Skip to content 
By: Linda Nakagawa | With Insights From: Brian Uzwiak | Published: Jun 4, 2026 | 
9 Min
Summary:
Before adopting AI, human services agencies should ask five critical questions about domain fit, data ownership, explainability, bias testing, and human oversight. With budgets tightening and caseloads growing, thoughtful AI adoption can return meaningful staff capacity. The principle is simple: AI recommends while humans decide.
Artificial intelligence is a conversation you can’t escape.
It is in everything from conference agendas, and board meetings, to vendor pitches, and federal guidance documents. The pressure to “have an AI strategy” can feel constant, and it can also feel premature when your team is already stretched thin and the next funding cycle looks uncertain.
The good news is that the most useful questions about AI in the human services sector are not technical: they are mission questions. They are about where your caseworkers spend their hours, what gets in the way of the work that improves lives, and how a thoughtful adoption plan can free up time, attention, and budget for the people you serve.
Here is a practical look at what agencies should be weighing before they implement AI, why the timing matters now, and what a human-centered approach looks like in practice.
The Administrative Tax on Caseworkers Is Real, And It Is Measurable
Anyone who has worked in case management knows that documentation eats the day. The research backs up what staff already feel.
According to the National Association of Social Workers, significant numbers of licensed social workers report spending more than half of their time on case management responsibilities, including the documentation that comes with them (NASW). A separate study cited by NASW Press asked social workers to estimate the share of their day spent writing, and they reported 50% or more (NASW Press).
The pattern shows up across the broader human services workforce as well. Administrative spending in the U.S. health and human services system has been estimated at 15 to 30 percent of total spending, with annual administrative costs reaching roughly $1 trillion (Health Affairs Scholar). Every dollar tied up in paperwork, duplicate data entry, and reporting is a dollar that does not reach the individuals and families a program is designed to serve.
These are not abstract numbers. Every hour spent retyping notes, hunting through unstructured case histories, or reformatting the same data for a different report is an hour not spent with an individual, a family, or a funder.
Every dollar tied up in paperwork, duplicate data entry, and reporting is a dollar that does not reach the individuals and families a program is designed to serve.
Why The Timing Matters in 2026
The financial pressure on human services agencies has intensified. Federal budget proposals in 2025 raised the prospect of billions in cuts to programs that nonprofits and human services agencies rely on (Center for Nonprofit Excellence), and a national survey of nonprofits described the sector as approaching a critical point as funding for community needs falters (Nonprofit Finance Fund).
In that environment, “do more with less” is not a slogan, but the operating reality. Agencies are looking at staffing freezes, longer waitlists, and harder choices about which programs to grow and which to scale back. The case for AI in this moment is that it can give time back to staff and create the capacity to look outward, toward additional funding sources, partnerships, and the deeper relationships with individuals that move outcomes.
What AI Is Already Doing Well, According to the People Using It
When organizations are asked where AI helps most, the answer is consistent. It is administrative work.
The 2025 Stanford AI Index Report found that 78 percent of organizations reported using AI in 2024, up from 55 percent the year before, and noted that a growing body of research confirms AI boosts productivity and, in most cases, helps narrow skill gaps across the workforce (Stanford HAI). A peer-reviewed study published through the National Bureau of Economic Research, looking at more than 5,000 customer support agents using a generative AI assistant, found a 14 percent average productivity increase, with a 34 percent improvement for newer and lower-skilled workers (NBER). The most consistent finding was that AI helped less experienced staff perform closer to the level of their more tenured peers, a meaningful pattern for human services agencies facing turnover and onboarding pressure.
A benchmark study of nonprofits found that 92 percent are using AI tools in some capacity, with the most common gains showing up in faster drafts, quicker research, and improved content quality (NonProfit PRO). The same study flagged an important caution: only 7 percent reported major strategic impact, suggesting that adoption alone does not produce transformation. Thoughtful design and clear goals are what move organizations from incremental gains to meaningful change.
The lesson for human services is that the highest-leverage starting point is rarely an exotic predictive model. It is the unglamorous, time-consuming back-office work that pulls staff away from the people they serve.
The most consistent finding was that AI helped less experienced staff perform closer to the level of their more tenured peers, a meaningful pattern for human services agencies facing turnover and onboarding pressure.
The Questions Every Agency Should Be Asking
To help agencies think beyond the hype and focus on practical adoption, we turned to Brian Uzwiak, Vice President of Cloud and Data Engineering at CaseWorthy. With deep expertise in data architecture, cloud infrastructure, analytics, and AI readiness, Brian works closely with human services organizations to help them build the secure, scalable foundations needed for responsible AI adoption. Based on his experience guiding organizations through complex data and technology challenges, he recommends starting with five critical questions before implementing any AI solution.
Does the AI actually understand human services — or is it a general-purpose tool in a vertical costume?
This is probably the question I hear least often, and it’s the one that matters most.
There are a lot of general-purpose AI tools being repositioned for human services right now. They can summarize documents, answer questions, and generate text. Some of them are genuinely impressive. But impressive isn’t the same as appropriate.
Human services has a vocabulary, a set of regulatory frameworks, and a set of program models that are genuinely specialized. Does the tool understand what an HMIS Universal Data Element is? Does it know the difference between a service plan in an HCBS waiver program and a case plan in a housing program? Can it distinguish between the compliance requirements that apply to a SNAP Employment and Training program versus a victim services program?
If the answer is “we can configure it to learn your terminology,” that’s a yellow flag. Configuration can get you a long way, but it’s not the same as a system that was built with human services in mind from the start. The difference shows up in the details — in whether the AI’s suggestions actually make sense for the work, or whether they require a caseworker to mentally translate generic output into something useful.
Ask vendors to show you real examples of the tool working in a context similar to yours. If they can’t, push harder.
Who owns the data — and does it ever leave your environment?
This question might feel like IT territory, but I’d argue every program director and executive should be asking it.
A lot of AI tools — especially general-purpose ones — use the interactions they have with customers to improve their models. That’s actually fine in many contexts. But in human services, your data includes protected health information, personally identifiable information, and in many cases, records about some of the most vulnerable moments in people’s lives. The idea that this data might be used to train a vendor’s AI model — or shared with a third party for that purpose — should give anyone pause.
Read the terms of service carefully. Specifically: Does inference (the process of generating AI responses) happen inside your environment, or does data get sent to an external server? Is your client data ever used to improve the vendor’s model? What are the data retention policies? Who has the right to access it, and under what circumstances?
The standard you should hold vendors to is this: your data stays in your environment. It is not used to train models. You control what is retained and when it is deleted. An AI tool that can’t make that commitment isn’t ready for human services.
Can you show me where the answer came from?
I’ve started asking this question in almost every AI conversation I have, and the answers are very revealing.
There’s a version of AI where the system produces a summary, a recommendation, or an answer — and you have no way to know what it was based on. The output looks polished. It reads confidently. And it might be completely wrong.
In human services, that’s not a minor inconvenience. A caseworker who relies on an AI-generated case summary that misrepresents a client’s history, or acts on a recommendation that was generated without grounding in the actual record, can cause real harm. And in a field already managing significant workforce turnover and onboarding challenges, new staff are especially vulnerable to over-trusting tools they don’t yet know well.
What good looks like: every AI output should be traceable. If Cara summarizes a client’s case, it should be able to show you the specific records it drew from — the assessment date, the enrollment record, the contact note. If it answers a question about your data, it should show you the underlying numbers, not just the conclusion. That transparency is what allows a caseworker to verify, to push back, and to override when something looks wrong.
Explainability isn’t a nice feature. In this field, it’s a prerequisite.
How has the AI been tested for bias — and who did the testing?
This is the question that I think deserves the most attention and gets the least.
Algorithmic bias doesn’t announce itself. It shows up quietly — in who gets flagged as high-risk, in whose case gets prioritized, in what service options get surfaced for whom. And for populations that have historically experienced unequal treatment in healthcare, housing, criminal justice, and social services, an AI tool that perpetuates those patterns isn’t just ethically problematic. It actively contradicts the mission of the organizations using it.
Here’s what to ask: Has this AI been tested for differential outcomes across race, ethnicity, gender, age, and language? Not “we monitor for bias” — what specific testing has been done, and what did it find? Who ran the test — the vendor’s internal team, or an independent auditor? When something was identified, what changed in the model?
And critically: is this ongoing? Bias testing isn’t a box you check at launch and never revisit. Model behavior can drift. Data distributions shift. The right answer from a vendor is a commitment to regular, documented fairness assessments after deployment, not just before it.
If a vendor is vague on this question, or can’t produce documentation, that tells you something important about how seriously they take it.
What does "AI-powered" actually mean — and what do my staff still decide?
“AI-powered” is on every vendor slide. It doesn’t tell you whether the AI makes decisions or supports them.
This distinction matters enormously in human services. There is a meaningful difference between an AI that drafts a case summary for a worker to review, edit, and approve — and an AI that automatically generates a service recommendation that gets acted on without human review. Both might be described as “AI-powered.” Only one of them keeps a human in the loop.
For organizations working with vulnerable populations, the answer to this question should always be: staff remain in control. AI augments human judgment. It never replaces it.
That means every AI output should be reviewable. Every recommendation should be editable. Staff should be able to see why the AI produced what it produced, disagree with it, and override it without friction. And there should be a clear audit trail showing what the AI suggested and what the human decided.
Ask vendors directly: Can my staff override any AI output? Is there any step in your workflow where an AI takes action without human review? What happens when a caseworker disagrees with what Cara recommends?
If the answer to any of those isn’t a clear, confident yes, keep asking.
The organizations that get this right will ask these questions first
AI has real potential in human services. I genuinely believe that.
The ability to surface a complete picture of a client’s history across every program they’re enrolled in in seconds rather than minutes could give caseworkers meaningful time back. The ability to flag clients who are overdue for contact, or identify patterns in a caseload that a human would miss, could improve outcomes in ways that matter.
But that potential is only realized when the technology is built the right way, for the right audience, with the right safeguards. The organizations that will benefit most from AI are the ones that adopt it intentionally — that ask hard questions, demand real answers, and choose tools that were built to earn the trust of the people using them and the people they serve.
If you’re exploring what AI might mean for your organization, I’d encourage you to start with these five questions. The vendors who can answer them confidently are the ones worth your time.
The Principle That Should Sit Underneath All of It: AI Recommends, Humans Decide
This is the principle that should guide any AI rollout in human services, and it is worth saying clearly. AI can recommend. Humans decide.
That goes beyond clicking approve on every action. It means the human always defines the rules: what the AI can do automatically, what needs a person to review before it moves forward, and what stays a human-only call. The organization draws the boundaries, and the AI works inside them. The human stays in the seat of authority.
This matters more in human services than in almost any other sector, because the decisions that affect a client’s safety, eligibility, placement, or care plan are not the kind of decisions you delegate to a model. They are decisions that require judgment, context, relationship, and accountability. AI can take notes off a caseworker’s plate, surface the right document in seconds, and draft a first version of a report. The caseworker still owns the call.
AI recommends. Humans decide.
Where Cara Fits
At CaseWorthy, this principle is the foundation of Cara, our AI copilot built specifically for human services. Cara is designed to reduce documentation time, surface insights, and expand staff capacity so teams can spend less time on administrative tasks and more time with the individuals they serve.
The work behind Cara goes well beyond the AI itself. The governing principle is non-negotiable: Cara recommends, humans decide. Customers configure what Cara can do automatically, what requires approval, and what stays a human-only call. The boundaries belong to the agency.
A few things we think are worth knowing as you evaluate any AI tool for human services, including ours:
Customer data is never used to train AI models without explicit consent
And there is no claim to cross-customer model improvement.
AI features are optional
Users can override suggestions, customize settings, or opt out of specific capabilities entirely.
All processing happens inside the vendor’s system
In our case, that is CaseWorthy's certified infrastructure, with SOC 2 Type II, HIPAA compliance, and HITRUST certification in progress.
We need to be perfectly clear: these are not selling points. They are the minimum standard any human services agency should expect from any AI vendor it considers.
The Bigger Payoff: Capacity, Not Just Speed
The most important reason to think seriously about AI right now is not that it shaves minutes off a task. It is that when you give caseworkers and program staff their time back, you create capacity for the work that does not fit into a form field.
Time with individuals and families. Time to look at outcomes data and adjust a program. Time for a development director to chase a new grant opportunity. Time for a director of programs to think about partnerships, expansion, or the next funding cycle. Time to support staff in a field where burnout is one of the largest threats to continuity of care.
In a year where budgets are tight and demand is rising, that capacity may be the single most valuable thing technology can give a human services organization. AI will not solve the funding environment, and it will not replace the relationships at the center of this work. Used well, with humans clearly in charge, it can give your team back the hours those relationships deserve.
ON-DEMAND WEBINAR:
Beyond the Buzzwords: How Human Services Organizations Should Approach
Artificial intelligence is evolving quickly, but many human services organizations are still determining where it fits. Watch this on-demand webinar to explore practical use cases, key governance considerations, and how to approach AI without losing sight of your mission.
About the Authors
Linda Nakagawa
Linda Nakagawa is a Senior Policy Analyst on CaseWorthy’s Industry Leadership team, where she focuses on policy trends, funding landscapes, and best practices in health and human services. In this capacity, she tracks emerging federal and state guidance, engages with provider and association partners, and translates complex policy and regulatory changes into clear, actionable insights for internal teams and clients. Linda is especially interested in how technology, data, and thoughtful policy design can strengthen outcomes for communities served by aging, disability, and other human services programs. She holds a Bachelor of Arts in Psychology and Politics from Brandeis University.
Brian Uzwaik
Brian Uzwiak is Vice President of Cloud and Data Engineering at CaseWorthy, where he leads the strategy and development of the company’s cloud infrastructure, data architecture, and analytics capabilities. With deep expertise in scalable platform engineering and data systems, Brian focuses on helping human services organizations unify data, improve reporting, and build a stronger foundation for operational insight and innovation. His work supports CaseWorthy’s mission to deliver secure, connected, and purpose-built technology for human services organizations
