Domain Atlas / Caseworker documentation & copilots
Trelleborg's Welfare Robot
In Trelleborg, Sweden, the first municipality to fully automate social-assistance decisions, peer-reviewed analysis reports that about 30 percent of digital reapplications are decided entirely by rules-based software with no human review and about 85 percent receive at least partial automated handling; decision time on reapplications fell from roughly two days to under a minute, and a human caseworker re-enters the path only by exception, when a routing rule detects significantly changed circumstances, a missing activity plan or job-seeking documentation, or a complex or negative case. No error, override, exception-routing, or appeal-rate figures for the automated path have been published, so the fraction of automated decisions that ever reaches a human cannot be established from the record.[3]
What happened
In 2015 the municipality of Trelleborg (about 40,000 residents) digitalized its social-assistance (försörjningsstöd) application process; by 2017 roughly three-quarters of applications arrived through the digital platform. In early 2017, building on that platform, Trelleborg became the first Swedish municipality to fully automate benefit decisions. The tool is robotic process automation (RPA) — deterministic, rules-based software, not machine learning, as the peer-reviewed record and the European Commission's AI-Watch use-case record describe it. It was built with the consultancy Valcon by filming a caseworker performing the standard workflow and encoding that sequence as rules; the vendor case study attributes the platform to UiPath (a vendor claim). For recurring monthly reapplications, the robot logs into the ProCapita case-management system as if it were a caseworker, copies the application data, cross-checks it automatically against national registers including tax-agency and housing-support data, computes eligibility, and issues the decision with no human review of the automated path. Peer-reviewed analysis reports that 85% of digital applications receive at least partial RPA handling and 30% are fully automated — about one in three reapplications decided entirely by software — and decision time on reapplications fell from about two days to under a minute. (The AI-Watch record separately notes a pre-automation average wait of 8 days, up to 20, and 300-plus applications a month; its own status field reads "Unknown," and no source reports the system being decommissioned.)
A single deterministic routing rule is the whole human safeguard: a caseworker re-enters the decision only by exception — when circumstances have significantly changed, when a required activity plan or valid job-seeking documentation is missing, or for complex or negative decisions — and rejected applications remain human-handled. Peer-reviewed analysis characterizes the result as a "human-technology hybrid actor" in which caseworkers' discretion over routine cases was eliminated by design rather than overridden case by case. The staffing impact is contested and reported both ways: AlgorithmWatch's 2019 account states the number of benefit caseworkers fell from 11 to 3, while the municipality and the vendor instead describe 2 employees redeployed from repetitive work with no cuts. The direction of the client outcome is also disputed — one report has the municipality "considerably reducing" the number of people receiving benefits, while the vendor claims "22% more people helped than the year before" — and neither is an independent evaluation. The sharpest documented workforce reaction came when the neighbouring municipality of Kungsbacka adopted the Trelleborg model in 2018: 12 of 16 social workers left their jobs in protest (some later returned). An appeals function, the main record-correction channel, was added to the platform in July 2019.
Accountability arrived from outside, and late. From June 2018 the union economist Simon Vinge (Akademikerförbundet SSR) asked Trelleborg how the algorithm decided cases, received only screenshots and a flowchart, and in September 2019 filed a complaint with the Parliamentary Ombudsman arguing the disclosures were not meaningful information about how decisions are made (the complaint's disposition is not documented in the sources gathered here). A journalist won an Administrative Court of Appeal ruling that the RPA source code is an official document under Sweden's public-access principle (offentlighetsprincipen) and must be disclosed over the vendor's trade-secret objection; the code comprised roughly 136,000 lines of rules across 127 XML files. When the municipality forwarded the request, a Danish software company shared the code — and the released code was then found to contain the personal data of about 250 individuals, including names and social-security numbers present since 2017 and visible to requesters and subcontractors; the municipality opened an investigation into the screening failure. A union survey of all 290 Swedish municipalities found only Trelleborg had fully automated decisions; before 1 July 2022, Swedish municipal law was widely read to permit only automated decision support and not automated decisions, "with the exception of Trelleborg and their lawyers." Amendments to the Local Government Act (Prop. 2021/22:125), in force 1 July 2022, permit municipalities to delegate decision-making to an "automated decision function" going forward — regularizing the practice prospectively, without adjudicating Trelleborg's earlier deployment. A 2024 peer-reviewed study of 800 applications across four anonymized Swedish municipalities using RPA in social assistance found that RPA handling correlates with applicants' country of birth, age, and duration of assistance receipt and coincides with less generous decisions that disproportionately affect financially vulnerable groups; its authors attribute the pattern to the administrative reorganization accompanying adoption rather than to the technology itself, and because the municipalities are anonymized the finding is not attributed to Trelleborg.
The sociotechnical reading
Most copilots in this atlas keep a human between the model and the record: the machine drafts or flags, and a person decides. Trelleborg removes that person from the routine path entirely and puts a rule in their place. What is left of human judgment is not an override — no human sees the automated decisions before they take effect — but a single deterministic routing rule that re-inserts a caseworker only on the exceptions it was written to catch. That makes the routing rule, not the model's accuracy, the entire discretion surface, and it fails in a way no other cell here does. First, the rule's blind spot is permanent: whatever it never learned to call "hard" gets no human, ever, and there is no standing audit of the automated path's decision quality anywhere in the record to notice. Second, and unique to a monthly reapplication, an error does not scatter — it renews. Each automated decision is written into the case record and read back next month as the "changed circumstances" baseline, so an unchallenged automated mistake becomes the reference state that reports nothing has changed and keeps the case on the no-human path indefinitely. The one downstream correction, appeals, fires only after payment and only when an applicant contests, which most do not.
The distinct lesson is about the safeguard itself: a conditional human-in-the-loop is not the same as human review, and it corrodes from the one direction its designers do not price. The exception edge depends on a residual human corps to route the hard cases to — and this is the shape where adopting the tool destroys exactly that capacity. Trelleborg's own caseworker count is disputed (11 to 3, or 2 redeployed with no cuts), but the adopting municipality's reaction is not: 12 of 16 social workers resigned in protest. A conditional loop whose human endpoint has walked out is a full-automation loop wearing a human-in-the-loop label. So governing this shape is three moves the Field Guide keeps naming and this deployment did none of by design: audit the routing rule's coverage, not the model's accuracy, because the cases it calls easy are the ones no one will ever see; reconcile a decision against its source evidence before it hardens into next month's baseline, so one wrong month cannot lock itself in; and protect the fallback capacity, because the safety valve degrades exactly as the throughput it guards grows. The tell that none of it was built in is where the accountability came from — a freedom-of-information lawsuit forced the code into the open years after deployment, and the code itself leaked about 250 people's data on the way out. The honest boundary is that nothing here measures the applicants whose benefits were decided; what this map moves is the shape of an institution's own decision-making, and whether the automation changed any single person's entitlement was, by every account gathered here, never independently evaluated.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.