Real-World Case Studies: AI in Social Services
From the course AI for Social Workers
Leading agencies and nonprofits have deployed AI to improve client outcomes and reduce caseworker burden. These documented examples illustrate what works — and what requires careful oversight.
Case Study 1: Allegheny County DHS — Predictive Risk Modeling
Allegheny County's Department of Human Services in Pennsylvania deployed the Allegheny Family Screening Tool (AFST), an AI model that scores child welfare referrals.
Approach: The model ingests administrative data from public assistance, mental health, and juvenile justice systems to generate a risk score for screeners to review alongside their own judgment.
Outcomes:
- Reduced time-to-screen from hours to minutes for routine referrals
- Enabled screeners to focus deeper attention on ambiguous cases
- Revealed racial disparities in underlying data, prompting a public audit
Key Lesson: Predictive tools can assist triage but must be paired with transparent bias audits and human override authority. Allegheny published its audit methodology — a model for the sector.
Case Study 2: Crisis Text Line — AI Triage at Scale
Crisis Text Line processes millions of texts annually. AI now flags conversations where the texter may be in immediate danger.
Approach: Natural language processing scans incoming messages for urgency signals, prioritising them in the counsellor queue rather than routing chronologically.
Outcomes:
- High-risk conversations reach a counsellor significantly faster
- Counsellor capacity is better matched to severity
- Data insights have informed national mental health research
Key Lesson: AI is most valuable in triage — matching urgency to capacity — not in replacing the human conversation itself.
Case Study 3: Benefits Data Trust — Automated Benefits Screening
Benefits Data Trust partners with agencies to identify people who qualify for benefits but are not enrolled.
Approach: AI cross-references administrative datasets to flag households likely eligible for SNAP, Medicaid, utility assistance, and other programmes. Outreach workers then make targeted contact.
Outcomes:
- Enrolled over one million people in benefits they were already eligible for
- Reduced outreach cost per enrolment substantially compared to broad campaigns
- Demonstrated that data-driven proactive outreach reaches vulnerable people traditional outreach misses
Key Lesson: AI's ability to find patterns across large datasets enables proactive service — moving from "waiting for clients to come to us" to actively reaching those most in need.
Common Threads Across All Three Cases
- Human judgment remains the final authority in every deployment
- Bias audits were built in from the beginning, not retrofitted
- Outcome measurement was tied to client welfare, not just efficiency metrics
- Transparency with clients and the public built trust in the programmes
Key Takeaways
- •Allegheny AFST shows predictive triage can work when paired with mandatory bias audits and human override
- •Crisis Text Line uses NLP to prioritise high-risk contacts — triage, not replacement of human support
- •Benefits Data Trust proves proactive AI outreach enrolls a million-plus people previously missed by traditional methods
- •All successful deployments share three features: human final authority, built-in bias review, and client-welfare metrics
Progress is kept in this browser only.