AI projects for government succeed when they start from a narrow administrative task, keep personal data under clear control and are procured with operations, accessibility and auditability built into the contract. The technology is rarely the bottleneck; data protection, the EU AI Act, procurement rules and long-term maintenance are, so plan for them from the first concept.
On this page
Public administrations face rising caseloads, staff shortages and citizens who expect digital services that work like the apps they use privately. AI can help with all three, but government projects operate under constraints that private companies rarely face at once: public procurement, strict data protection, transparency obligations, accessibility law and systems that must run reliably for many years. This guide covers what makes public sector AI development different and how to structure a project so that it reaches production instead of ending as a pilot.
Where AI helps in administration
The most productive use cases are narrow and document-heavy: classifying and routing incoming mail, summarising case files for a clerk, drafting standard replies for review, extracting data from forms, answering citizen questions from published regulations, and searching internal knowledge bases. These tasks share a pattern: a human stays responsible for the decision, AI reduces the reading and writing effort, and quality can be measured against existing case outcomes. Starting with such tasks also builds the internal experience and trust needed for more ambitious projects later.
Use cases that make or strongly influence decisions about individuals, such as eligibility for benefits, require far more caution. Under the EU AI Act several of them fall into the high-risk category, with obligations for risk management, data governance, human oversight and documentation. Our practical primer on the EU AI Act for software teams explains how to classify a system.
Data sovereignty and protection
Public bodies process sensitive personal data, and many are required or strongly expected to keep it within the EU or under their own control. That rules out casually sending case files to any model API. Practical options include models hosted in EU data centres under a data processing agreement, open-weight models running on infrastructure the administration controls, and anonymisation or pseudonymisation before data reaches any model. Local anonymisation tools such as Textwash Pro remove names, addresses and identifiers on premises, so that only cleaned text enters the AI pipeline. In Germany, cloud providers are frequently assessed against the BSI C5 criteria catalogue; ask any vendor which attestations their hosting holds.
Platform first, use cases second
A recurring mistake is buying a separate AI tool for every use case. Each one needs its own procurement, data protection assessment and operations contract, and the administration ends up with a patchwork. A more sustainable approach is a governed AI platform with central access control, logging, model choice and policy settings, on which individual use cases are built. KalaidoCreate is designed for this: an all-in-one AI platform for enterprises and the public sector with compliance features, governance and custom APIs that connect to existing specialist procedures. Central governance also makes it far easier to answer audit questions and parliamentary inquiries about how AI is used.
Procurement and contracts
Public procurement rules require transparent competition above certain thresholds, and AI projects fit poorly into specifications written years in advance. Two approaches help. First, separate a short, fixed-scope proof of concept from the later production contract, so that the specification for the larger tender is based on evidence. Second, write outcome-oriented requirements: accuracy on a defined test set, response times, availability and data residency, rather than prescribing a specific model that may be outdated by the time the contract starts. Include operations in the tender. A production AI system needs monitoring, model updates and incident handling, and these belong in a service level agreement; our SLA and public sector page describes what such agreements typically cover.
Accessibility, transparency and staff acceptance
Citizen-facing services must meet accessibility requirements, in the EU usually measured against EN 301 549 and WCAG, and in Germany additionally the BITV 2.0. Chat interfaces are not exempt; they need keyboard operation, screen reader support and plain language. Transparency matters as well: citizens should know when they are interacting with an AI system, and clerks should see which sources an AI summary is based on. Finally, involve staff councils and the people who will use the system early. Projects that present AI as support for overloaded clerks, with clear human control, gain acceptance far faster than those perceived as replacement.
A delivery path that reaches production
- Select one narrow task with measurable volume and a clear human owner.
- Run a data protection impact assessment and AI Act classification before building.
- Build a proof of concept on real but protected data, with a test set drawn from past cases.
- Decide hosting, model and platform based on results, then procure production with operations and SLA included.
- Roll out to one department, measure, and only then extend.
Development teams need new skills for this too. Our article on agentic coding training for public sector IT teams describes how internal IT can build and maintain such systems. OnlineAppCreator, part of the ISO 9001 certified jocapps GmbH, delivers B2G projects through its custom development services, with contracts and SLAs shaped for public sector requirements.
Frequently asked questions
Can public authorities use cloud-based AI models?
Often yes, if hosting is in the EU, a data processing agreement is in place and the data protection assessment allows it. Sensitive data may require anonymisation first or self-hosted models.
Is a citizen chatbot a high-risk system under the EU AI Act?
A chatbot that answers general questions is usually not high-risk but has transparency obligations. Systems that influence decisions about benefits or rights can be high-risk.
How should an AI pilot in government be procured?
Separate a small, fixed-scope proof of concept from the production contract, and write the production tender with outcome-based requirements, operations and an SLA.
Which accessibility standard applies to public sector AI interfaces?
In the EU, EN 301 549 referencing WCAG is the common benchmark, and German federal bodies additionally apply BITV 2.0.