Benefits casework: faster decisions, with a human on every single one
An agency cut its backlog by automating document intake and evidence assembly — while keeping every eligibility decision with a caseworker and auditing the system for disparate impact.
The problem
A public agency administering a benefits program had a backlog measured in months. Each case arrived as a pile of documents — pay stubs, tenancy agreements, identity documents, medical letters — that a caseworker had to read, extract, cross-check, and assemble into an eligibility determination. The work was 80% assembly and 20% judgment, and the assembly was consuming the judgment.
The constraint was as much political as technical. Automated decision-making about benefits is, correctly, subject to intense scrutiny. Any system that could be characterized as an algorithm denying someone's housing support was a non-starter.
The hypothesis
Separate assembly from adjudication. If AI reads the documents, extracts the evidence, and presents a structured, cited case file, caseworkers get their judgment time back — and the decision itself never leaves human hands.
The build
- Document intake and classification — inbound documents identified by type, with extraction models pulling the relevant fields and flagging anything illegible for human review rather than guessing.
- Evidence assembly with citations — the system builds a structured summary where every fact links to the exact page and region it came from. The caseworker verifies rather than re-reads.
- Gap and inconsistency detection — missing documents and conflicting figures are surfaced explicitly, with a drafted applicant request for the missing item.
- A hard architectural boundary — the system has no eligibility decision function. It cannot approve and it cannot deny. It assembles; the caseworker determines.
Design choice that mattered: refusing to build the decision model was the reason the project succeeded. It removed the legal and political objection entirely, and — quietly — it was also the better system. The judgment cases are exactly where human discretion belongs.
Rollout
We piloted with a volunteer group of caseworkers on a single benefit type, measuring extraction accuracy, time to assemble, and — critically — whether assembled files led to different outcomes than manual ones. They did not, which was the point. We also ran a disparate-impact analysis across protected characteristics before expanding, and committed to repeating it quarterly with published results.
Results
Case assembly time fell by more than half, the backlog moved from months toward weeks, and caseworkers reported spending their time on the cases that actually needed thought. No decision in the program is made by a model, and the agency can say so plainly.
What we'd tell you
- In high-stakes public services, the boundary between assembly and adjudication is the whole design.
- Choosing not to automate a decision is a legitimate architectural decision — often the right one.
- Run disparate-impact analysis before scaling, then on a schedule, and publish it.
- Citations to the source page are what let a caseworker move fast without moving carelessly.
More from the field.
Consumer lending, decided in hours
Early-warning credit risk
The 6 a.m. advisor brief
Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.