Administrative Burden and Algorithmic Equity: A Randomized Trial of AI versus Human Assistance in Public Benefits Enrollment

Both arms deliver the same service (eligibility screening, application completion, and submission for SNAP, Lifeline, and WIC through Link Health's standard pre-submission human review) through Link Health's remote benefits-assistance channel. Both arms are SMS conversations entered by a tap-to-text link carrying the assignment code. The AI arm is a conversation with LEO, Link Health's AI enrollment assistant, which identifies itself as automated, is continuously available in English and Spanish, and hands off to a human navigator on request, on fallback thresholds, and on distress flags. The human arm is a conversation with a trained Link Health remote navigator, staffed in shifts with response fulfillment targeted within 48 hours. The contrast is the modality of assistance as each is deployed, service levels included, with the channel held constant. Full description in the Analysis Plan §3.

2026-12-01

2028-03-14

Verified approval of at least one benefit application submitted through Link Health (SNAP or Lifeline) within 60 days of randomization; the sole primary endpoint, tested without multiplicity adjustment. See Analysis Plan §6.

A confirmatory family of three, adjusted by Romano–Wolf stepdown: application submission within 90 days; completion of eligibility screening; exit at the opening block. Four further families, all exploratory and reported unadjusted with confidence intervals, are defined in Part B §6: exploratory enrollment outcomes (including per-program approval and submission for SNAP and Lifeline); component-capability measures (approval conditional on submission; review-stage correction rate; assistant-to-application fidelity); RQ2 process measures; and RQ3 heterogeneity.

Two-arm, individual-level randomized controlled trial delivered through Link Health's remote benefits-assistance channel. Each partner CHC texts its own patients, from its own platform, about free help enrolling in public benefits; the message carries a link coded to that health center and to the outreach language. Tapping the link is the single entry action: the study server logs the click, assigns the arm, and opens the participant's messaging app with a short text pre-filled, carrying the assignment code, addressed to the assigned arm's Link Health number; no page or form is shown and no phone number is entered. The mechanic is identical in both arms.

- **AI arm.** An SMS conversation with LEO, Link Health's AI enrollment assistant (inference on AWS Bedrock; SMS on Amazon SNS with AWS End User Messaging, in Link Health's AWS environment — a study build, since LEO's operational deployment is a web chat). LEO identifies itself as automated, delivers the research disclosure, screens for eligibility, completes applications conversationally, and submits through Link Health's standard human-review step. Participants may reach a human at any time on request; LEO also hands off on distress flags and fallback thresholds.
- **Human navigator arm.** An SMS conversation, on the same service, with a Link Health remote navigator, whose first message carries a one-line self-introduction and the same disclosure; the navigator screens, assists, and submits through the same review step.

**Service levels.** The AI arm is continuously available and responds immediately, in English and Spanish. The human arm is staffed by 5–20 trained navigators per campaign, shifts between 8am and 8pm Central, response fulfillment targeted within 48 hours of a wave — the service levels as currently staffed; the levels in force at launch are confirmed then and reported with the results. The asymmetry is part of the treatment as it exists: the estimand is the two modalities *as deployed*, service levels included, and response latency differs between arms by construction — an interpretive point for the process outcomes, not a limitation to adjust away. A participant who stalls or abandons mid-application receives, at minimum, follow-ups at approximately 48 hours and 7 days, in both arms, on an identical schedule, delivered channel-natively; the schedule is a floor — if Link Health's documented standard follow-up practice is denser, that practice becomes the floor, applied identically in both arms.

**What is and is not contrasted.** The arms differ in *who or what* provides the assistance and in nothing else the study controls: both are SMS conversations on Link Health's numbers, entered by the same tap-to-text mechanic, carrying the same four-message disclosure, with the same follow-up schedule and the same outcome contact. The estimand is AI assistance versus human assistance with the channel held constant — the modalities as deployed, service levels included. Both arms are new SMS deployments for the study (LEO's SMS delivery; the navigator pathway), which is what the QA wave exists to verify (§5).

Not available

Server-side assignment at the participant's entry tap, 1:1, by permuted blocks of randomly mixed sizes 4 and 6 within health-center × outreach-language strata, from pre-generated sequences held by the technical custodian; no human involvement in assignment; generation procedure and realized balance reviewed by Analyst Institute. See Analysis Plan §4.

Individual

No

4,300–7,000 randomized individuals, from a planned 6–10 CHCs sending roughly 90,000–150,000 messages; the registry record carries the planned range and is updated with the realized counts at closure. The denominator is everyone who clicks the study link, including those who leave immediately.

4,300–7,000 randomized individuals, from a planned 6–10 CHCs sending roughly 90,000–150,000 messages; the registry record carries the planned range and is updated with the realized counts at closure. The denominator is everyone who clicks the study link, including those who leave immediately.

1:1 allocation — approximately 2,150–3,500 randomized individuals per arm over the planned range.

Two-sided α = 0.05, 80% power, over the randomized-N range 4,300–7,000. The ITT denominator is all clickers, while the ~40% operational enrollment baseline was measured among engaged applicants in the human-navigator channel, so the effective baseline is 40% × Pr(engage | click) and MDEs are registered across engagement scenarios: | Engagement rate | Effective baseline | MDE, main effect (ITT) | |---|---|---| | 30% | 12% | 2.2–2.8 pp | | 50% | 20% | 2.7–3.4 pp | | 70% | 28% | 3.0–3.8 pp | Subgroup MDEs scale with subgroup share: half the universe, 4.2–5.4 pp; one quarter, 6.0–7.7 pp; one tenth, 9.5–12.1 pp. The on-site LEO deployment's landing-page-to-conversation rate of ~27% is stated as a **floor** on the engagement rate; at that floor the effective baseline is ~11% and the main-effect MDE is approximately 2.2–2.8 pp, the same as the 30% row. See Analysis Plan §5.