AI for Public Sector Bodies: Records, Citizen Communication, and Procurement

Few areas carry as heavy a text load, combined with as much of a staffing shortage, as public administration. Applications, statements, memos, notices, meeting papers, citizen inquiries – the work consists mostly of reading, summarizing, and drafting documents. At the same time, special requirements apply here: proper record-keeping, traceability, equal treatment, data protection, and, depending on the task, ties to specialized procedures.

This doesn’t rule out AI. But it does determine how it’s used.

The framework conditions that determine everything else

Data sovereignty and hosting

For public bodies, where processing takes place isn’t a matter of preference. The requirements of supervisory authorities, municipal data protection rules, and often the regulations of the respective state mean that, in practice, processing within Germany – or at minimum within the EU – under a solid data processing agreement forms the baseline. Anyone who doesn’t clarify this question at the start clarifies it at the end – and more expensively.

Proper record-keeping

Whatever forms the basis of a decision belongs in the file. If an assistant helps draft a memo, the result must find its way into the file, and it must remain traceable what it’s based on. Systems that cite sources are clearly at an advantage here over ones that only output an answer.

No decisions by the machine

Administrative acts are issued by people. An assistant prepares, drafts, finds references – it doesn’t decide. This separation should not only be practiced but put in writing.

Involving the staff council

What applies to a works council applies equally to a staff council (Personalrat). Early involvement avoids late roadblocks.

Use cases that have proven themselves

Research in your own holdings

The least spectacular case, and at the same time the most effective. Anyone with a question about a past decision, a bylaw, a policy, or a comparable matter today finds the answer by asking experienced colleagues. That exact knowledge is lost with every retirement.

An assistant connected to your own holdings makes cases, templates, and rulebooks searchable – with source citations and within the relevant permissions. This isn’t automation, it’s knowledge preservation, and it works immediately.

Meeting papers and minutes

Drafting a paper in the office’s standard structure from the current status, background, and proposed resolution saves considerable time. So does summarizing long attachments for committee members. Professional and political responsibility naturally stays with the office.

Citizen communication

Two levels need to be distinguished. Internally: drafts for reply letters to recurring inquiries, reviewed and approved by a caseworker. This is uncritical and provides noticeable relief.

Externally: an assistant on the website that answers questions about opening hours, required documents, fees, and responsibilities. Three points are mandatory here – a clear notice that this is an AI system, as Article 50 of the EU AI Act has required since August 2026; limiting it to maintained, approved content; and a visible path to a human. Keep expectations realistic: such an assistant answers standard questions and relieves the switchboard. It doesn’t replace advice.

Plain language and accessibility

An often overlooked, very rewarding case: converting notices and information texts into plain or easy-to-understand language. The need is legally anchored, the manual effort is high, and the results are easy to check. Multilingual versions of information sheets can also be produced much faster this way – with expert review before publication.

Procurement documents

On the procurement side, an assistant helps draft and review specifications: Are the requirements free of contradictions? Is the description worded in a product-neutral way? Is any information missing that bidders need for their calculations? For incoming bids, it can help compile the review against formal criteria.

Evaluation itself remains the procurement office’s task. That’s not just caution, it’s mandatory: procurement decisions must be justified and verifiable.

What doesn’t work

• Automated notices without a human decision. Even where administrative procedure law allows fully automated processes within narrow limits, those are bound decisions made under clear rules – not a field for generative systems.
• Assessment of people. Whether for hiring, appraisals, or performance evaluations: a high-risk area under Annex III, with full requirements from December 2027.
• Information without a verified source. Incorrect legal information given to a citizen isn’t a minor flaw.
• Sensitive proceedings. Social benefits, youth welfare services, discretionary regulatory matters – this is not where you should start.

A realistic starting point

The case with the best benefit-to-risk ratio is almost always internal research. One office, one clearly defined set of documents, one assistant answering questions about it with source citations. No external effect, no decision, no personal-data risk – but immediately noticeable relief.

This case builds all the groundwork for the next ones: hosting clarified, a vetted data processing agreement, staff council involvement, employee training under Article 4, and a permissions model that holds up.

As a corporate group with roots in the systems-house business, we know the framework conditions for public sector clients from practical experience – from the question of where processing takes place to involving the staff council. Get in touch if you’d like to examine a first use case.

The preliminary check that should precede every project

Before a public body starts talking about tools, a short preliminary check along five questions is worthwhile. It can be answered in a single meeting and prevents most later roadblocks.

• Which task, exactly? Not “AI in public administration,” but a named process with a named organizational unit.
• Which data? Internal to the administration, personal, especially sensitive – and in which procedure does it sit?
• Who decides? The separation between preparation by the system and decision by the human is put in writing.
• Where is it processed? Processing location, contracting party, subprocessors – coordinated with the agency’s data protection officer.
• Who needs to be involved? Staff council, data protection officer, and, where applicable, IT security officer and equal opportunities officer.

Procurement: what belongs in the specification

Anyone putting such a project out to tender should explicitly include four requirements beyond the usual points: checking permissions at the moment of the query, citing sources for every answer, logging with defined access rights, and the ability to export configuration and content at the end of the contract.

These four points distinguish usable bids from ones that cause problems in operation – and they can be worded in a product-neutral way.

Frequently asked questions

Are we even allowed to use a cloud solution?

That depends on your state’s regulations, your supervisory authority, and the type of data. Blanket answers don’t help. In practice, the requirements often lead to processing within Germany with a clearly named contracting party – clarify this early with your agency’s data protection officer.

What about fully automated procedures?

Administrative procedure law allows them within narrow limits – for bound decisions made under clear rules. That’s something different from a generative system, and not a use case for a language assistant.

How should we handle a citizen-facing chatbot?

Three things are mandatory: a clear notice about the AI system (a transparency requirement since August 2026), limiting it to maintained, approved content, and a visible path to a human. Set expectations realistically: standard questions yes, advice no.

What should we start with?

With internal research in a clearly defined set of documents. No external effect, no decision, immediate relief – and it builds the groundwork for further cases.

Accessibility as a natural starting point

A use case that’s regularly overlooked in public administration, even though it’s professionally uncontroversial and legally required: converting texts into plain and easy-to-understand language.

The need is real. Notices, forms, and information sheets are hard to understand for a significant share of recipients, and the requirements for accessible administrative communication are increasing. Doing this manually is labor-intensive, because it demands linguistic care and subject-matter understanding at the same time.

This is a favorable case for AI use for three reasons. First, the professional substance already exists and has been reviewed – it’s about rewording, not content. Second, the result is easy to check: a human reads it and immediately sees whether it’s understandable. Third, there’s no decision risk, because the legal effect stays with the original notice.

Two rules should apply here: the simplified version supplements the legally binding text, it doesn’t replace it. And every version is reviewed and approved by an expert before release – a mistaken simplification of a legal consequence is worse than a hard-to-understand original text.

The same pattern applies to multilingual information sheets. Here too: translation as a service, the German version stays binding, and approval comes from a human with sufficient language proficiency.

Inter-municipal cooperation

For smaller municipalities, the effort for a standalone project is often hard to justify – the groundwork is required regardless of population size. That’s why it’s worth looking at shared approaches.

Three forms are particularly practical. A joint procurement: several municipalities tender jointly and share the conceptual work, but each operates its own separate holdings. Then using a municipal IT service provider, provided one offers a suitable service. And exchanging experience without shared technology – often valuable in itself, because requirement lists, vendor questionnaires, and works agreements don’t have to be created from scratch each time.

Important for all shared approaches: the data holdings stay separate. What’s shared in the cooperation is concepts and contracts, not case files.

Working with the staff council

Involving the staff council isn’t a formality to be checked off at the end. In practice, it determines whether a project goes live within a quarter or within a year.

It works well to involve the staff council already when selecting the use case – not for approval, but for shaping it together. The points typically covered there need to be clarified anyway: purpose and scope, an explicit exclusion of performance and behavior monitoring, access to logs, retention periods, employee training, and a fixed date on which the agreement is jointly reviewed.

Anyone who takes this path generally ends up with a workable agreement. Anyone who plans involvement as the last step before rollout ends up with a delay.

A word on managing external expectations

Public administrations are under particular scrutiny on this topic. A citizen assistant that gives incorrect information isn’t a footnote – it becomes the example of “the administration now having machines handle citizens.”

That’s why it pays to be understated in external communication. An assistant announced as help with standard questions, which visibly refers to a human for everything else, doesn’t create inflated expectations. One marketed as a modernization project has to deliver what the announcement promises.

Internal communication matters just as much: employees should learn what’s being introduced and why before the public does – not from the newspaper.