Los Angeles County phones people at risk of losing their homes. A computer model decides who gets the call.

Problem

The help often starts with a letter, then a phone call that nobody asked for. A county case worker offers support with rent, bills and health care. Many people assume it is a scam, CalMatters reported in 2024. So the team tries again: it makes three call attempts to every prospective client. “Our clients didn’t raise their hand and ask to enroll in this program,” said Dana Vanderford, associate director of homelessness prevention at the county’s Department of Health Services.

Most prevention programs wait for people to apply, and people in crisis often do not know how, so many are missed. Homelessness is also rare across the whole population, even among people under strain, so it is hard to aim limited help at the people who would otherwise lose their homes.

AI solution

To find people before they lost their homes, the county turned to the California Policy Lab, a University of California research lab, which began testing machine learning on county data in 2019. The county approved and funded a pilot in 2020, and the unit started serving clients in 2021.

The lab trained the model on the records of people who later became homeless and tested it on older data before using it live. It looks at 580 factors drawn from linked, anonymized county records, such as emergency room visits, arrests, contact with probation and applications for public benefits. With them it ranks about 90,000 people who recently used the county’s health or mental health services, from highest to lowest risk. The county gets a new list four times a year.

“You sort of let the computer learn what it finds to be predictive over time,” said Janey Rountree, the lab’s executive director.

The ranking reaches people that other programs miss. In the lab’s tests, 24% of the 10,000 people ranked highest became homeless within 18 months, compared with 7% of all 90,000. In 2019, only 42 people appeared both on the model’s high-risk list of 10,672 and among the 2,522 people in the county’s other prevention programs.

The team also checked for bias. In 2019 data, the model missed 56% of Black people who went on to become homeless, compared with 63% of White people and 65% of Hispanic or Latino people. Only the gap for Black people was statistically significant, and the team chose not to adjust the model. “We know that at its worst, AI predictive analytics can be used to inflict harm on communities,” Vanderford said.

The model only picks who gets a call, and the support comes from staff. Each case manager works with about 15 clients, over four to six months according to the county, or six months on average according to the lab. Clients receive an average of $6,469 in flexible aid, which can cover rent, utility bills, car repairs or debt, and they are linked to health care, mental health services, legal aid and help finding work.

Impact

The program is deployed and county-run, and its results so far are early. The lab studied a pilot period from May 2022 to February 2023, comparing 335 people who enrolled with 1,285 eligible people who did not. Within 18 months, enrolled people were 71% less likely to use street outreach or interim housing such as shelters. In plain numbers: under 2% of the enrolled group needed those services, compared with a little over 6% of the others.

This is an observational comparison, not a randomized trial, and it comes from the same lab that built the model. People who answer the phone and sign up may differ from those who do not. The lab itself points to the limits of its method, and its randomized trial is expected to report in 2027.

Other figures come from the program. The enrollment rate rose from 21% to 35% after the unit changed how it works, including setting up a dedicated outreach team. By July 2025, the unit had served 1,498 people, and 86% kept their housing when they left the program, a self-reported figure. In 2024 alone, the county counts 712 households served and says 90% kept their housing or moved to other permanent housing at exit. By July 2025, eight in ten participants were people of color.

In 2024, the program ran on roughly $26 million in federal COVID relief money from the American Rescue Plan and was expected to end in 2026. In March 2026, CalMatters reported that the county had added funding from Measure A and was starting a prevention program for young people.

People

Dana Vanderford is the county’s associate director of homelessness prevention and co-wrote the lab’s 2025 report. The county’s Chief Information Office, where Max Stevens is chief analytics officer, is a partner, and so is the Department of Mental Health. At the California Policy Lab, Till von Wachter and Robert Santillano lead the evaluation as co-principal investigators, with Janey Rountree on the team.

Genice Brown, a case worker, told CalMatters that once people believe the call is real, about nine in ten sign up. One client put it simply: “I just really needed the help.”

“This is what smart homelessness prevention can look like, using predictive modeling and supportive services to act early, not after someone ends up homeless,” Rountree said in 2025.

Why it matters

Prevention programs are spreading across California, and CalMatters calls Los Angeles County’s approach unique because it uses AI to pick who to contact. The model has a narrow job: choosing who gets a call. The help itself is a case worker and flexible cash, offered before people lose their homes. The trial expected in 2027 should show whether the early numbers hold, and the team has been open about the limits and the risks along the way.

What we could not confirm (9)

• Whether the program causes the drop in shelter and street outreach use. The 71% figure comes from an observational comparison by the lab that built the model; the randomized trial is expected to report in 2027.
• How many people the unit serves in 2026, and how much Measure A money it now receives.
• How consent and privacy work in practice. The lab says it uses anonymized, linked county records under a data-sharing agreement, but we found no details on whether or how people are told their records are used to flag them.
• How people who are flagged but cannot be reached, or who say no, are followed up.
• Which county department runs the unit today. The 2024 lab report places it in the Department of Health Services; its official page now sits on the LA County Homeless Services & Housing website.
• How long support lasts: the county says four to six months; the lab gives a six-month average; one case worker told CalMatters in 2024 of working with each client for three or four months.
• How “people” (1,498 served by July 2025) and “households” (712 in 2024) relate, so the two totals cannot be added or compared.
• The housing-retention figures (86% and 90%) are measured when people leave the program and are self-reported; we found no longer-term follow-up.
• Whether any other county has adopted the model. Santa Clara and San Diego counties showed interest in 2024; we found no adoption.