Public-administration case-processing agent (Antragsbearbeitung)

Municipal and state authorities process large volumes of applications for benefits, permits and grants. Many applications arrive incomplete, and caseworkers spend much of their time chasing documents and reading evidence rather than deciding. An agent checks completeness, writes the requests to citizens, maps evidence to the legal criteria and drafts the decision proposal and notice. A human caseworker decides and signs. In the German BMDS pilots this approach cut processing time substantially, while the legal frame keeps the decision with people.

Problem & value

• Today: applications reach the authority through OZG online services, email, post and counter. Staff check forms by hand, send standard letters for missing documents, read payslips, contracts and certificates, calculate entitlements in the specialist procedure and write the Bescheid. Backlogs and long waits are the result; staff shortages in municipalities make this worse.
• AI value: better-prepared applications, fewer rounds of "missing document" correspondence, evidence summarised with page references, consistent application of the same criteria, drafts instead of blank pages.
• Evidence (Germany): the BMDS Agentic AI Hub ran 18 to 20 municipal pilots in 2026. In the Berlin Steglitz-Zehlendorf pilot on Hilfe zur Pflege (startup Myosotis, product formfix, with a care provider), the district reports a 45 % shorter overall throughput time and 31 % less processing time spent by the office on each document [3]. Decisions and approval stayed with caseworkers; the report notes that interfaces to e-file and specialist systems were not consistently available, so workarounds were used [1][2][3]. These are pilot figures from one office, not general benchmarks.

Variants

• Social benefits: care assistance, housing benefit (Wohngeld), basic income, child supplements.
• Permits: building permits (see DOC-016 for the rule-checking step), trade and gastronomy licences, parking and event permits.
• Grants and subsidies: funding programmes, research or cultural grants with scoring criteria.

How it works

1. application intake (OZG/EfA service, BundID, post scan)
2. document parsing + classification (DOC-003)
3. completeness check against procedure checklist
4. agent drafts request for missing documents
5. caseworker releases
6. citizen portal / letter
7. evidence analysis: extract facts, map to legal criteria with page citations
8. retrieval over statutes, rules, internal guidelines (RAG)
9. calculation in the specialist procedure (deterministic, not by the LLM)
10. draft decision proposal + hearing letter if adverse (§ 28 VwVfG)
11. caseworker reviews, edits, decides, signs; Bescheid with reasons (§ 39 VwVfG)
12. E-Akte filing + audit log

• Deterministic core: entitlement calculations and thresholds run in rule engines or the existing Fachverfahren. The LLM extracts facts and drafts text; it never computes the amount.
• Retrieval: statutes, ordinances, administrative guidelines and the authority's own precedent decisions are indexed with hybrid search and a reranker; every claim in the draft cites a source.
• Tools: the agent may read the e-file, query registers where lawfully permitted, and write drafts. Sending letters, issuing decisions and payments are tools gated behind human approval.
• Human-in-the-loop: release of outgoing letters, every decision, any discretionary or judgment-based assessment, and every citizen objection. The caseworker sees the evidence matrix and can overrule each field.

Data preparation

Data needed: incoming applications and attachments, procedure checklists, statutes and internal guidelines, past cases with outcomes, and templates for letters. In authorities this lives in specialist procedures (Fachverfahren), the e-file system (E-Akte/DMS), file shares, mail and still a lot of paper. Typical gaps: scans of poor quality, mixed formats, duplicated and outdated guidelines, knowledge that sits only in experienced caseworkers' heads, inconsistent case categories, and strict access rights per procedure and department.

Concrete steps: scan and OCR remaining paper ([PREP-006], [DOC-014]), build one authoritative, versioned set of rules and guidelines ([PREP-013], [PREP-014]), define the case categories and checklist structure ([PREP-019]), extract outcomes from closed cases as labels ([PREP-020]) and build a test set with caseworkers ([PREP-018]). Pseudonymise historical cases used for testing ([PREP-022]) and check the legal basis for reuse ([PREP-025]). Tools: Paperless-ngx or the existing DMS, OCR/document parsers, a vector store, spreadsheets for the test set.

Rough share of effort: 40 to 50 percent of the project. Expect the rule base and test set to be the bottleneck, not the model.

Implementation paths

Which path for whom

• Public sector: B. Build on a sovereign EU or German cloud with open or EU-hosted models: case data is sensitive and procedures are specific, so no product fits off the shelf, but full self-hosting rarely pays off.
• Enterprise: B. Large permit or claims-like processes need integration with own case systems and rules; cloud APIs in an EU region with the vendor's AVV are the pragmatic route.
• Developer: C. Open-weight models and open-source parsing and RAG run on own infrastructure, which suits prototypes and public-sector delivery projects.

Path A — Buy (off-the-shelf)

• Specialised GovTech startups (DE/EU): Myosotis (formfix, Berlin), plus vendors in the Agentic AI Hub cohort [1]; municipal IT providers such as Dataport, AKDB, ekom21, KDO, regio iT and ITDZ Berlin bundle AI modules and EfA services.
• E-file / procedure vendors: Fabasoft, d.velop, enaio (OPTIMAL SYSTEMS), Axians eGov and other Fachverfahren suppliers add AI assistants to their systems; check their roadmap first.
• Sovereign hosting: openDesk, Delos Cloud, STACKIT, IONOS.
• Choose A when: the procedure is standardised and your Fachverfahren vendor or IT provider already offers it.

Path B — Build on cloud APIs

• Models: Mistral (EU provider, e.g. Mistral Large and Mistral OCR), Aleph Alpha stack via PhariaAI for public sector, or frontier APIs (Claude, Gemini, GPT) in EU regions with zero data retention and a data processing agreement. Classified or highly sensitive social data may exclude non-EU providers.
• Orchestration: LangGraph, Microsoft Agent Framework or n8n with human approval steps; tool access via MCP servers wrapping the E-Akte and Fachverfahren APIs with least privilege.
• Choose B when: you need custom logic across your own systems and your legal basis permits a processor.

Path C — Self-host (open source)

• Models: open-weight LLMs such as Mistral Small/Medium class, Qwen3, Llama 3.x or gpt-oss, served via vLLM; Docling or PaddleOCR for parsing; BGE-M3 embeddings with Qdrant, pgvector or OpenSearch.
• Frameworks: LangGraph, Haystack (deepset, Berlin), Camunda or Flowable for BPMN with human tasks.
• Hardware: HW3 for a 30 to 120B model with several caseworkers in parallel; HW4 for large municipal data centres. Public bodies often prefer this or a sovereign cloud (Path C or B on STACKIT/IONOS).

Path D — Custom training

Only for high-volume, narrow procedures: fine-tune extraction models on the authority's own scanned forms, or a small classifier for routing. Requires labelled historic cases with legal-privacy review. Do not train a model to predict outcomes of individual applications.

State of the art (2026)

• Assistive agents, not deciders: the pilots in the Agentic AI Hub focus on guided application, pre-checking and preparation for caseworkers rather than automated decisions [1][2].
• Reasoning LLMs map free-text evidence to legal criteria well when the rules are in context, and citation-grounded RAG over statutes is the standard design. Weak points: unusual documents, conflicting evidence, outdated rule versions, long case histories and legal interpretation.
• Integration is the bottleneck: missing APIs into E-Akte and Fachverfahren limited the pilots; MCP-style tool wrappers and EfA service standards are the current answer.
• Limits and failure modes: hallucinated legal citations, automation bias (caseworkers rubber-stamp drafts), prompt injection through uploaded documents, inconsistent treatment of similar cases, and disadvantage for citizens with poor-quality documents. Mitigations: source-linked outputs, sampled quality review, evaluation sets built from decided cases, and injection filters on all uploaded content.

Shared resources & combinations

• Same stack: DOC-003 (forms), DOC-016 (rule checking of permits), DOC-014 (RAG over legal texts), one VLM and LLM server for all.
• Front end: CONV-011 (citizen chatbot and guided application) hands complete cases to this agent.
• Governance: AGENT-025 (agent observability, evaluation, audit), AGENT-013 (multi-agent orchestration).

Effort, cost & KPIs

• Path A: months for procurement, public tender rules and integration; licence per procedure or user.
• Path B: 6 to 12 months for one procedure incl. legal assessment, data protection impact assessment and integration.
• Path C: 9 to 18 months, plus a GPU server and a small platform team.
• KPIs: time to decision, number of document-request rounds per case, share of complete applications at first submission, caseworker override rate on drafts, error rate in sampled reviews, objection and appeal rate, citizen satisfaction.

People & change

Involve the department lead and the caseworkers from day one: they own the procedure, define the checklists and later review every proposal. Bring in IT and the information security officer for integration with the specialist procedure and for hosting; the DPO for a DPIA (special categories and vulnerable citizens are likely) and the AVV with any processor (GUIDE-005). Legal/compliance covers administrative-procedure law and the high-risk duties from 2 Dec 2027 for Annex III uses (GUIDE-014).

Staff council: yes. Case-processing tools log caseworker actions and can measure throughput, so co-determination under the BPersVG or state law applies, analogous to §87(1) no. 6 BetrVG; inform early, ideally with a written agreement on no individual performance evaluation (GUIDE-004). In private companies with a works council, §90(1) no. 3 and §80(3) apply.

Start a pilot with one procedure and a small volunteer team (GUIDE-009), name AI champions, and be open that the agent drafts but the caseworker decides and signs. Typical concerns: job loss (position it as relief from routine checks), liability (decision stays with the human), citizen trust (transparent notice). AI literacy: support staff skills through training on limits and automation bias (GUIDE-006, GUIDE-008).

Risks, regulation & compliance

• EU AI Act (high-risk): Annex III point 5(a) covers AI used by or for public authorities to evaluate eligibility for essential public benefits and services, and to grant, reduce, revoke or reclaim them. Art. 6(3) allows a derogation for narrow procedural or preparatory tasks that do not materially influence the outcome, but systems that profile persons stay high-risk. A completeness checker may fall under the derogation; an evidence analyser with a decision proposal usually does not. Document the assessment [5][6]. Deployers that are public bodies must also carry out a fundamental-rights impact assessment (Art. 27). Timing: under the Digital Omnibus on AI (Reg. (EU) 2026/1744) the Annex III obligations apply from 2 Dec 2027.
• Prohibitions (Art. 5): social scoring of citizens and risk profiling for fraud based solely on profiling are prohibited or highly restricted. Do not add benefit-fraud scoring to this agent.
• Administrative law: § 35a VwVfG allows fully automated administrative acts only where a statute permits them and no discretion or margin of appreciation exists [4]. Discretionary and judgment-based cases need a human decision. Also § 24 (investigation principle), § 28 (right to be heard), § 39 (reasons) and the ban on purely automated decisions in GDPR Art. 22.
• GDPR: social and health data are special categories (Art. 9) or protected by SGB X and social secrecy (§ 35 SGB I); legal basis in public-task law, DPIA, no training on citizen data by providers, EU processing.
• Mitigations: release gates, rationale logs, regular bias and error audits, works-council agreement, citizen information.

Getting started

1. Pick one rule-based procedure with high volume and many incomplete applications; map it as BPMN.
2. Run the AI Act and GDPR assessment first (Annex III 5(a), Art. 6(3), DPIA, § 35a VwVfG).
3. Build an evaluation set from 100 to 300 past cases with the final decisions.
4. Pilot the preparation steps (completeness, requests, evidence summary) before any decision proposal.
5. Measure override rate and time saved with caseworkers; expand only when the interfaces to E-Akte are stable.

Sources

1. BMDS — Agentic AI Hub — https://bmds.bund.de/themen/kuenstliche-intelligenz/agentic-ai-hub
2. BMDS — Ergebnisbericht der 1. Pilotierung des Agentic AI Hub — https://bmds.bund.de/fileadmin/BMDS/Dokumente/48-Anlage-Ergebnisbericht_Agentic_AI-Hub_.pdf
3. Bezirksamt Steglitz-Zehlendorf — Pilotprojekt formfix erfolgreich abgeschlossen — https://www.berlin.de/ba-steglitz-zehlendorf/aktuelles/pressemitteilungen/2026/pressemitteilung.1689443.php
4. § 35a VwVfG — https://www.gesetze-im-internet.de/vwvfg/__35a.html
5. Regulation (EU) 2024/1689 (AI Act), Annex III — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
6. Bundesnetzagentur — High-risk AI systems — https://www.bundesnetzagentur.de/EN/Areas/Digitalisation/AI/09_HighRisk/start.html