AI for Municipalities: Making Policy Documents Searchable
Policy papers, council records, permit files and local regulations become searchable by meaning instead of just keywords with RAG, with citations back to the original document. Why the Woo and data sovereignty matter, and when a municipal AI application can qualify as high-risk under the AI Act.
Municipalities can make policy papers, council records, permit files and local regulations substantially easier to search by using a RAG system: an AI assistant that retrieves relevant municipal sources before generating an answer. Instead of relying on what a language model happens to know, the system grounds its response in the municipality's own documents and can point users back to the original file, document date and, where available, page number.
That distinction matters. A municipal knowledge assistant should not merely sound convincing; its answers need to be verifiable. If a policy officer asks, "What conditions apply to converting a property into multiple units?", the useful output is not just a summary. The user needs to see which regulation, policy rule or council decision supports it. If the document collection does not contain sufficient evidence, the system should say so rather than invent a plausible municipal policy.
Why municipal information is difficult to search
Municipalities rarely suffer from a lack of information. The more common problem is fragmentation.
Council information systems, case-management software, document management platforms, network drives, archives and public websites may all contain parts of the institutional record. The content itself ranges from PDFs and Word files to scans, appendices, decisions, policy papers and multiple historical versions of regulations.
Traditional keyword search works reasonably well when users already know the wording used in the source. Policy questions do not work that way.
Someone may search for "rules for dormer windows in a conservation area", while the relevant planning document refers to "alterations to roof surfaces within protected townscape". A literal search engine may struggle to connect the two.
Versioning creates another problem. A policy document from several years ago may have been amended or replaced. Finding a relevant paragraph is therefore not sufficient: users also need its date, status and context. An answer based on the wrong version can be technically well retrieved and operationally useless.
RAG searches by meaning, not just keywords
Retrieval augmented generation, or RAG, combines information retrieval with generative AI.
Municipal documents are first processed and divided into smaller sections. Those sections are represented in a form that allows the search system to compare their meaning. When a user asks a question, the retrieval layer selects the passages that are most relevant to the request. Only then does the language model generate an answer.
Suppose an employee asks: "Which exceptions apply to restaurant terraces in the town centre?" The system might retrieve relevant sections from the local ordinance, hospitality policy, implementation rules and a council decision. The language model can turn those passages into a concise explanation while being instructed not to make claims that are unsupported by the retrieved material.
A well-designed municipal RAG system can then show:
- the source document;
- the supporting passage;
- document date or version;
- page number;
- a link to the original.
This is fundamentally different from asking a general-purpose chatbot to answer questions about local policy.
Citations are part of the control mechanism
For public-sector knowledge systems, traceability should be architectural rather than cosmetic.
Language models can produce statements that appear authoritative while being wrong. Grounding the model in retrieved documents reduces that risk, but RAG does not make hallucinations impossible by itself. Additional safeguards are required.
The system should be able to recognise when no sufficiently relevant source has been found. A citation shown to the user should actually support the claim being made. The retrieved passage also needs to belong to the correct version of the policy or regulation.
When the municipal collection cannot substantiate an answer, "I cannot determine this from the available sources" is preferable to an educated guess.
The Woo creates an information challenge, not permission to expose every file
Searchability is closely connected to a broader public-sector responsibility: maintaining a usable information estate.
The Dutch Open Government Act, the Wet open overheid or Woo, applies to municipalities. It covers both disclosure on request and proactive publication, while also placing importance on government information being properly organised and findable. Mandatory proactive publication for specified information categories is being introduced in phases.
That does not mean every municipal document should be added to a publicly accessible AI search service. Permit files, objections, correspondence and case documents may contain personal data or information subject to disclosure restrictions. The Woo itself includes grounds for withholding information, including protection of personal privacy. Documents intended for publication may therefore require a separate publication version from which protected information has been removed or redacted.
A sensible architecture separates public and internal collections. A citizen searching the municipal website may have access only to published documents. An authorised caseworker may be permitted to search additional records. RAG should inherit access controls, not bypass them.
Data sovereignty is an architectural decision
For municipalities, data sovereignty involves considerably more than the physical location of a server. Important questions include:
- where source documents and vector indexes are stored;
- where prompts and generated answers are processed;
- which parties can access the data;
- which logs are retained;
- whether data may be used for model training or product improvement;
- which subprocessors are involved;
- who controls administrative access and encryption keys.
A hosted AI service may be entirely reasonable for information that is already public. Internal policy material and case files containing personal or confidential information require a different risk assessment.
An on-premises architecture provides the highest level of technical control because document storage, retrieval, embedding models and language models can all remain inside infrastructure controlled by the organisation. Private hosting within the EU can provide a useful middle ground where fully local operation would create too much operational overhead.
A US-based AI provider should not be rejected solely because its headquarters are in the United States. The actual processing locations, contractual arrangements, data flows, retention policies and technical controls matter more. For sensitive municipal data, however, an architecture in which those flows can be explicitly governed and audited will generally be preferable to an unrestricted consumer-oriented cloud AI service.
Searching policy is not the same as making decisions about citizens
A RAG assistant that retrieves and summarises municipal documents is a very different system from an AI application that evaluates residents or influences administrative decisions.
Under the EU AI Act, certain public-sector AI applications can qualify as high-risk. Annex III includes, among other categories, specific systems used in relation to essential public services and benefits and certain law-enforcement applications. Whether a municipal system falls into one of these categories depends on its intended purpose and use; the fact that AI is used for enforcement, supervision or decision support does not by itself make every such system high-risk. That distinction should be considered before deployment.
"Find the applicable sections of this subsidy regulation" is not equivalent to "decide whether this resident qualifies for the subsidy". Likewise, giving an inspector fast access to the relevant rules is materially different from using AI to score individuals or automatically determine who should be investigated or sanctioned.
Procure auditability, not just an AI chatbot
When municipalities procure these systems, the open-source-versus-vendor debate is only part of the decision. The more important question is how much control remains with the municipality.
Procurement teams should ask practical questions. Can the system show exactly which documents supported each answer? Does it preserve document versions? Can an incorrect answer be reconstructed later? Are retrieval and generation logs available for audit? Can permissions from source systems be enforced? Can the search index be rebuilt or exported? Can the language model or hosting provider be changed without rebuilding the entire knowledge architecture?
Open-source technology can make local deployment, inspection and supplier independence easier. The trade-off is that the municipality or its implementation partner assumes more responsibility for security, maintenance and upgrades. A managed vendor platform can reduce operational complexity, but the contract then needs to address data portability, processing conditions, logging, audit access and dependency on proprietary components.
For municipal information systems, the most impressive chatbot demonstration is rarely the best procurement criterion. A system that can consistently reconstruct its sources, access permissions, document versions and answer history is often far more valuable.
A practical starting point is a clearly bounded collection such as policy papers, local regulations and public council documents. Add a RAG layer that links every substantive answer back to the original source and refuses to speculate when supporting evidence is missing. Keep public information separate from internal and person-specific case files, and design authorisation into the retrieval architecture from the outset. For sensitive collections, favour an architecture in which storage, processing and logging remain demonstrably under organisational control, whether on-premises or in an appropriately configured EU-hosted environment. Once AI is used for supervision, enforcement, access to public services or decisions that materially affect citizens, its legal classification needs a separate assessment. The same applies to determining what must or must not be disclosed under the Woo: have this reviewed by a lawyer.