Contributor: Kira Radinsky, PhD
To learn more about Kira, click here.
Source: Bigstock |
Health systems and payers have invested heavily in predictive analytics — readmission risk, care-gap likelihood, and deterioration indices. Yet in many organizations, these insights remain stranded on dashboards. Without a mechanism that translates prediction into action, clinical teams face alert fatigue, slow follow-through, and muted outcomes.
This article summarizes evidence from multi-site deployments (2023–2025) of agentic AI — systems that not only identify work but also execute it within defined safety bounds. Across programs, we observe faster gap closure, higher clinician uptake, and meaningful operating margin expansion in care-management functions.
Prediction Isn’t Performance
Model outputs — typically risk scores or propensities — often sit apart from the downstream work of outreach, scheduling, documentation, and billing. In that configuration, the burden of orchestration falls to already stretched staff. The result is “insight without infrastructure.”
Agentic AI reframes the problem. Rather than ending at prediction, the system prioritizes, plans, executes, and learns within a governed scope. That shift — from scores to closed loops — is what converts analytic lift into clinical and financial performance.
The Anatomy of an Agentic Care Workflow
The agent operates as a domain-specific service embedded in the existing stack (EHR, CRM, communications). A typical loop:
- Identify: Multimodal models analyze claims, notes, labs, and social determinants to surface at-risk patients or unmet care needs.
- Prioritize: Patients are ranked by clinical urgency, program eligibility, and expected ROI.
- Act: The system initiates interventions (reminders, follow-ups, refills, prior auth, referrals), requests consent where needed, and coordinates hand-offs.
- Document: Actions and rationales are recorded and posted to the EHR and billing systems with appropriate codes.
- Learn: Outcomes and clinician feedback update routing, thresholds, and content under version control.
Agents do not replace clinicians; they remove administrative drag and escalate when confidence or consent requirements dictate human review.
Case Study: Intake at an NCI-Designated Breast Cancer Center
At an NCI-designated breast cancer center, every new patient undergoes AI-assisted intake led by care coordinators. The agent compiles history from multiple systems, highlights potential care gaps, drafts documentation for review, and proposes individualized next steps (imaging, labs, multidisciplinary referrals). Integration occurs inside the EHR; no additional portals are required.
Operational impact (pilot → scale, 11 months; N≈all intakes):
| Outcome | Baseline | With Agentic AI |
| Intake processing time (median, minutes) | 58 | 32 (-45%) |
| Manual chart reviews per intake (count) | 3.0 | 1.9 (-37%) |
| Time to close flagged care gaps (median, days) | 8.3 | 3.2 (-61%) |
| Clinician acceptance of AI actions (%)¹ | 52-60 | 93-96 |
| Documentation completed ≤24 hours (%) | ~74 | 100 |
| Staffing change vs. volume growth | — | 0% increase |
¹Acceptance is defined as clinician-endorsed or directly executed actions/actions proposed.
These gains were achieved through co-creation with coordinators, nurses, and physicians: embedded recommendations in the EHR, structured override and feedback loops, and rapid iteration on trigger logic and phrasing. The system succeeded by aligning with clinical goals — fewer delays, fewer missed steps, and more time for patient interaction — rather than by adding a separate tool.
Data are de-identified and aggregated; “data on file,” 2023–2025.
Financial Impact: From Cost Center to Value Engine
Care management teams often run on thin margins, constrained by manual workflows and headcount-limited throughput. By automating administrative steps and accelerating completion, agentic systems convert a variable-cost function into a leveraged service.
| Metric (per 1,000 covered lives) | Traditional Ops | Post-Agentic AI |
| Patients actively managed per FTE | ~85 | ~200-250 |
| Average time to close a gap (days) | 8-12 | 2-4 |
| Clinician acceptance of actions (%) | 45-60 | 90-96 |
| Cost per closed gap | Baseline | ↓ 40–60% |
| EBITDA from care mgmt programs | Low/negative | ↑ 2.1–3.0× |
Mechanisms include: (i) fewer touches per case, (ii) higher first-pass yield on scheduling and prior auth, (iii) higher completion rates, and (iv) better targeting of high-value gaps. For risk-bearing entities (MA plans, ACOs, direct-contracting), these operational improvements translate directly into improved contribution margin and competitiveness.
Governance, Safety, and Model Stewardship
Because agentic systems act, governance must be explicit and auditable:
- Auditability: Every action records trigger, data sources, rationale, actor (human/agent), and outcome; logs are immutable and reviewable.
- PHI protections: Protected health information (PHI) is secured via encryption, least-privilege access, and de-identification during model training.
- Decision boundaries: Scope is enforced through policy rules (e.g., “agents may propose but not order diagnostic imaging”).
- Change control: Model and policy versions are tracked; promotions require safety checks and sign-off.
- Equity monitoring: Performance by demographic/clinical subgroup is monitored; drift or disparity triggers review.
- Incident response: A playbook defines detection, containment, disclosure, and remediation for safety or privacy events.
Governance is not a compliance afterthought — it is the operating system that sustains trust and enables scale.
Why This Matters Now
Three forces converge: (1) workforce constraints in nursing, coordination, and administrative roles, (2) the continued shift to value-based arrangements, where proactive engagement is financially material, and (3) mature integration rails (APIs, FHIR resources, payer connectivity) that allow agents to work inside core systems rather than beside them. In combination, these conditions make execution-first AI both feasible and necessary.
Limitations and Preconditions
Results depend on data quality and integration readiness; organizations typically complete a short “plumbing sprint” to connect source systems and define decision boundaries. Site-specific policies and state regulations can affect which steps may be automated. Finally, early gains may taper as programs saturate easy wins; ongoing A/B testing and model stewardship are needed to sustain lift.
Conclusion
Dashboards did not move margins because they did not move work. Agentic AI closes that gap by turning prediction into governed execution: identifying, prioritizing, acting, documenting, and learning — inside the tools clinicians already use. For operators, the outcomes are tangible: faster gap closure, higher clinician acceptance, and 2.1–3.0× improvements in care-management EBITDA under value-based contracts.
When AI executes within guardrails, clinicians get time back, patients see fewer misses, and care management shifts from cost center to margin engine.
Contact Kira at: [email protected]