The human side of human services modernization

The risk isn’t the technology—it’s failing to redesign the work around it

Consider an AI agent that uses historical risk indicators to perform a preliminary case review. In this example, the agent could identify discrepancies across case records, flag potential errors, and route its findings to a worker for resolution. But if the workflow doesn’t clearly show what the agent checked—and what it didn’t—workers may rely on an incomplete review and miss important issues. Those issues may surface later through appeals, federal reviews, or complaints, undermining trust and forcing staff to reopen work they believed was complete. Modernization intended to save time can instead create a cycle of correction and reprocessing.

When agencies invest in agentic AI without investing in the people and processes around it, they can create systems that look efficient on paper while oversight weakens, rework grows, and outcomes become harder to sustain. The dashboard may improve even as the underlying work becomes less reliable.

Agentic AI amplifies this risk because it operates at workflow scale, not task scale. A generative AI tool that produces a flawed summary may affect one document, whereas an agent with the ability to retrieve, update, and route information within case management, document, and communication systems can propagate an error across multiple steps before anyone notices. The more autonomous the system, the more important it becomes to design clear human checkpoints, escalation paths, and accountability.

The key is to deploy with discipline by understanding what agentic AI changes, what it requires from the workforce, and what agencies should put in place before scaling.

Two strategies running in parallel: Technology and workforce readiness

Sustainable modernization requires investment on two fronts simultaneously.

The first is people-enabled technology, which refers to systems designed from the start for traceability, accountability, and meaningful human oversight. As agentic AI takes on more operational activity, the challenge is to make informed oversight the path of least resistance. Workflows should prompt humans at the right moments—not as a formality after the agent has acted, but as a genuine checkpoint supported by the context that workers need to intervene effectively. It also means treating frontline input as a design requirement. Caseworkers and eligibility staff are often the first to see where an agent’s logic breaks down in real-world conditions. Their observations can help refine workflows, improve safeguards, and strengthen adoption over time.

The second is tech-enabled people, a workforce with the proficiency, discernment, and adaptability needed to turn technology investments into performance gains. In an agentic environment, this requires more than general AI literacy. Workers need to understand how agents plan and prioritize, what data they use, and when an output warrants scrutiny. Supervisors need to manage processes in which some steps are handled autonomously while others require human judgment. Leaders need to define which decisions should never be delegated to an agent and communicate those boundaries clearly.

These strategies give leaders three actions to take, which include redesigning work, enabling the workforce, and making accountability visible.

What agentic AI makes possible in HHS

AI agents differ meaningfully from many tools agencies have used before. Unlike rule-based automation, which often relies on predefined logic and structured data, AI agents can interpret unstructured information, plan next steps, and recommend or execute bounded actions under defined policies. Unlike standalone language models, agents can also use connected systems and tools to carry out those steps with defined permissions and human oversight.

For HHS organizations, four AI capabilities are relevant.

Workflow orchestration across the case life cycle

Agents can manage multistep processes from intake through determination, connecting eligibility systems, document repositories, and communication platforms without requiring workers to manually hand off every step. For example, an agent could aggregate information across sources and prompt more detailed interview questions. It could automatically trigger interface validations, update the case, and route for worker confirmation. It could also flag data discrepancies and require worker resolution before a case moves to certification. It could automatically nudge clients with incomplete or missing information via calls or text messages. Used within clear authority limits and audit trails, these capabilities have the potential to shorten processing timelines and reduce the administrative burden that pulls caseworkers away from direct client engagement.

Adaptive exception handling

Traditional automation often flags an exception and stops. Agents can interpret an exception and evaluate available options, but their authority to act is bounded by predefined policies. Therefore, routine, low-risk discrepancies within defined thresholds may be resolved autonomously and logged for review, while anything outside those thresholds is escalated to a human with a plain-language explanation of what requires attention. In quality control, this could include identifying payment discrepancies, performing automated data comparisons, recommending corrective actions, or initiating predefined and approved actions to correct errors ahead of federal review.

Real-time decision support

While workers are processing a case, agents can retrieve relevant policy, case history, and risk indicators and surface recommendations in real time—rather than requiring staff to search across multiple systems. This has the potential to support more consistent determinations and faster response times.

Automated documentation

Agents can draft case comments and documentation as workers complete activities in case management systems, reducing time spent on transcription and manual notetaking. With appropriate human review and controls, consistent documentation aligned with applicable state and federal standards can improve review readiness and reduce documentation gaps.

AI adoption and workforce transformation

Drawing on implementation experience across state health and human services organizations, the New Mexico Health Care Authority offers a useful illustration of this approach. The agency’s AI road map didn’t begin with a technology mandate. It began with an operating challenge: how to process cases faster, more accurately, and more sustainably without adding strain to a workforce managing increasingly complex workloads. The premise was simple but consequential. Technology and workforce investments would have to move together.

That premise continues to shape the agency’s choices. The road map provides a framework for prioritizing and implementing AI capabilities that sit close to the work of case processing. These include tools like AI-assisted case documentation and case analyzers that help caseworkers identify potential quality issues, while keeping staff responsible for reviewing information and making eligibility decisions. Implementation includes change management, role-based training, and an incremental rollout, allowing the tools and road map to be refined and tested based on user experience rather than being released as a one-time deployment.

This implementation design is as important as the technology itself. Stakeholder input, workforce feedback, training needs, and testing cycles are built into the process from the start. The objective is to help workers realize the full value of the investment—not only by using AI tools, but by knowing when to direct them, when to question their outputs, and how to reallocate the efficiencies they create to complex case processing and customer service.

The New Mexico agency example underscores a broader lesson for HHS leaders. The value of AI depends on whether frontline staff can use it with confidence in daily decisions. Caseworkers are often the first to encounter operational reality—where policy meets lived circumstance, exceptions emerge, and automated logic may miss important context. Our experience across the HHS domain shows a workforce that can engage critically with AI and help preserve program knowledge, strengthen adoption, and sustain performance as both the technology and operating environment evolve.

From capability to impact: Five actions for leaders

These capabilities can expand what HHS agencies are able to do, but the impact depends on the operating conditions leaders build around them. That includes a clear vision for how work will change, explicit choices about which decisions remain with humans, and deliberate preparation of the workforce expected to use and oversee the technology.

Leadership commitment is foundational to adoption. Employees look to leaders not only for direction, but also for signals about trust, accountability, and expected behavior. Agencies that communicate a coherent vision, demonstrate visible sponsorship, address concerns openly, and model responsible AI use are more likely to sustain adoption than organizations that focus primarily on deployment.

For HHS leaders, that means advancing five priorities at the same time:

1. Analyze work before redesigning jobs. Map where human judgment creates value, where AI can augment decisions, and where automation can safely execute activities within defined controls. The goal isn’t to digitize existing work, but to intentionally analyze and redesign how work is performed across people and technology.

2. Design AI around the work. Focus on solutions that reduce administrative burden, support decision-making, and fit naturally within day-to-day operations. The most effective technologies don’t replace professional judgment. They make it easier to apply consistently and at scale.

3. Build tailored AI fluency through continuous, role-based learning. The workforce has different needs. Workers need confidence using AI in daily decisions, supervisors need visibility into AI-enabled work, and leaders need the judgment to govern risk, accountability, and adoption. Because agentic AI evolves quickly, agencies should move beyond one-time training events toward agile learning journeys built around real job tasks, tailored to each role, and designed to measure whether proficiency is improving.

4. Make accountability visible. Clearly define where AI agents can act autonomously, where human review is required, and what conditions warrant escalation. As work becomes distributed between people and intelligent systems, transparency and traceability become critical to maintaining trust and program integrity.

5. Treat workforce voice as a core control. Create structured channels for frontline feedback, define how that feedback will be reviewed and acted on, and use worker insights to identify failure modes that model evaluations may miss. The goal is to protect program integrity by treating the people who interact with agents every day as an early warning system for risk, quality, and adoption.

The next era of HHS likely won’t be defined by how quickly agencies deploy agentic AI. It will be defined by whether they redesign work, build workforce capability, and create operating environments where the technology strengthens human judgment rather than compromises it.