Public sectorDepartment Director

Casework with an AI Colleague

A backlog that hiring freezes made untouchable fell by half — inside EU data rules.

The scenario

A Nordic public agency's casework department carried an eleven-week backlog, a hiring freeze and a political mandate to improve service times. The director's constraint set was unforgiving: strict EU data protection rules, union agreements, and zero tolerance for errors in citizen-facing decisions.

The agency deployed a narrowly scoped AI colleague inside its existing case system. For each case, it produces a structured summary of the file — chronology, key facts, applicable rules with citations — and drafts routine correspondence for the caseworker to edit and send. A knowledge lookup answers procedural questions from the agency's own regulations, always with source references. Everything runs on EU-hosted infrastructure; no case data trains external models; every AI action is logged and auditable.

The boundaries were negotiated with the data protection officer and union before go-live: AI summarises and drafts, caseworkers decide, and citizens are told how AI is used in their case. Caseworkers were trained as reviewers, not replaced as decision-makers — a distinction the union helped sharpen.

Nine months in, the backlog is down by half, average handling time per case has dropped 40%, and the error rate in reviewed decisions has not moved. The agency now fields questions from peers across Europe about doing this properly rather than quickly.

How AI enters the workflow

Case summarisation
AI
On assignment, the AI produces a structured case summary: chronology, facts, open questions.
Rule lookup
AI
Applicable rules and precedents are retrieved from the agency's own regulatory library with citations.
Caseworker assessment
Human
The caseworker verifies the summary against the file and forms their own assessment.
Correspondence drafting
Human + AI
Routine letters are drafted by the AI, edited and approved by the caseworker before sending.
Decision
Human
Every decision on a citizen's case is made and signed by a caseworker, with reasoning recorded.
Quality audit
Human + AI
Random samples of AI summaries are audited monthly against source files; findings tune the system.
Transparency
Human
Citizens are informed how AI is used in casework and can request human-only handling.

Ask about this workflow

ADA, the taskforce deputy, explains exactly how human and AI share the work — ask anything.

The human + agent taskforce

Team leader — approves every deliverable

Taylor

Orchestrator Agent

SENTINEL

Governance Agent

SAGE

Knowledge Architect

PRISM

Readiness & Assessment Agent

Agents propose and execute. The human leader always approves the final result.

Outcomes

Metric	Before	After
Case backlog	11 weeks	5.5 weeks
Handling time per case	3.2 hours	1.9 hours
Decision error rate (audited)	2.1%	1.9%

Key takeaways

Narrow scope inside hard rules beat broad ambition: summarise and draft, humans decide.
DPO and union involvement before go-live converted potential blockers into co-designers.
EU-hosted infrastructure and no-training terms made the legal review survivable.
Quality audits showed AI assistance did not degrade decisions — the number that mattered most.