Real-Time Guidance for Public Caseworkers: The Software Exists, the Rulebook Is the Bottleneck

The problem

The anchor paper found that an AI tool suggesting replies to contact centre agents raised issues resolved per hour by 15% on average at one US software firm, with the gains concentrated among the least experienced agents and none among the most skilled (measured without a randomised control group, 2020-21)¹². That was technical-support chat, where a wrong suggestion costs an agent a few minutes. This paper adds a different setting: the caseworker who decides, or advises on, who receives food assistance, housing help or social care, where a wrong answer can change whether a household is paid.

Wrong answers are common at scale. The national payment error rate in the US food-assistance programme SNAP for fiscal year 2025 was 10.62%, against a congressional threshold of 6%. Counting both overpayments and underpayments, that represents $10.1 billion in improper payments (official quality-control data, announced 24 June 2026)³. The rate measures how accurately states decide who is eligible and how much they should receive; the release gives no causes, and this paper does not attribute the errors to caseworkers³.

Staffing adds pressure. State officials told a news organisation in March 2026 that Pennsylvania had nearly 400 open positions in county human services offices, Idaho 40 eligibility-worker vacancies, and Montana had filled 39 of 59 projected positions (state-reported)⁴. Missouri's agency director said the agency had 1,000 fewer front-line workers than about a decade earlier, with more than double the number of Medicaid and SNAP enrollees⁴. Reported CMS data for December 2025 showed about 30% of Medicaid and children's health insurance applications in Washington DC and Georgia taking more than 45 days⁴. In England, child and family social worker turnover was 12% in the year ending 30 September 2025, down from a 17% peak in 2022, and the vacancy rate was 14%, down from 20% (official statistics, published 19 March 2026)⁵. The picture differs by country and role: the English figures are improving, while the US figures describe shortages.

The rules are also moving. Medicaid work requirements of 80 hours a month take effect on 1 January 2027 in most states, with verification every six months instead of yearly⁴.

What already works

The building block is agent assist, described in the anchor paper: a model reads a live conversation and suggests replies and documents to a human who decides¹.

For caseworkers, one evaluation was found. Nava, a US public-benefit technology firm, with Georgetown and Cornell researchers, tested a chatbot that explains programme rules in plain language, retrieves from pre-vetted sources, gives citations, and answers that it does not know when a question is out of scope (developer-led evaluation, published 18 March 2026)⁶. The design combined a randomised trial with 125 caseworkers, who were shown AI-generated responses to hypothetical client questions, and a 14-week quasi-experimental pilot with 61 caseworkers in six organisations in Los Angeles County⁶. The tool was estimated to improve caseworker accuracy by an average of 40%, with larger gains on harder questions; the page does not say which part of the study produced the figure and gives no confidence interval⁶. About 65% of caseworkers with access used it, averaging 14 prompts each, but usage declined over time and varied by site, and the Net Promoter Score (a 0-to-100 style recommendation measure) was 11⁶. Evidence of reduced administrative burden was "promising but inconclusive" and not statistically significant⁶. Answers averaged a 10th to 12th grade reading level, above the recommended 8th, while the source policy manuals require college-level reading⁶.

In plain terms: on made-up questions the assistant helped people answer more accurately, and in real use about two-thirds tried it, but fewer kept using it, and the benefit in daily work is not yet shown. Time saved and error rates were not reported on the page read⁶.

What does not yet exist in the sources found: a published test of a tool that listens to a live conversation with a member of the public and guides a public caseworker. Nava lists a call notes summarizer at proof-of-concept stage as of December 2025⁶.

The possibility

One way this could work: a caseworker in a benefits office or a council housing team takes a call. Speech recognition transcribes it. A language model, limited to the agency's current legislation, policy manuals and forms, shows the relevant rule with a plain-language explanation and a pointer to the manual paragraph, and lists the documents to ask for. If the question lies outside those sources, it says so. After the call, it drafts a case note for the caseworker to correct. In this design the caseworker would decide, and the tool would never decide eligibility.

The role of AI. Three parts would depend on AI: transcribing speech, reading the conversation to find which rule applies, and phrasing the explanation. Most of the rest is ordinary systems and policy work: keeping the rulebook current and structured, connecting to the case-management system, recording and consent notices, access controls, and training and supervision. If the rules were held in a well-maintained, searchable form, part of the benefit might come without AI; no comparison was found.

If the pattern in the anchor study carried over, newer caseworkers would gain most. No study found tests that in benefits work.

What would have to be true

1. The rules would have to be kept current and in a form the tool can be restricted to. The source manuals require college-level reading⁶, and rules change, as the 2027 Medicaid work requirements show⁴. A tool grounded in last year's manual would give last year's answer.
2. Accuracy would have to be shown on live cases. The 40% gain came from a study part that used hypothetical questions, and the page does not separate it from the pilot⁶.
3. Caseworkers would have to keep using it. Usage fell over time. Sites with training, organisational reinforcement and peer support produced more active users⁶.
4. The legal status would have to be settled. The EU AI Act recitals say systems used to determine whether public benefits and services are granted, denied, reduced, revoked or reclaimed are to be classified as high-risk, and set out a derogation for narrow procedural or preparatory tasks that requires documented assessment and registration, and does not cover profiling⁸. The Commission states that Annex III high-risk rules apply from 2 December 2027, after the AI Omnibus entered into force on 27 July 2026⁷. Only the recitals were read, so whether a tool that surfaces rules during a decision counts as preparatory could not be settled here.
5. Recording and transcription rules would have to be met. A US federal court declined on 10 February 2025 to dismiss a class action over contact-centre AI that gives agents a real-time transcript and suggestions, on an allegation that callers were not told; the order decided only that the claims were adequately pleaded⁹.
6. Someone would have to pay and be accountable. No price per caseworker was found in the sources opened.

What could stop it

What the evidence shows

• A public rules chatbot failed. New York City's MyCity chatbot, released in fall 2023 to help business owners check city rules, was found in 2024 testing to suggest that landlords could discriminate against tenants with Section 8 vouchers and that refusing cash was acceptable despite a 2020 city law¹⁰. It was taken down by 4 February 2026; the mayor said it cost around half a million dollars (investigative news report)¹⁰. It served businesses and was not a staff tool, so it is a comparison, not a like-for-like case.
• Wrongly encoded rules harm people at scale. The Royal Commission into Australia's Robodebt scheme reported on 7 July 2023 that the scheme's income averaging neither produced accurate results nor complied with the income calculation provisions of the Social Security Act 1991 (as summarised in a catalogue entry; the report itself was not read)¹¹. That scheme made decisions automatically; guidance for a human is a different design, but both depend on the rule being right.
• Skilled staff may not gain. In the anchor study the most skilled agents saw no productivity gain².
• Money pressure is already set. States at or above the 6% SNAP error threshold are to cover 5%, 10% or 15% of benefit costs, in most cases from 1 October 2027, with FY2025 the first year that can be used³.
• Gaps. The research found no published test of live-call guidance for public caseworkers, no time-saved or error-rate figure for caseworker assistants, and no cost data.

Reasoning

Who bears the risk? If a caseworker, trusting a suggestion, gave a wrong answer, the applicant would carry the cost, not the supplier. The anchor study found adherence to suggestions of about 35 to 38%², so a tool could be right and still be ignored, or wrong and still be followed.

Who would pay, and what would they want? The cost-share rule gives agencies a financial reason to cut recorded errors³. If the tool were tuned mainly to prevent overpayments, which protect the budget, rather than underpayments, which protect claimants, it could reduce error rates while harming applicants. Which way it was tuned would need to be visible and audited.

Would something simpler do the job? If most errors came from rule complexity, simpler rules or a fixed decision tree might remove part of the problem. Where turnover is falling, as in English children's social work⁵, the case rests more on vacancies and training time than on turnover.

The anchor tool borrowed its value from top performers' conversations¹. Here the value would be borrowed from the rulebook and the people who maintain it, and that upkeep could cost more than the model.

Signs to watch

• Nava's full report and its 2026 Public Benefit Report, listed for launch on 29 October 2026⁶.
• The start of Medicaid work requirements on 1 January 2027 in most states, and how agencies staff the change⁴.
• The first SNAP cost-share calculations from FY2025 error rates, in most cases from 1 October 2027³.
• Application of AI Act Annex III high-risk rules from 2 December 2027, and any guidance on the preparatory-task derogation⁷.
• The next Department for Education workforce release, due February 2027⁵.
• A published trial of live-call guidance for public caseworkers.