BRENNEN

Industries / Government, education and community / Government / GovTech

Case management

The systems that caseworkers and agencies use to take in applications, decide who qualifies, keep each case's record and send notices, for welfare, health coverage, unemployment pay and child protection.

7

Milestones achieved, from 1980 to 2026.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1980

    Washington pays for computers

    A 1980 federal law let the US government pay 90% of what states spent building computer systems for Aid to Families with Dependent Children (AFDC), then the main cash welfare program, and the Department of Health and Human Services offered states a model design called FAMIS. In a February 1994 review of states' automated systems for federally supported welfare programs, GAO counted 52 "states" (its word) running a statewide automated AFDC system by June 1993, and found that federal agencies had contributed over $6.8 billion to such systems from 1984 to 1992. In 1981 GAO had warned that FAMIS "has not been pilot tested to demonstrate its feasibility".

    LessonFederal money paid for a decade of state building before the model design had been tried.

    gao.govgao.govgao.justia.com

    Credit

    US Congress (Social Security Disability Amendments of 1980); US Department of Health and Human Services

  2. 2009

    A call-center welfare plan ends

    On 27 December 2006 Indiana signed a 10-year, $1.3 billion contract with IBM and partner firms to decide welfare eligibility through call centers and the web, "without the need for face-to-face meetings with a case worker". The state ended the contract for cause, effective 14 December 2009, and "moved eligibility determinations back to the local office". In March 2016 the Indiana Supreme Court held that IBM's breaches "were material in light of the MSA as a whole".

    LessonEligibility work moved to call centers still needed local caseworkers, and the state had to bring them back.

    cases.justia.com

    Credit

    Indiana Family and Social Services Administration; IBM and its partners

  3. 2015

    Fraud found by default

    Michigan's unemployment agency switched on MiDAS, a system built on FAST Enterprises' GenTAX software, in October 2013. Its specification called for it to "automatically adjudicate all issue types without staff intervention", including when a claimant did not answer a questionnaire. A later review of October 2013 to September 2015 found that 85% of the 40,195 fraud findings the system made on its own were wrong, IEEE Spectrum reported. The agency shut off the automatic decisions in August 2015, and in October 2022 the state agreed to pay $20 million to settle a class action, shared by about 3,200 people.

    LessonA rule that reads silence as guilt, run without review, turns into tens of thousands of false accusations.

    mied.uscourts.govspectrum.ieee.orgmichiganpublic.orgbridgemi.com

    Credit

    Michigan Unemployment Insurance Agency; FAST Enterprises; findings reported by Robert Charette in IEEE Spectrum

  4. 2019

    A risk score evaluated

    In 2016 Allegheny County, Pennsylvania, began giving the screeners on its child-abuse hotline a predictive risk score, the Allegheny Family Screening Tool, and it contracted Stanford University for an independent evaluation. The evaluation, summarized by the county in 2019, found that the tool "did not lead to increases in the rate of referred children screened-in", that more of the children screened in were those later found to need services, and that the gap between Black and white children in case opening rates narrowed, but it found "no evidence" that screening became more consistent.

    analytics.alleghenycounty.us

    Credit

    Allegheny County Department of Human Services; Vaithianathan, Putnam-Hornstein and colleagues at the Centre for Social Data Analytics, who built the model; Jeremy Goldhaber-Fiebert of Stanford, who led the evaluation

  5. 2021

    Load-tested at 2%

    Florida launched CONNECT, its unemployment claims system built by Deloitte, on 15 October 2013. It took more than 5.19 million claims after 15 March 2020, and a draft report by Florida's chief inspector general found that the contract called for "a minimum of 200,000 concurrent external users" while testing covered "approximately 4,200", WKMG News 6 reported. In March 2021 the state Auditor General found that "many of the findings disclosed in prior audits" of the system remained uncorrected.

    LessonA system tested at about 2% of its contracted load fails when the full load arrives.

    flauditor.govclickorlando.com

    Credit

    Florida Department of Economic Opportunity; Deloitte, which says it has had no connection to the system for over five years; Florida's Auditor General and chief inspector general

  6. 2023

    Renewals judged per household

    When US states resumed checking every Medicaid enrollee's eligibility in 2023, the Centers for Medicare & Medicaid Services (CMS) found that some states' systems ran automatic renewals, called ex parte renewals, for whole households instead of for each person, so a child who still qualified on the data the state held could lose coverage because a parent did not return a form. On 30 August 2023 CMS told states "to test the renewal logic in your eligibility system", pause such terminations and reinstate the people affected. Thirty states reported the problem, and about half a million children and family members regained coverage, by CMS's count.

    LessonThe law decides eligibility person by person, so a system built around the household has to be tested for that difference.

    medicaid.govcms.govcms.gov

    Credit

    CMS (letter signed by Daniel Tsai)

  7. 2026

    AI notes add false facts

    The Ada Lovelace Institute, an independent research body, published "Scribe and prejudice?" in February 2026, a study of the AI transcription and note-taking tools that UK councils already use in social work, among them Beam's Magic Notes and Microsoft Copilot. In 39 interviews held from March to October 2025, social workers described invented content reaching official records, in one case a note of suicidal thoughts when "at no point did the client actually, you know, talk about suicidal ideation". The institute found that councils' own evaluations focus "on efficiency over the impact on people"; it did not test any tool itself.

    LessonCouncils are measuring the time these tools save, not how often their notes are wrong.

    adalovelaceinstitute.org

    Credit

    Oliver Bruff and Lara Groves, Ada Lovelace Institute

Watch for

  1. California's new child welfare system

    California is replacing CWS/CMS, its statewide child welfare case system, with CWS-CARES, built in separate modules by several vendors instead of under one large contract. When the approach began in 2016 the state's Legislative Analyst's Office (LAO) gave the cost as $449 million and full rollout as April 2019; in October 2026 it said the system "currently is scheduled to go live in December 2026" and put total costs at "around $2 billion through 2026‑27".

    Achieved whenevery county works its cases in the new system and the old one is retired, confirmed by the LAO, the State Auditor or the federal Children's Bureau.

    lao.ca.govlao.ca.gov

    Claimed by

    California Department of Social Services, as reported by the LAO, October 2026

  2. Michigan's MiDAS replaced

    MiUI, Michigan's new unemployment system, built by Deloitte under a $78 million, 10-year contract to replace MiDAS, moved employers over on 23 February 2026 after a delay and was due to go live for benefit claims on 9 November 2026.

    Achieved whenUS Department of Labor timeliness data for the first quarters, or a review by Michigan's Auditor General, shows claims paid on time without a rise in wrongful fraud or overpayment findings.

    sbam.orginsurancebusinessmag.comstatescoop.com

    Claimed by

    Michigan Unemployment Insurance Agency, January and October 2026

  3. Hours saved by AI notes

    Beam, which makes Magic Notes, says the tool is used by "more than 100 local authorities" and saved staff "on average seven hours per week", and Somerset Council says practitioners in a 20-person pilot submitted assessments "65% faster". Neither published a measure of errors in the notes.

    Achieved whensomeone not paid by the vendor publishes measured time savings together with rates of omitted and invented content.

    techuk.orgsomerset.gov.uk

    Claimed by

    Beam, in a techUK article, December 2025; Somerset Council, September 2025

Open

  1. 01

    Test households for renewals

    A public set of made-up households, each with the correct renewal result for every member under federal Medicaid rules: renewed automatically, or sent a form asking only for what is missing. Include children with higher income limits than their parents, members with missing documents and families of mixed immigration status, and ship a deliberately household-level version that the tests must catch.

    The brief
    Credit · unclaimed
  2. 02

    Social-work note benchmark

    A public set of recorded, scripted home visits about made-up families, played by consenting adult actors in several accents, each with a reference note, a list of facts that must appear and a list of claims that must not, and a script that scores AI note-taking tools on what they leave out and what they invent.

    The brief
    Credit · unclaimed

Open 01 · the brief

Test households for renewals

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 households, each with the correct renewal result for every member under federal Medicaid rules: renewed automatically, or sent a form asking only for what is missing. Include children with higher income limits than their parents, members with missing documents and families of mixed immigration status, and ship a deliberately household-level version that the tests must catch.

Why it is still open

CMS told every state to test its renewal logic but supplied two worked examples, a family of two and a family of three. PolicyEngine's open model of US benefits computes Medicaid eligibility but has no renewal or redetermination variables, and KFF, a health policy research group, reports automatic renewal rates ranging "from 99% in North Carolina to 3% in Wyoming" without any public test of the logic behind them.

medicaid.govgithub.comkff.org

Built with what exists

the federal renewal rule (42 CFR 435.916) and the CMS letter; PolicyEngine-US or OpenFisca for income limits; Python with pytest; AI models such as Claude, Gemini or GPT to draft variants from the rule's text, each checked by a person.

  1. First step

    Encode CMS's two examples as machine-readable cases and write ten variants of each.

  2. Done looks like

    at least 100 households, each case citing the clause it tests, a reference version that passes all of them, the household-level version failing at least 20, and sign-off from a Medicaid policy reviewer outside the project.

Open 02 · the brief

Social-work note benchmark

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 recorded, scripted home visits about made-up families, played by consenting adult actors in several accents, each with a reference note, a list of facts that must appear and a list of claims that must not, and a script that scores AI note-taking tools on what they leave out and what they invent.

Why it is still open

The Ada Lovelace Institute "did not conduct a formal technical evaluation of any tool". Public test sets for AI notes exist for doctors' consultations, PriMock57 with 57 mock primary care consultations and ACI-BENCH with 207 doctor–patient conversations, but I found none for social care.

adalovelaceinstitute.orggithub.comgithub.com

Built with what exists

speech-to-text models such as Whisper or ElevenLabs Scribe v3 (see the Note "Fewer word errors"); AI models such as Claude, Gemini and GPT as baseline note writers and first-pass graders; two human raters; text-to-speech voices for extra accents.

  1. First step

    Write ten visit scripts with their fact lists and record three as a pilot.

  2. Done looks like

    at least 30 recordings in at least five accents, invented content labeled by two raters with agreement (Cohen's kappa) of at least 0.7, and scores published for at least three tools or models.