AI agent training for caseworkers in public sector in Germany

German public sector caseworkers need to learn to check the agent's file summaries and draft notices against the record, use their discretion where the law requires it, and refuse to let any decision about a person's rights leave their hands. The EU AI Act requires AI literacy for these staff, and German administrative law sharply limits automated decisions.

What does a caseworker in German public administration do?

Caseworkers in job centres, social welfare offices, youth welfare offices and housing benefit departments process applications, assess entitlement, issue formal decisions, manage ongoing cases and meet clients. Every formal decision is a written notice with reasons and an explanation of how to challenge it, and an objection triggers a review that tests whether the reasons hold.

The files are long and the rules dense. A single benefit decision may depend on household composition, income from several sources, housing costs, prior periods and obligations agreed with the client. Caseworkers also exercise discretion: whether to grant a one-off payment, how to respond when an obligation is missed, whether a hardship applies. That discretion is part of the law, not an inefficiency to remove.

Where can agents help, and at what depth?

File summaries (depth 1). The agent reads the electronic file and summarises household, income, prior decisions and open items, each point linked to its source document.

Calculation drafts (depth 2). For rule-bound calculations, such as income offsets, the agent drafts the calculation and shows each input and its source. The caseworker checks and adopts it.

Notice drafting (depth 2). After the caseworker decides, the agent drafts the notice with reasons and the required explanation of how to object. The caseworker reviews, signs and issues.

Missing document requests (depth 3, once certified). Requesting standard documents with agreed wording is routine work an agent can do after certification on the authority's own cases.

Payment changes (depth 4, held). Any change to what a person receives is held for the responsible caseworker.

What must caseworkers never let the agent do?

Never let it make or issue a decision, exercise discretion, record that a client has failed to meet an obligation, or communicate a sanction. Never let it treat a third party's tip-off or an online post as fact. What caseworkers should never let an AI agent do sets out the list with reasons. The authority sets these as its own limits, so the agent refuses or holds such actions, and the record shows why.

The point is not distrust of the agent. German administrative procedure law permits a fully automated administrative decision only where a statute allows it and no discretion or assessment is involved. Most casework involves both. An agent that drafted a decision a caseworker signed without reading would undermine the decision's lawfulness as well as its quality.

What should caseworkers learn, and what stays with them?

They need to read a verdict and the bound that decided it, to weigh provenance, and to correct the agent well. Provenance in casework is concrete: a pay statement from the employer is signed; data from another authority received through a formal exchange is sealed or signed depending on the channel; the client's statement in an interview is declared; an anonymous tip is external and cannot drive a decision by itself.

Correction matters because rules change and local practice differs. When a caseworker finds the agent applying an outdated allowance or misreading a household rule, the fix belongs in its brief and sources, by the method in teaching caseworkers to bank corrections.

What stays with the caseworker: every decision, every exercise of discretion, every judgement about a person's situation, and the conversation with the client. That makes caseworkers the real control and explains why AI agents will not replace caseworkers: they take reading and drafting, leaving more time for people.

Which German and EU rules shape oversight and literacy?

The EU AI Act (Regulation (EU) 2024/1689) applies directly. Article 4 requires deployers to take measures to ensure sufficient AI literacy among staff who operate AI systems. Under the Act some public uses are high-risk, and anything that bears on eligibility for benefits deserves a careful check against its lists. For high-risk systems, Article 14 requires human oversight by people who understand the system and can override it, Article 26 sets deployer obligations including assigning oversight to competent, trained people, and Article 27 requires certain deployers, including bodies governed by public law, to carry out a fundamental rights impact assessment before use.

GDPR applies throughout. Article 22 restricts decisions based solely on automated processing with legal or similarly significant effects, Article 35 requires a data protection impact assessment for high-risk processing, and Article 30 requires records of processing. German social law gives social data additional protection, and the federal and state data protection authorities oversee public bodies. Where a staff council exists, introducing a system that can monitor employees' work usually requires its involvement.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

A 90-day plan

1. Weeks 1 to 4. Involve the staff council and the data protection officer. Complete the impact assessments the use requires. Caseworkers on the pilot team complete foundation learning with anonymised files. The agent runs at depth 1 on summaries.
2. Weeks 5 to 9. The agent drafts calculations and notices at depth 2. Every draft is reviewed against the file. Weekly sessions discuss misread rules; corrections go into the brief. Objection rates on agent-assisted notices are watched closely.
3. Weeks 10 to 13. Caseworkers who approve held items complete certification and supervised practice. Document requests are tested in Attest. Payment changes stay held.

Use the AI agent training plan template for caseworkers to assign owners. For team leaders, legal departments and IT, see AI agent skills by role.

Where the Academy fits

Article 4 asks authorities to ensure literacy; Pryme Intelligence Academy offers a documented route. Foundations, free and about four hours, covers why agents must be governed and the decision chain. Caseworkers who approve held items take the Operator track, with a two-week practicum on real decisions signed off by an assessor. Certificates last 12 months.