo team delivering redactions 40% faster

A mid-sized municipality receives structurally more Woo requests than it did under the Wob. The legal team has limited capacity. An NLP pipeline scans the requested documents, identifies personal data, confidential business information and draft decision-making documents (Woo articles 5.1 and 5.2), and produces an initial redaction proposal. The lawyer reviews and adjusts it, with no more manual searching for every BSN or email address.

BSN-free anonymisation for research

A municipality wants to analyse patterns in WMO applications to underpin policy. BSNs and name and address details are converted through a pseudonymisation pipeline before the data reaches the analytics platform. The analysis team works with keys that cannot be traced back to individual citizens. The raw data stays within the source application, in line with Common Ground.

A multilingual citizen chatbot for the 14+ service desk

A municipality with a diverse population receives enquiries by phone, email, WhatsApp and website. A chatbot draws on the indexed content of the municipal website, product catalogue and FAQs, answers in five languages, and escalates complex cases to the right department. The backend logs every interaction for audit and quality monitoring; no conversation is used for training without explicit review.

A council information system that summarises agenda items

Ahead of the council meeting, members receive an agenda with dozens of documents. An AI layer generates a summary per agenda item with key points, previous decisions, financial implications and relevant motions. Councillors click through to the full document when needed. For the council clerk's office, this means fewer questions in advance and better quality of debate.

Frequently asked questions about AI in government

Can an AI system make decisions for a municipality?

Not independently. The General Administrative Law Act (Algemene wet bestuursrecht) requires a reasoned decision attributable to a specific administrative body. AI may produce proposals, summarise case files and flag exceptions, but the final decision is taken by an official or administrator who endorses the reasoning. Human-in-the-loop is not an additional safeguard; it is the legal starting point.

Does our application fall under the high-risk category of the AI Act?

That depends on the application. Annex III of the AI Act lists access to public services, social security and essential services among its high-risk domains. A chatbot providing product information is generally not high-risk; a model that makes proposals on granting benefits is. We carry out this assessment during the DPIA phase at the start of the project.

Can we use cloud LLMs such as OpenAI?

For data without personal data or commercially confidential information, this can be considered. For processes involving citizen service numbers (BSN), BRP data or case data, we recommend open-source models that run within EU regions or on-premise. Schrems II and the requirement for data residency carry significant weight; it is also difficult to explain to citizens why their data travels via a US cloud.

How do we prevent the model from reproducing bias?

By building bias testing into model validation, rather than as an after-the-fact check. We measure disparate impact across relevant demographic dimensions, check whether the model uses proxy variables for protected characteristics, and set a threshold below which the model may not go into production. Where in doubt, the system falls back on rule-based logic with human assessment: better a slower process than a repeat of Syri.

How does AI integrate with DigiD, eHerkenning and Common Ground?

DigiD and eHerkenning handle authentication for citizens and businesses; AI applications authenticate through these channels when the interaction is personal. For data exchange, we connect to Common Ground APIs: data stays at the source, and reuse happens through standards such as StUF and NLX. We don't build shadow databases; we integrate.

What if the model, contrary to the intent of the Awb, still makes an implicit decision?

That is exactly where interface design matters. We ensure that officials are never presented with AI output as a ready-made decision. The system shows a proposal with justification, explicitly asks for confirmation or amendment, and logs the human action. In this way, an official can always demonstrate why a decision was taken: not "the system did this", but "I took this decision, and this was the input I used".

How long does an AI project take at a municipality?

The technical build is often shorter than the governance processes. A proof of concept is typically ready within a few weeks. A DPIA, advice from the data protection officer (FG), assessment by the IBP coordinator and alignment with the information security policy take longer in calendar time. We plan these in parallel: governance starts at the beginning, not at the end.

Do you work with VNG or GovTech suppliers?

We are used to working in landscapes where VNG-Realisatie, GovTech suppliers and municipal ICT departments are already active. Our role is usually that of AI partner: we build the application, while the municipality or its suppliers provide the existing infrastructure. No rip-and-replace; we integrate with existing case management systems, DMS platforms and CRM platforms.

Want to deploy AI in your municipality or public sector organisation?

Discuss your case with a team that is at home in both AI technology and the Awb, the AI Act and BIO. Together we will map where AI can contribute safely and with measurable impact, and where it is better not deployed.

Schedule a conversation