3,611 AI use cases in US federal agencies.

The count doubled in a year

US federal agencies reported 3,611 AI use cases for 2025, up from 1,757 a year earlier, according to the OMB federal use-case inventory. 858 of them use generative AI and 117 agentic AI. 445 are classed as high-impact, the category that triggers risk practices for agencies. Government AI is no longer a lab exercise. It is a portfolio, and portfolios need owners.

The topic mix tells you where the work is. Science leads with 21.0% of use cases, followed by administrative functions and IT. Service delivery, the part citizens actually see, is still small.

Source: OMB inventory raw data, status 13 Apr 2026; ETHORITY tally; 16.7% without topic.

The UK moves in the same direction from a lower base. The National Audit Office found in March 2024 that 37% of 87 central government bodies had deployed AI, and 70% were piloting or planning it. Analysing about 500 public consultations a year takes UK government roughly 75,000 working days, which is why a consultation tool became one of the first AI projects with a clear time account.

Germany's driver is people, not technology

Germany's public service employs about 5.5 million people, 12% of everyone in work, according to Destatis. The civil servants' federation dbb counts 570,000 missing posts today and 1.4 million retirements within ten years. 43% of municipalities report an acute need for staff, 61% of large cities.

570,000 posts missing in Germany's public service (dbb, Jan 2025)

That is the frame for every AI decision in German administration. AI is a capacity strategy first. Every hour it returns has to land in case work, citizen contact or the backlog, or it is not value at all.

Citizens are ahead of their administrations. In Bitkom's October 2026 survey, 74% of Germans say authorities take too long and 70% want AI used when their application is processed, up from 63% a year earlier. Only 12% want AI to decide alone. That is the public mandate in one line: AI prepares, a person decides.

The people side breaks first when agendas move fast: half-adopted tools, skills that lag the system, staff who were not involved and decisions no one owns. A pilot charter therefore needs a workforce annex. It names the case-work and service roles the pilot touches, measures today's processing times, backlogs and error rates before anything changes, plans training for operating, reviewing and challenging the system, and names the official accountable for every decision that affects a citizen. It also says where returned time goes: complex cases, citizen contact, backlog or training, and who decides that. Without it, saved minutes evaporate into the next queue.

Municipal adoption trails. 13% of German municipalities used AI or automated systems in 2024, up from 8% a year earlier, the DStGB/iit Zukunftsradar shows.

Build once, share widely

The strongest German pattern is shared, sovereign platforms rather than a hundred separate chatbots. Hamburg's assistant LLMoin reaches 40,000 staff, and four more states adopt it. Baden-Württemberg's F13 is open source. In May 2026 the federal government commissioned a sovereign AI platform from Deutsche Telekom and SAP as part of the Deutschland-Stack.

For a state ministry or a district office, the buying question changes. It is no longer "which assistant?" but "which shared platform do we join, where is the model hosted, and how do we leave?". Hosting, model choice and exit terms decide long-term control more than any feature list.

Seven AI offerings for citizens, four for police and rescue

The Good AI Prism maps 22 offerings in the Public Sector area, each with a dated source from the provider's own pages. Seven serve citizens through chat, phone and participation, among them Citibot, CivicPlus Agent, convaise Studio, fonio and Go Vocal. Five are administrative assistants and sovereign platforms: Assistent.iQ from Bundesdruckerei, F13, Fabasoft eGov-Suite, LLMoin and PhariaAI from Aleph Alpha. Four support police and emergency services, including tools from Axon, Mark43 and Motorola Solutions. Four are data and decision platforms such as Palantir Gotham and Tyler AI, and two handle documents and procurement. 10 of the 22 offerings come from providers headquartered in Europe, seven of them in Germany, and 12 from providers headquartered in North America. The order is alphabetical, not a ranking. The full map sits in the Public Sector sector brief.

The market moves while you buy. Axon acquired Prepared in October 2025. Cohere and Aleph Alpha agreed to merge in September 2026. A procurement file written in spring can describe a different company by autumn, so tender documents need change-of-control and exit clauses as standard.

The rules have dates

Europe's AI Act, as amended by the Digital Omnibus on AI in force since 27 July 2026, sets the calendar. Prohibitions, including social scoring, apply since 2 February 2025. Since 2 August 2026, people must be told when they deal with an AI system. From 2 December 2027, AI that decides eligibility for public benefits, supports law enforcement, migration, asylum and border control, or the administration of justice carries high-risk duties under Annex III, and public deployers run a fundamental-rights impact assessment under Article 27. High-risk systems already in use by public authorities must fully comply by 2 August 2030.

2 Dec 2027 EU high-risk duties for AI in benefits, policing, migration and justice apply

Germany adds its own building blocks: the federal guideline on AI with human oversight, disclosure and registration; the MaKI marketplace and register; and section 35a of the Administrative Procedure Act, under which a fully automated administrative act needs a legal basis and leaves no room for discretion. The UK made algorithmic transparency records mandatory across departments. Australia's policy since 15 December 2025 requires a use-case register, an accountable owner and an impact assessment.

The AI Act classifies uses, not products. The same assistant is minimal risk when it writes a letter and falls under the high-risk duties when it decides a benefit claim. That is why the inventory comes before the tender.

In practice

Use case: one assistant, many administrations. Hamburg's assistant LLMoin reaches 40,000 staff, and four more states adopt it. One shared platform settles hosting, model choice and exit once, instead of in a hundred separate tenders, and every new administration starts with a tested setup.

Scenario: from letter to decision. A district office uses an assistant to write letters to citizens, a minimal-risk use. When the same assistant is asked to pre-sort benefit claims by eligibility, it becomes a high-risk use under Annex III from 2 December 2027, with registration, human oversight and a fundamental-rights impact assessment. Because the inventory records uses rather than tools, the change is visible before it happens.

What to do Monday

1. Inventory every use. List each AI use, including tools staff bring themselves, and record it in a register such as MaKI or the UK standard.
2. Separate assistance from decisions. Mark every use that touches benefits, policing, migration or justice, and plan the impact assessment for December 2027 now.
3. Join before you buy. Check shared platforms such as LLMoin, F13 and the federal platform before a new tender; write hosting, model choice and exit into every contract.
4. Measure what citizens feel. Record processing time, backlog, appeals and satisfaction before a pilot, and scale only on the measured change.
5. Bring people in early. The staff council, data protection and the named official accountable for each decision belong in the pilot charter from day one.

The Good AI Prism Global Market Study puts the public sector next to ten other sectors; the Education topic area shows the same rule dates applied to schools.