Who answers after the referral?
A referral sent is not a service received. What happens when the partner on the other end changes?
2023–2024 studies; the original navigator organization closed in 2024. No benefit awards or employment impacts were established.
Prepared October 10, 2026 · Independent reader review waived for this collection; no independent reader test was conducted.
The Decode (30 seconds)
A referral can look complete in a report while the person is still waiting for help. At two hospitals, 20 of 79 patients who got follow-up texts called a benefits helper within two weeks. None of the 81 patients given only a flyer did. The texts helped people make contact. The study did not show whether they received benefits. And the nonprofit that answered those calls closed in 2024. For workforce boards, the next step is not just to send a text. It is to know who will answer and how staff will check that help began.
The story (3-minute overview)
A referral can be complete in one database and unfinished in a person’s life.
Consider the last time your team sent someone to a partner for housing, food assistance, training or another service. You may know when the referral went out. Do you know who answered, whether the person could reach them, or what happened when they tried?
In Philadelphia, Penn Medicine researchers worked with Benefits Data Trust (BDT), a nonprofit with human benefits navigators available by telephone. From November 2023 to April 2024, research staff at two emergency departments screened patients for possible public benefits. Everyone enrolled received a paper flyer with the navigator phone number. Those randomized to additional follow-up received four automated text reminders over 14 days. Kilaru and colleagues, JAMA Health Forum, February 2026.
The difference in contact was striking. Within 14 days, 20 of 79 people offered the texts reached a navigator, compared with zero of 81 offered the flyer alone. 11 of 79 text-group participants submitted at least one benefits application in that period, compared with zero controls. Those are meaningful early steps. They are not evidence that anyone actually received benefits, improved earnings or found a job.
A follow-up study examined a different way to reach people. In May and June 2024, four hospitals automatically identified 1,429 patients for text outreach without the earlier study’s in-person benefit screen. One month used an opt-out message approach, the next used active choice. Across both groups, researchers recorded 121 logged navigator calls (8.5% relative to the enrolled cohort), and reported 43 patients (3.0%) submitting at least one application. The researchers also counted 78 applications across those 43 people. Do not mistake the 78 applications for 78 people, or applications for approvals. Dixon and colleagues, JAMA Network Open, June 2026.
It would be tempting to say the automated rollout did worse because it removed human screening. The studies do not support that causal claim. The second reached a different population, did not randomly compare its two messages and had no untreated control group. The original staff-screened trial and broader rollout cannot isolate the effect of the missing human step.
But there is another lesson that often gets left out of a program summary. BDT closed in August 2024, before either paper was published. Philadelphia later arranged a new BenePhilly Helpline through Clarifi and a network of benefits centers in 2025. That is documented service continuity work, not evidence that the replacement produced the same results. WHYY, 2024 · Clarifi/Philadelphia, 2025.
The closure did not cause the randomized findings, and the experiment did not cause the closure. It does remind us that an intervention can show an encouraging result in one period and still depend on a partner arrangement that no longer exists.
Go deeper (10-minute reading)
Count the steps, not just the referral
These two studies are unusually useful because they count more than outreach attempts. Their stages can be separated:
| Step in the pathway | What the research measured | What remains unknown |
|---|---|---|
| Screen or identify potential need | Researcher-assisted eligibility screening in the small trial; electronic records and rules in the later rollout | How many excluded people could have benefited |
| Deliver information | Paper flyer in both randomized groups; additional texts in one group | What each person understood or could act on |
| Contact a human navigator | 20/79 versus 0/81 by day 14 in the randomized trial; 121 logged calls, equal to 8.5% of the 1,429-person cohort, in the rollout | All possible unlogged contacts and later access |
| Submit an application | 11/79 versus 0/81 within 14 days; 43/1,429 people in the rollout | Whether submissions were complete or successful |
| Receive benefits | Not measured as a verified trial outcome | Award, start date, amount, retention of benefit and household stability |
The difference between calling and receiving help matters. A referral funnel that ends with “information provided” cannot tell a manager whether an appointment was booked or a service began. A funnel that stops at “application submitted” cannot say the household received the benefit.
The historical service was not simply an automated messaging platform. Penn had researchers screening eligible patients in the first study, a staffed hotline, reporting arrangements and a data-sharing agreement. The later rollout reduced screening effort, but navigator capacity remained essential. Neither paper supplies a complete, transferable 2026 cost per person actually helped. JAMA Health Forum trial methods.
A stronger caution than the response-rate comparison
Who was excluded? The smaller trial required English reading ability, mobile-phone access, qualifying insurance and other clinical criteria. The wider rollout also required a mobile phone, English as primary language and Medicaid or Medicare coverage. These choices limit how the evidence applies to California communities with different languages, phone access or enrollment circumstances.
What was randomized? In the first trial, the offer of follow-up texts was randomized. In the second, message type changed by calendar month, not random assignment. Differences between the two studies might reflect eligibility, recruitment, timing, service burden or other factors. They are not proof that “more automation led to less help.”
Who held the data? BDT could report calls and application totals by message group, but the larger study could not link each application back to a specific enrolled patient due to confidentiality limits. The published call figure represents logged calls and should not silently become verified unique callers. Keeping a missing outcome blank is more honest than turning it into a zero or assuming all callers completed an application.
Who was still there to answer? BDT’s 2024 closure and the 2025 launch of Clarifi’s helpline are both documented. The newer helpline was not the intervention evaluated in these papers. Before a California WDB copies the model, the first check should be the present partner’s responsibilities, language access, referral capacity, response times, and continuity plan.
California translation: repair one handoff before buying another system
Borrow now: Select one approved aggregate AJC-to-partner referral pathway. Define what counts as a successful connection from the receiving partner’s perspective. Ask how unresolved referrals are reviewed and whether someone is responsible when contact never happens.
Adapt first: The workforce board and partner should agree on the denominator, a response deadline, a safe way to reconcile missing outcomes, language options, and when to escalate a handoff. Worker consent and minimum-necessary data access need local review. Do not assume the research grants eligibility authority, approves information sharing or authorizes texting clients.
Explore further: Only after the partner has the staff and a working process should the team consider whether simple consented reminders could improve the specific missing stage. Choose a comparison, identify opt-outs and undelivered messages, count staff response time, and decide what result would make the approach worth continuing.
Our Partner Handoff Review Workbook already tracks referrals that started, closed without starting, remain pending or are unknown, and it keeps unknown outcomes in the denominator. Start there with approved aggregate counts. Do not build a second generic referral ledger. An application-to-award measure might warrant an extension only after a practitioner establishes a distinct need and tests it.
One question for the next partner meeting
If someone acts on our referral tomorrow, who will answer, and how will we know help actually started?
This question still matters if the organization changes its phone number, its contractor, or its funding source. Text reminders can help people make contact. A sustainable workforce partnership has to be ready when they do.
Sources and limitations
- Kilaru et al., JAMA Health Forum (February 6, 2026), small randomized trial conducted November 2023–April 2024; outcomes are calls and applications, not benefits awarded.
- Dixon et al., JAMA Network Open (June 17, 2026), prospective nonrandomized rollout in May–June 2024; phased groups and aggregate applications cannot identify the effect of removing in-person screening.
- University of Pennsylvania, historical BDT profile, notes closure; WHYY closure reporting (September 2024).
- Clarifi and Philadelphia (March 2025), describes replacement BenePhilly Helpline at the time of launch. No equivalence with the trial is claimed.