Drafting a single home study can take a caseworker two to three hours. In one Minnesota county, AI now produces a usable first draft in 30 minutes. Results like that are showing up across the country as agencies adopt AI: it gives caseworkers back the hours they lose to documentation. Before any procurement starts, though, agencies face a more basic question. Which kind of AI do they actually need? Two different tools share the same label. Administrative AI handles documentation. Predictive AI generates risk scores and recommendations. The two call for different oversight, and the choice between them shapes everything that follows.
The administrative burden is concrete. A caseworker wraps up a home visit at 4:15 p.m. She took good notes. She knows what she saw and what the family needs. Turning those notes into the documentation her system requires will take the rest of the afternoon: three case updates, a safety plan, all formatted for different parts of the system. By the time she finishes, she has lost the window to prepare for tomorrow’s family team meeting.
That lost afternoon is what administrative AI is built to solve. It handles the documentation so caseworkers can spend more of their time on the work that requires human judgment. Decisions about children and families stay with the worker.
What’s the Difference Between Administrative AI and Predictive AI?
Administrative AI handles documentation. Predictive AI generates risk scores and recommendations. The two call for different oversight, and conflating them is what makes “AI in child welfare” feel risky to the agency leaders weighing it.
What is administrative AI?
Administrative AI handles tasks like documentation, transcription (the tools that draft from a recording are often called an AI “scribe”), form drafting, translation, and case file search. It supports the existing work without generating predictions or recommendations about families. A 2025 report from the UK government’s National Workload Action Group examined transcription and documentation tools being tested in social care settings to reduce administrative workload.
These tools work on a specific task, and a human reviews the output. The AI does not make a decision about the family. It drafted a form, and the caseworker decides whether the draft is accurate.
What is predictive AI?
Predictive AI generates risk scores, predicts outcomes, or recommends interventions for a case. Federal policy is now encouraging this category. The November 2025 executive order on child welfare directs agencies to expand the use of predictive analytics and AI tools to improve caregiver recruitment, retention, and the matching of children with families.
When a predictive tool surfaces a risk score or a suggested next step, that output should arrive in draft form for a caseworker or supervisor to weigh, accept, or override. The human-in-the-loop standard that governs administrative AI applies here too, with more riding on it.
Because predictive output can shape consequential decisions about a family, this category carries a higher bar before deployment: independent evaluation of how the tool performs across the populations an agency serves, plain documentation of what it does, and community input into whether and how it is used. Splitting “AI in child welfare” into these two categories, each with oversight matched to its stakes, makes every procurement conversation that follows more productive.
What Are Government Agencies Doing with AI in Social Services?
The clearest results come from agencies that picked one specific administrative task and applied AI to it.
San Bernardino County, California, the fifth-largest county in the state, uses Binti AI for transcription during family interviews and home visits. The tool drafts case documentation from the recording, and the caseworker reviews and approves every record before it enters the file. A social worker in the county’s resource family approval unit reports saving about two hours of report writing after each interview. In her words, “It saves about two hours writing reports after interviews.” That time goes back into direct work with families.
Stearns County, Minnesota uses generative AI to draft outlines for child welfare safety plans. The county enters no sensitive case data; the tool creates an outline that workers complete with case-specific details. What used to take two hours now takes 30 minutes, returning 90 minutes per safety plan to the conversations and family contact that move a case forward. The county’s published AI policy spells out which tools and uses are allowed and which are not, a useful model for other agencies considering the same step.
The pattern is the same in both. AI handles one specific administrative task. A human reviews and approves the output. The value shows up as hours redirected from paperwork to direct work with families.
Why Does the Workforce Crisis Make AI Adoption Urgent?
Administrative AI is urgent because the workforce is breaking. Caseworker turnover and burnout have crossed thresholds where the work itself gets harder every year, and documentation time is one of the few levers an agency can pull this quarter.
Casey Family Programs considers annual turnover at or below 12% optimal for human services and estimates that replacing a single caseworker costs approximately 70% to 200% of their annual salary. Federal workforce data from 2021–2022 found caseworkers spending an average of 4.3 hours a day on paperwork and documentation. NASW projects a nationwide deficit of 74,000 social workers by 2035. Burnout among mental health professionals, including social workers, has been measured at rates from 21% to 67% across studies, with substantial variation by population and setting.
The connection to administrative AI is direct. Every hour AI shaves off documentation is an hour returned to home visits, family team meetings, or court preparation. Whether a caseworker works from her own notes or, when a family consents, from a recording of the visit, she can review a structured draft at her desk instead of rebuilding the details from memory hours later. The grandmother’s exact words about how she will support the children make it into the record. The father’s clarifications about the service plan are documented while they are still fresh.
Administrative AI gives caseworkers more hours for the relationship-based work with families that drew them to the field. That matters when an agency is losing people faster than it can hire them. Reducing administrative burden does double duty: it returns time to direct casework and gives workers a reason to stay.
What Federal AI Governance Applies to Social Services Agencies?
Federal AI policy has shifted since 2023, and the direction now favors adoption. Several risk-management frameworks agencies built procurement plans around have been rescinded, and a November 2025 executive order directs agencies to expand AI and predictive analytics in child welfare. A baseline still applies: OMB Memorandum M-25-21 sets risk-management requirements for federal agency AI use, and the NIST AI Risk Management Framework remains the active voluntary standard that other federal documents reference.
What changed in federal AI policy?
Executive Order 14110, issued under the Biden administration, directed agencies to address AI risks to civil rights, civil liberties, and equity. That order is no longer the active federal baseline. OMB memorandum M-24-10, which required agencies to meet minimum AI risk management practices or stop using non-compliant AI, was rescinded by OMB M-25-21.
The federal posture then moved further toward adoption. The November 2025 executive order, Fostering the Future for American Children and Families, directs the Department of Health and Human Services to modernize child welfare information systems and expand states’ use of technological solutions, including predictive analytics and AI tools. The federal signal to agencies is now one of encouragement.
That makes the agency’s own protections more important, not less. Notice to families, appeal rights, human oversight, and equity review continue to serve the children and families involved regardless of which executive order is in force, because the risks those protections were built around did not change. The federal floor moved. The risks did not.
What’s still active?
OMB M-25-21 still requires agencies to stop using AI where they cannot properly manage the risk. The standard is proportional: higher-risk uses require stronger safeguards. For child welfare agencies, that distinction matters. The NIST AI Risk Management Framework remains the main voluntary technical standard that other federal documents reference.
States are still working out their own rules, and the picture is unsettled. Colorado passed the first comprehensive state AI law in 2024, then repealed and replaced it in 2026 with a narrower notice-and-transparency framework before the original requirements took effect. Agencies should expect state requirements to keep shifting.
If your agency built protections around the earlier federal frameworks, keep them. Children who experience foster care, birth parents, and resource families are no less vulnerable because an executive order changed.
How Should Your Agency Adopt AI Responsibly?
Governance should match the kind of AI an agency is deploying. Administrative tools that draft documents need lighter oversight than predictive tools that influence decisions about a family. Four practices scale with that risk and separate deployments that hold up to public scrutiny from those that don’t.
Establish governance before procurement. The NIST AI Risk Management Framework breaks risk management into four functions: Govern, Map, Measure, Manage. The Govern function, which covers policies, accountability, and workforce training, underpins all the others. Putting governance in place first means an agency avoids discovering gaps after a tool is already deployed, when course-correcting is more expensive and more public.
Ensure human oversight is substantive, not nominal. The Government Accountability Office (GAO) AI Accountability Framework calls for procedures that maintain human oversight and intervention throughout automated processes. In practice that means caseworker judgment supersedes any algorithmic output, override rates are audited regularly, and the workflow makes the override the natural action, not the friction. This matters most for predictive tools, where the output can influence a decision about a family.
Publish plain-language transparency documentation. Families and caseworkers both need to understand what AI tools do, what data they use, and what they do not do. Making this information accessible on the agency website, and available to any family who asks, builds the public trust that sustains adoption beyond the first administration.
Engage communities. That means convening structured listening sessions with families, specifically families of color and families with disabilities, before finalizing procurement decisions, and documenting how their input shaped those decisions. The documentation is both an ethical requirement and a practical record when a deployment is questioned later.
Together, these practices give an agency the documented record it needs when AI adoption is questioned by an oversight committee, the press, or the families themselves.
These principles also describe how Binti builds AI for child welfare. Binti AI does not make decisions or recommendations about children, families, or cases. It drafts case documentation from recordings and case notes for caseworker review and edit, and it lets caregivers complete forms in their preferred language, with responses translated back to English for workers. Every output appears in draft mode, awaiting human approval before it enters the case file. Binti is HIPAA-compliant and SOC 2 Type II compliant, and Binti AI does not store or train on customer data.
What This Means for Your Agency
Administrative AI is producing measurable results in agencies that pick specific tasks, maintain human oversight, and build governance before procurement. Recall the caseworker from the start of this article: at 4:15, still holding the details of a home visit in her head. With administrative AI, she files the update in 30 minutes instead of two hours, and uses the time she saved to prepare for tomorrow’s family team meeting.
Binti brings this same approach to the rest of the platform: Caregiver Licensing to increase available families for placement, and Family Finding & Engagement to expand kinship connections, so agencies gain time back without losing the human judgment that keeps children safe.
FAQs About AI in Government Social Services
Where is the safest place for an agency to start with AI in child welfare?
Administrative AI for specific tasks like transcription, documentation drafting, translation, and case file search. These tools handle administrative work without generating predictions or recommendations about families, and they are simpler to govern than predictive tools.
What’s the difference between administrative AI and predictive AI?
Administrative AI drafts paperwork, transcribes meetings, translates forms, and searches case files. Predictive AI generates risk scores, outcome predictions, or intervention recommendations. Because predictive output can shape decisions about a family, it calls for a higher bar: recommendations delivered in draft mode for human review, independent evaluation of the tool, and community input before deployment.
Are agencies already using AI in social services?
Yes. In San Bernardino County, California, social workers use Binti AI to transcribe family interviews, saving about two hours of report writing each. Stearns County, Minnesota cut safety-plan drafting from two hours to 30 minutes using generative AI with no sensitive case data entered. Both focus on a specific administrative task, with human review and approval of every output.
What governance practices matter most before procurement?
Establish governance before procurement, keep human oversight real and documented, publish plain-language transparency materials, and involve affected communities in procurement decisions. Match the level of scrutiny to the kind of AI: predictive tools that influence decisions warrant more review than administrative tools that draft documents.
How should agencies think about federal AI policy changes?
Federal policy now leans toward encouraging AI adoption. Several earlier frameworks were rescinded, and a November 2025 executive order directs agencies to expand AI in child welfare. The reasons agencies built protections in the first place, including notice to families, human oversight, and appeal rights, have not changed. Keep them regardless of which administration is in office.
