Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

Trelleborg's Welfare Robot

The robot that decides: automation and the vanishing human loop

This copilot is not a copilot - it is the caseworker. For recurring monthly social-assistance reapplications, a rules-based robot logs into the case system as if it were a person, cross-checks the registers, and issues the decision in under a minute, deciding about one in three cases with no human at all. A single deterministic routing rule sends a case back to a human only by exception - changed circumstances, a missing activity plan, a complex or negative outcome. Modeled on the Trelleborg municipality's welfare-decision robot. That one rule is the whole governance surface, and nothing audits it: an error on the no-human path writes straight into the record and the payment, then comes back next month as the baseline that says nothing has changed - so it stays automated. The catch beneath the catch is that the human the exception edge routes to is the first thing the automation removes.

Stylized model of a documented deploymentCaseworker documentation & copilots

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Trelleborg-RPA-class full-decision automation with a conditional human loop network: 8 components and 17 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 4 assumed · 2 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    This models the full-decision-automation-with-a-conditional-human-loop pattern documented in the Trelleborg RPA case file — a rules-based robot that decides recurring reapplications end to end and routes only exception cases back to a human — not a reconstruction of the actual municipal system. These tools are deterministic rules-based automation, not machine learning, and the modeled routing balance is a parameterization, not a measurement.

  • baseline

    The split between the no-human automated write and the exception re-entry to a human is the routing rule, and its sensitivity is the whole governance surface. No published exception-routing, override, appeal, or error rate exists for the automated path, so the balance drawn between the two halves is bounded by the documented roughly one-in-three automation split rather than measured — tightening the rule re-humanizes the system at linear labour cost, loosening it converts caseworker time to unreviewed throughput.

  • baseline

    Each automated decision is read back as the next month's changed-circumstances baseline, so an unchallenged automated error becomes the reference state that keeps the case on the no-human path; the appeals function, added in July 2019, is the main correction channel and fires only downstream, after payment, and only on applicant-contested cases.

  • assumed

    Peer pathways are authored on both signs: one predefined rule set decides every matching reapplication so its defect is systematic rather than idiosyncratic, the downstream appeals review survives as a low check on the upstream decisions, and the standing quality audit of the automated path is drawn as a documented absence — no source records any audit of the automated path's decision quality or of the routing rule's coverage.

  • assumed

    Staffing impact is a documented dispute carried on the map, not adjudicated: AlgorithmWatch reported the benefit-caseworker corps fell from 11 to 3, while the municipality and the vendor describe 2 employees redeployed with no cuts. The residual-caseworker node is the human capacity the exception edge depends on, and the workforce-exit shock at the adopting municipality (12 of 16 caseworkers resigned in protest) shows that capacity can evaporate at adoption time, which is what the staff-turnover stressor represents here.

  • assumed

    Served social-assistance applicants, and the income support they do or do not receive, are not in the dynamics; this Lab models institutional propagation only. A 2024 national study reports RPA case-selection correlating with applicants' country of birth, age, and duration of receipt and coinciding with less generous decisions, but its municipalities are anonymized and its authors attribute the pattern to administrative reorganization rather than the technology, so it is not attributed to Trelleborg and no differential client harm is estimated here; those outcomes are documented in the case file and measured outside any diagram like this one.

What this example does not show

  • Served social-assistance applicants, and the income support they do or do not receive, are not modeled here; the Lab models institutional propagation only, and those outcomes are documented in the case file and measured outside any diagram like this one.
  • No published error, override, exception-routing, or appeal-rate data exists for the automated path. The balance drawn between the no-human write and the human exception re-entry, and any figure for what fraction of machine error bypasses a human, is a parameterization bounded by the documented roughly one-in-three automation split - it is not a measurement, and the system is rules-based automation, not machine learning.
  • The staffing impact is a documented dispute, not adjudicated here: AlgorithmWatch reported the benefit-caseworker corps fell from 11 to 3, while the municipality and the vendor describe 2 employees redeployed with no cuts. The outcome direction is likewise disputed - one report has the municipality considerably reducing the number of people receiving benefits, while the vendor claims 22 percent more people helped - and neither pole is an independent evaluation; the vendor's figures (a 94 percent time saving, 22 percent more people helped) are vendor claims.
  • The 2024 national finding that RPA case-selection correlates with applicants' country of birth, age, and duration of receipt and coincides with less generous decisions comes from four anonymized municipalities, and its authors attribute the pattern to administrative reorganization accompanying adoption rather than to the technology itself; it is not attributed to Trelleborg, and no differential client harm is computed on this diagram.
  • Legality is time-split: before 1 July 2022 the practice was widely read as unlawful for municipal decisions, with only Trelleborg and its lawyers dissenting; the Local Government Act amendment (in force 1 July 2022) permits delegation to an automated decision function going forward and did not adjudicate Trelleborg's earlier deployment.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • 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.

    empirical
    • Investigative AlgorithmWatch, Central authorities slow to react as Sweden's cities embrace automation of welfare management (2020) https://algorithmwatch.org/en/trelleborg-sweden-algorithm/
    • Academic Ranerup and Henriksen, Digital Discretion: Unpacking Human and Technological Agency in Automated Decision Making in Sweden's Social Services (Social Science Computer Review, 2022;40(2):445-461) https://journals.sagepub.com/doi/full/10.1177/0894439320980434
    • Government European Commission Joint Research Centre, AI-Watch use-case record: Trelleborg automated social welfare decisions (2021) https://ai-watch.github.io/AI-watch-T6-X/service/90131.html

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.
  • Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
  • Staff turnover — Experienced skepticism leaves; new staff calibrate their trust on the tool itself.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Caseworker documentation & copilots domain page.

Levers available here and the patterns behind them

Documented case histories