A practical guide to augmenting clinical teams, connecting fragmented data, and scaling responsible automation.

Health plans do not need another AI experiment disconnected from the work of managing care.

They need a disciplined way to embed intelligence into the moments when nurses, care managers, and utilization management teams assemble information, assess risk and decide what should happen next. That is the promise of AI agents: software that can pursue a defined goal across data and tools, complete multiple steps, and return a result or recommended action.

In a clinical workflow, an AI agent might gather relevant member information, prepopulate an assessment, identify missing documentation, search approved benefit and community-resource directories, and present a concise set of next-best actions. The professional remains accountable; the agent reduces the friction surrounding the decision.

This distinction matters. A generative AI model produces content or insights. An agentic system combines that capability with workflow orchestration - the ability to retrieve information, apply rules, call approved tools and advance a task within defined boundaries.

The terminology is rapidly evolving, but the executive question is straightforward: What work may the system perform, what evidence must it show, and where must a human approve the outcome?

The Opportunity: Make Intelligence Part of the Workflow

Health plan clinical operations are rich in data but often poor in connected context. Clinical notes, claims, authorizations, benefits, care plans, quality gaps, pharmacy activity, and community-resource information may live in separate systems. Staff spend time searching, reconciling and documenting before they can act. An AI agent layered over governed, interoperable data can help convert those fragments into a usable member story.

This direction aligns with the broader move toward interoperable payer data. The CMS Interoperability and Prior Authorization Final Rule advances API-based exchange among payers, providers and patients, creating an important foundation for more timely, connected workflows.^[1] HHS's current AI strategy likewise emphasizes shared, reusable infrastructure; reducing information silos; workforce augmentation; and outcomes-first modernization rather than isolated AI tools. ^[2]

        Four Roles for Agents in Clinical Operations

The highest-value design is rarely a single, all-purpose agent. It is a coordinated set of narrow agents and deterministic services, each with explicit permissions, reliable source access and a defined handoff to people. A benefit-lookup agent should not make a medical-necessity determination; a clinical-summary agent should not independently initiate a high-impact intervention.

A Practical Example: A Maternity Care Workflow

Consider a pregnant member entering a care-management program. Today, a nurse may review multiple records, verify completed tests, identify complications, check available benefits and search for local support. An agent-enabled workflow could:

• assemble relevant clinical, claims, authorization and engagement history into a structured assessment;
• highlight missing or conflicting information and cite the source behind each field;
• surface risk signals - such as documented gestational diabetes or a gap in recommended follow-up -without converting a signal into an unreviewed clinical conclusion;
• identify plan benefits and approved community resources that match the member's circumstances; and
• draft a care-manager action plan and outreach message for review, then document the approved action.

The value is not simply faster data entry. It is a better-prepared human decision: more complete context, fewer missed opportunities and less time spent navigating systems. Similar patterns can support transitions of care, complex-case management, quality-gap closure, prior authorization preparation and benefit-accumulator inquiries.

The Differentiator is the Operating System, Not the Language Model

Access to a leading model will increasingly be table stakes. Sustainable differentiation will come from the capabilities around it:

• A unified, governed data foundation. Agents need timely access to normalized clinical, claims, benefit and operational data - with identity resolution, lineage and permissioning. Without that foundation, they automate fragmentation.
• Deep workflow integration. The agent must operate inside the systems and queues where work occurs, not in a separate chat window that creates another handoff.
• Health-plan-specific knowledge. Clinical policies, benefit rules, care pathways, network resources and organizational procedures must be retrievable, current and version-controlled.
• Evidence and traceability. Every material suggestion should show its sources, distinguish retrieved facts from generated language and create an auditable record of actions.
• Continuous measurement. Plans need to monitor accuracy, override rates, subgroup performance, workflow adoption, time saved and downstream clinical and operational outcomes—not just model benchmarks.

Clinical Agents Require Bounded Autonomy

The more consequential the action, the less appropriate it is to rely on an unverified generative output. NIST identifies risks including confabulation, privacy exposure, harmful bias, automation bias and opaque third-party components; its Generative AI Profile recommends managing risk across governance, mapping, measurement and ongoing management. ^[3]

Health IT policy is moving in the same direction. Federal HTI-1 requirements for predictive decision-support interventions emphasize transparency about a model's intended use, development, performance and ongoing risk management. ^[4] Peer-reviewed implementation guidance also argues that healthcare organizations must address workflow integration, governance, evaluation and implementation - not model performance alone. ^[5]

For health plans, bounded autonomy should be a design principle, not a temporary compromise. Define what the agent may read, recommend and execute. Use deterministic checks where rules can be explicit. Require human approval for clinical judgments and other high-impact actions. Make escalation easy. Log inputs, sources, outputs, tool calls and approvals. Test failure modes before deployment and monitor performance after workflows change.

A Five-Step Path From Pilot to Scale

Moving from experimentation to enterprise adoption requires more than a successful technology pilot. Health plans need a disciplined path that connects each use case to clear outcomes, appropriate oversight and reusable controls.

1. Start with a workflow, not an AI feature. Choose a high-volume process with measurable friction, accessible data and a clear accountable owner. Map the current work before introducing an agent.

2. Separate assistance from authority. Specify which steps are data retrieval, summarization, recommendation, decision and execution. Assign human approval thresholds based on risk.

3. Build the evidence layer. Connect only approved sources, enforce least-privilege access and require citations or provenance for material outputs. Resolve data quality and identity issues early.

4. Evaluate in the real workflow. Use representative cases, including edge cases and diverse populations. Compare against the existing process and track safety, quality, productivity and user trust.

5. Scale reusable controls - not one-off pilots. Standardize identity, permissions, model evaluation, monitoring, audit logs, incident response and vendor requirements so successful patterns can be reused.

The Leadership Imperative

AI agents can help health plans move from fragmented information and manual navigation toward more proactive, coordinated clinical operations. But the strongest business case is not “autonomy” for its own sake. It is the disciplined augmentation of clinical teams: giving people better context, eliminating avoidable work, and helping the organization act earlier and more consistently.

The plans that create durable advantage will do more than add a model to an existing process. They will connect their data, redesign the workflow, define the boundaries of machine action and measure whether the combined human-and-agent system produces better outcomes.

The health plan of the future will not be run by agents - but its people will increasingly be supported by them.

Questions Executives Should Ask

Before selecting a platform or launching a pilot, leaders should align on the operating problem, decision rights and evidence required for trust. The following questions can help frame that discussion:

• Which clinical workflows consume substantial time assembling information rather than applying judgment?
• What data must be unified - and what quality, consent, privacy and access controls must be resolved - before an agent can be trusted?
• Which actions may the agent take, which may it only recommend, and which always require human approval?
• How will we demonstrate source provenance, audit every action, and respond when the system is uncertain or wrong?
• What outcome will justify scale: faster cycle time, fewer touches, more completed interventions, improved quality, better experience or lower avoidable cost?

In Conclusion

Agentic AI offers health plans a practical path to more connected, proactive and efficient clinical operations -but only when it is deployed as part of a governed workflow. The objective is not to remove human judgment; it is to surround that judgment with better data, faster execution and clearer evidence. Plans that begin with focused use cases, bounded autonomy, and measurable outcomes will be best positioned to turn AI agents from promising technology into durable clinical and operational value.

References

1. Centers for Medicare & Medicaid Services. CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F).

2. U.S. Department of Health and Human Services. Artificial Intelligence Strategy (2025).

3. National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1, 2024).

4. Office of the National Coordinator for Health Information Technology. HTI-1 Final Rule: Requirements for Decision Support Interventions and Predictive Models.

5. Hua Y. et al. Generative AI in healthcare: an implementation science informed translational path on application, integration and governance. Implementation Science (2024).