BRENNEN

Industries / Government, education and community / Social services / Foster care and adoption

Case management

Case management is the running work of the caseworker assigned to a child in care: visiting the child, making and reviewing the plan with the family, reporting to the court and keeping the file.

7

Milestones achieved, from 1989 to 2024.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1989

    Families plan with the state

    New Zealand's Children, Young Persons, and Their Families Act 1989 (now the Oranga Tamariki Act) made the family group conference (FGC) the main way to decide what happens after care and protection concerns: the child's wider family, or whānau, meets with social workers and other professionals and agrees a written plan. Aroturuki Tamariki, New Zealand's Independent Children's Monitor, says the conferences were "a direct response to concerns of institutional racism raised in the 1988 report Pūao-Te-Ata-Tū" and were described at the time as "unprecedented". By its count about 4,500 children and young people had a care and protection FGC in 2023/24, about two-thirds of them Māori, and conferences are "not always carried out as intended", with plans not always followed through.

    LessonA legal right for the family to make the plan still depends on staff time and services to carry the plan out.

    aroturuki.govt.nz

    Credit

    New Zealand Parliament (the 1989 Act); reviewed by Aroturuki Tamariki

  2. 2005

    Turnover and time in care

    A January 2005 review for the Bureau of Milwaukee Child Welfare found that turnover among the private agencies' case managers ran at 34% to 67% a year in 2003 and 2004, and that 40% of the 2,899 children still in care at the end of September 2004 had had more than one worker in nine months. Of the 679 children who entered care from January 2003 and left it for a lasting home by September 2004, 74.5% had had one case manager, 17.5% two and 0.1% six or seven. Permanency means such a lasting home: back with parents, adopted or under a guardian. The report read these figures as worker changes lowering the chance of permanency, and said that changes "force clients to start over with new workers".

    LessonFigures drawn only from children who had already left care within 21 months describe those children, not each child's chances, and need that limit stated when they are quoted.

    uh.edu

    Credit

    Connie Flower, Jess McDonald and Michael Sumski

  3. 2007

    Monthly visits counted

    Since federal fiscal year 2007, under the Child and Family Services Improvement Act of 2006, every US state has had to report how often caseworkers visit each child in foster care. Since fiscal year 2015 the visits must reach at least 95% of the visits that would be made if every child were seen once a month, and at least half must take place where the child lives. A state that falls short loses one to five percentage points of federal matching money for each measure.

    acf.govcwoutcomes.acf.hhs.gov

    Credit

    US Congress; the Children's Bureau of the Administration for Children and Families (ACF), which sets the counting method

  4. 2018

    Federal reviews find gaps

    In the third round of the federal Child and Family Services Reviews (CFSRs), held in fiscal years 2015 to 2018, the Children's Bureau read at least 65 cases in each state and interviewed staff and partners. A state conforms on an outcome only if 95% of the cases it applies to are rated a strength. By the CHAMPS campaign's reading of the Bureau's aggregate report, no state reached substantial conformity on the outcomes that most directly affect children in foster care (permanency, placement stability and family connections), and only one state was rated a strength on placement stability.

    LessonReading the case files shows gaps that counts of visits and deadlines do not.

    fosteringchamps.org

    Credit

    Children's Bureau, ACF; summary by the CHAMPS campaign for foster parenting policy (January 2022)

  5. 2020

    Risk models fall short

    What Works for Children's Social Care, England's What Works Centre for the field, spent 18 months building and testing machine-learning models meant to predict from councils' case records whether risks to a child would escalate. Its September 2020 report found that none of the models predicted this accurately, and the centre's own headline was that the models "miss four out of five children at risk".

    LessonTesting models on councils' own records before anyone relied on them showed they would miss most of the children they were meant to find.

    khub.net

    Credit

    What Works for Children's Social Care

  6. 2022

    Oregon drops its risk score

    From 2018 Oregon's Department of Human Services gave hotline workers a numerical score, from its Safety at Screening Tool, when deciding whether a report of abuse or neglect should be investigated; the tool was modelled on Pennsylvania's Allegheny Family Screening Tool. After an Associated Press review raised concerns about racial disparities in such tools, the department told staff in May 2022 that hotline workers would stop using the score at the end of June "to reduce disparities", and replaced it with Structured Decision Making, a set of structured checklists used by many US agencies.

    LessonAn agency that could not show what a score did to which families were investigated chose to stop using it.

    npr.org

    Credit

    Oregon Department of Human Services; reporting by the Associated Press; Senator Ron Wyden, who had pressed the department on bias

  7. 2024

    A chatbot in court papers

    In December 2023 the Department of Families, Fairness and Housing (DFFH) in Victoria, Australia, reported that a child protection worker had used ChatGPT to draft a Protection Application Report filed with the Children's Court. The Office of the Victorian Information Commissioner (OVIC) found that the report described a doll the child's father had used for sexual purposes as evidence that the parents provided "age appropriate toys"; the error did not change the department's or the court's decision. A departmental review found 100 cases in the worker's unit with signs that ChatGPT may have been used, and nearly 900 staff, about 13% of the department, had visited the ChatGPT website from July to December 2023. In September 2024 OVIC ordered the department to bar child protection staff from web-based AI text tools, among them ChatGPT, Claude, Gemini and Microsoft 365 Copilot, and to block them until 5 November 2026.

    LessonStaff were already putting case files into general chatbots before the agency had rules for it, and the error came to light because the report was read.

    ovic.vic.gov.auovic.vic.gov.au

    Credit

    OVIC, Privacy and Data Protection Deputy Commissioner

Watch for

  1. Risk-score pilots in ten places

    ACF says it is giving $6 million, about $600,000 each, to ten jurisdictions (Indiana, Missouri, Nebraska, New Jersey, North Carolina, Ohio, Oklahoma, Texas, the Muscogee (Creek) Nation and the District of Columbia) for three-year pilots of predictive analytics, due to start on 30 September 2026. Uses range from sorting hotline calls to foster placements, permanency and cutting caseworker paperwork, and ACF says each grantee must set up governance, train staff and evaluate results, with the tools used for "decision support, not decision making".

    Achieved whenan evaluator independent of the grantees and of ACF publishes, for each pilot, how accurate the predictions were and whether decisions and children's outcomes changed, broken down by race.

    statescoop.com

    Claimed by

    Alex Adams, ACF Assistant Secretary, August 2026

Open

  1. 01

    Handover files for testing

    A public set of made-up foster care case files, each 100 to 300 pages as a real file might be after two years in care (referral, assessments, visit notes, court reports, placement moves, family contact records), with a list of the facts a new caseworker must not miss (safety concerns, court dates, the child's own wishes, siblings, contact orders), and a script that scores AI handover summaries for facts left out, facts changed and facts invented. The test covers reading a whole file when a case changes hands, not turning a recorded visit into a note.

    The brief
    Credit · unclaimed
  2. 02

    Case clocks as tested code

    Open code that takes a made-up child's dates (entry into care, moves, trial home visits, running away, hearings, visits) and returns every deadline the law sets, each with the clause it comes from. For the US: the federal test of 15 of the most recent 22 months in care, which triggers a petition to end parental rights unless an exception applies, and the monthly visit rule. For England: case reviews within 20 working days of entering care, the second within three months, then at intervals of up to six months. Every rule ships with test cases.

    The brief
    Credit · unclaimed

Open 01 · the brief

Handover files for testing

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of made-up foster care case files, each 100 to 300 pages as a real file might be after two years in care (referral, assessments, visit notes, court reports, placement moves, family contact records), with a list of the facts a new caseworker must not miss (safety concerns, court dates, the child's own wishes, siblings, contact orders), and a script that scores AI handover summaries for facts left out, facts changed and facts invented. The test covers reading a whole file when a case changes hands, not turning a recorded visit into a note.

Why it is still open

Files change hands often: in the Milwaukee review, 40% of children in care had more than one worker in nine months. Published tests of AI models on child welfare records use real agency records that are not released, such as a large Canadian agency's case notes (Moon and seven co-authors, January 2026) and four 500-case sets of Michigan investigation summaries (December 2025). The nearest public synthetic set found is 1,000 made-up East African child helpline call transcripts with summaries. A search for a public set of synthetic case files found none.

arxiv.orgarxiv.orghuggingface.co

Built with what exists

AI models such as Claude, Gemini and GPT to draft the files from family histories written by people, each file checked by a qualified social worker; open-weight models an agency could run on its own machines, such as Mistral Large 4 (see the Note "A trillion open parameters") or small LFM3 models (see the Note "One GPU to fine-tune"), run locally through llama.cpp on ggml (see the Note "A first stable number for local AI"); Python for scoring. Only invented children and families are used.

  1. First step

    Write three family histories with a qualified social worker, generate one file from each, and draft their must-know fact lists.

  2. Done looks like

    At least 30 files, each with 20 or more must-know facts; two raters agreeing on the scoring (Cohen's kappa of at least 0.7); published scores for at least four models, at least two of which run on an agency's own computers.

Open 02 · the brief

Case clocks as tested code

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Open code that takes a made-up child's dates (entry into care, moves, trial home visits, running away, hearings, visits) and returns every deadline the law sets, each with the clause it comes from. For the US: the federal test of 15 of the most recent 22 months in care, which triggers a petition to end parental rights unless an exception applies, and the monthly visit rule. For England: case reviews within 20 working days of entering care, the second within three months, then at intervals of up to six months. Every rule ships with test cases.

Why it is still open

ACF publishes a reviewer brief that explains how to count the 15 of 22 months by hand, leaving out partial months and time on trial home visits, with worked scenarios. The largest open dataset found, Project Milk Carton's US Child Welfare Decision Chain Dataset (version 1.1.1, 2026; 2,142 decision nodes in 51 jurisdictions), records deadlines as text with statute citations and does not compute them from dates. A GitHub code search found only scattered one-line approximations, one of which counts 15 of 22 months as a fixed number of days.

acf.govgithub.comlegislation.gov.uk

Built with what exists

Python, in the rules-as-code style of OpenFisca; the statute and regulation texts; AI models such as Claude or Gemini to draft test cases from the rule text, each checked by a person; GitHub Actions to run the tests on every change.

  1. First step

    Encode the ACF brief's counting steps and turn its two worked scenarios into the first tests.

  2. Done looks like

    US federal rules, three states and England encoded; at least 200 test cases, each citing its clause and including the ACF brief's own scenarios, all passing; a review of the rules by a child welfare lawyer outside the project.