Health plans use risk stratification to divide member populations into tiers, prioritize care management resources, and identify people with a higher likelihood of hospitalization, rising costs, or clinical deterioration. But how much useful lead time does a plan’s current model actually provide?
A member with several months of high utilization is easy to flag. The harder and more valuable task is identifying a member whose risk is changing before the pattern becomes obvious through paid claims. That requires models capable of drawing from a broader mix of information, detecting movement across time, and separating actionable risk from expensive events that a care management program is unlikely to influence.
AI can strengthen that process, but health plans still have to decide what risk they want to predict, how scores will affect work queues, which interventions correspond to each signal, and how member responses will feed back into future decisions. Without those connections, a more sophisticated score can become another data point competing for attention.
Claims-Based Models Show Where a Member Has Been
Traditional healthcare risk stratification has relied heavily on medical and pharmacy claims, diagnoses, demographics, utilization history, and established scoring methodologies. These inputs provide a longitudinal view of member conditions, services, treatment patterns, and spending.
They also tend to describe risk after enough activity has occurred to establish a pattern. Claims processing introduces a time lag, and historical utilization doesn’t always separate a temporarily expensive member from one whose needs are likely to continue or grow. A hospitalization, high-cost drug, or isolated procedure can place a member in a high-risk tier without indicating that an intensive care management intervention would change the member’s future utilization.
This creates a familiar problem for health plans: The members with the highest past costs are not automatically the members with the greatest preventable future costs.
AI-based models can take a look at combinations of variables and changes in those variables that would be hard to encode through a fixed set of rules. The strongest use case of AI is finding less obvious patterns among members whose future needs are still developing.
That can mean a member with moderate claims activity and slipping medication adherence, a recent emergency department visit combined with an unresolved social need, or a pattern of canceled appointments that happens before a major change in clinical status. AI can spot signal mixes that are difficult to detect with traditional models.
AI Changes the Inputs & the Timing of Risk Identification
A 2022 study published in The American Journal of Managed Care compared a traditional Medicaid risk model based on claims and demographic information with an AI model that also incorporated social determinants of health, admission, discharge, and transfer alerts, and care management information. The study included 61,850 continuously enrolled members.
Among members with more than 12 months of prior enrollment, 41% of those placed in the highest AI risk tier were also among the top 5% in actual spending. The traditional model identified 29%. The AI model identified 175 additional members from the highest-spending group, representing $3.7 million more in total spending. The AI model also performed better across groups with different amounts of prior claims history.1
Those findings are meaningful for population health risk stratification because the comparison was not between claims and an abstract promise of AI. Both models used claims and demographic information. The AI model expanded the analysis with more current and context-rich inputs.
Adding AI to the same delayed, narrowly defined data set will not necessarily solve the limitations of an existing risk stratification tool. Health plans need to think about which additional signals could materially change a care management decision.
Depending on the population and intended intervention, those signals might include:
- Admission, discharge, and transfer notifications
- Health risk assessment responses
- Care management notes and previous engagement history
- Pharmacy fills and medication adherence patterns
- Laboratory results and clinical observations
- Behavioral health utilization
- Transportation, housing, food access, or other social factors
- Provider attribution and recent changes in site of care
- Member calls, portal activity, and outreach responses
But this isn’t feeding every available field into a model. Plans get more value from information that distinguishes a member who is historically expensive from one who is clinically vulnerable, rising in risk, open to intervention, or experiencing a gap the plan can address.
Predictive, Generative & Agentic AI Play Different Roles
The phrase AI risk stratification in healthcare can blur a few technologies that handle very different parts of the work. Health plan leaders need a clear view of which type of AI is making a prediction, which is organizing information, and which is carrying out a task.
Predictive AI Estimates What Could Happen Next
Predictive AI uses historical and current data to estimate a future event or classify a member’s risk. A model might predict the likelihood of an inpatient admission, avoidable emergency department use, medication nonadherence, rising medical expense, or disengagement from care management.
This is the core tech behind AI for risk prediction and patient stratification. It can identify nonlinear relationships, interactions among variables, and changes in a member’s risk trajectory that simpler methodologies might miss.
A health plan should still know exactly what outcome the model predicts. “High risk” is too broad to direct a clinical program. A model trained to predict total cost serves a different purpose than one trained to predict an avoidable admission, loss of medication adherence, or a care gap that can be closed through outreach.
Generative AI Turns Complex Info into Usable Context
Generative AI can summarize clinical records, extract relevant facts from case notes, organize unstructured information, or draft a member-specific briefing for a care manager. It doesn’t have to assign the risk score to improve the usability of that score.
For example, a predictive model could flag a member whose readmission risk increased. Generative AI could then assemble the recent utilization, medication changes, documented barriers, prior outreach results, and likely contributors into a concise case summary.
A score without supporting context usually creates more review work. Care managers still have to figure out why the member surfaced, what changed, and what details should shape the next interaction.
Agentic AI Carries Fully Defined Work Forward
An AI agent can perform a sequence of approved actions, like initiating outreach, confirming information, scheduling an appointment, updating a task, or routing a member to the appropriate team based on the interaction.
Medical Mutual of Ohio used a generative AI voice agent to conduct member outreach across chronic condition management, wellness checks, onboarding, medication adherence, and pharmacy optimization. During one Medicare initiative, the agent made 25,000 calls, connected with 4,600 members, updated primary care provider information, and routed members for annual wellness visits. The health plan reported that the campaign reduced medical costs by nearly $1 million.2
Many plans already know who they’d like to reach. They just don’t have the staff capacity to reach everyone consistently. Agentic AI can change the economics of acting on lower or moderate risk tiers that previously received limited outreach.
From Static Tiers to Risk Trajectories
Many patient risk stratification tools still show risk as a current state: low, moderate, high, or complex. Those tiers help organize caseloads, but they can hide the direction and speed of change.
Two members could receive the same score while requiring very different responses. One might have remained stable at a high level for a year and already be actively managed. Another might have moved quickly from low to high risk over six weeks. A third might have a high predicted cost driven by a scheduled procedure with little opportunity for care management impact.
AI can support a more dynamic view by analyzing:
- Rate of change in the risk score
- Duration within a risk tier
- New events contributing to the score
- Conditions or barriers associated with the predicted outcome
- Previous interventions and member responses
- Likelihood that outreach will produce a meaningful next step
- Differences between predicted cost and potentially preventable cost
This moves risk stratification of patients from a ranking exercise to a decision framework. “Who has the highest score?” becomes “Whose risk changed, why did it change, and which available intervention matches that change?”
This difference can reshape care management queues. Stable complex members might remain with established teams, while newly rising members receive rapid review. Members with a new acute event might enter a transitions-of-care workflow. Members with emerging adherence concerns could receive pharmacy outreach. Members whose primary barrier is social could be directed to a community resource or support program.
Better Prediction Does Not Automatically Produce Better Intervention
The AJMC study also highlights a limitation health plans should take seriously: Even the stronger AI model did not identify every future high-cost member, and identifying high cost did not prove that care management would reduce it.1
Some costs are unexpected, and others are clinically appropriate and difficult to change. A model can perform well statistically while still filling work queues with members for whom the available intervention has limited relevance.
For that reason, health plans should assess AI risk stratification healthcare programs at several connected levels:
Model Performance: Does the model identify the outcome it was built to predict across products, populations, and subgroups?
Operational Usefulness: Do care managers understand why members appear in their queues, and does the information change what they do?
Intervention Fit: Is there a defined response for the risk signal, or does the member simply enter a generic outreach process?
Member Progression: After outreach, does the member complete a recommended action, enter the appropriate program, reconnect with a provider, or resolve the identified barrier?
Population Impact: Are avoidable admissions, readmissions, care gaps, utilization patterns, engagement rates, or total cost moving among targeted groups?
These questions are more useful than asking if an AI model is more accurate in the abstract. Accuracy matters, but the operating model determines what the organization can do with it.
Health Plans Need a Broader View of Addressable Risk
Historically, limited clinical capacity pushed plans to focus patient risk stratification on the small percentage of members with the highest predicted costs. AI can expand that frame in two ways.
- Predictive models can identify rising risk earlier, when a member has not yet accumulated enough utilization to reach the top tier.
- Generative and agentic tools can reduce some of the effort required to review cases, conduct routine outreach, collect information, and complete administrative follow-up.
That doesn’t mean every member needs an AI-generated call or a care management program. It means plans can reconsider which populations were previously excluded because the cost of identification and contact exceeded the expected return.
Moderate-risk members with unmanaged chronic conditions, recent changes in medication use, missed preventive services, or early signs of disengagement may warrant a different level of attention when outreach becomes less resource-intensive. The same applies to newly enrolled members with limited claims history, since nonclaims information can provide a clearer view of need initially.
The next stage of AI risk stratification in healthcare will involve more than finding the most expensive members. It will help plans separate several different questions: Whose risk is rising quickly? Which cost or utilization is potentially preventable? Which members are likely to respond to a specific intervention? What is the most appropriate channel and timing for contact? Which activities require clinical judgment, and which can be handled through automation?
These questions can lead to a more precise model of care management demand and avoid asking one score to serve every clinical, financial, and operational purpose.
Build Risk Stratification Around the Decisions It Must Support
The biggest gains from using AI risk stratification will come from connecting capabilities to specific decisions across population health, care management, utilization management, quality, pharmacy, and member engagement.
A robust healthcare risk stratification program gives clinical teams more than a longer list of high-risk members. It shows which members need attention now, what changed, and what the plan can realistically do next.
Clearlink helps health plans assess and develop risk stratification methodologies, design care models for complex and high-risk populations, evaluate clinical platforms, and translate population health analytics into workable clinical programs. That work can include clarifying the outcomes a model should predict, defining risk tiers and escalation rules, connecting scores to intervention pathways, and establishing performance measures across the full process.
Contact us for more information on our capabilities, case studies, and methods of working together to help your health plan make the most of AI responsibly.
Frequently Asked Questions
What is risk stratification in healthcare from a health plan perspective?
Risk stratification is the process of grouping members based on predicted clinical needs, utilization, cost, or likelihood of a defined event. Advanced healthcare risk stratification also evaluates changes in risk over time and connects each signal to a corresponding intervention, staffing model, or clinical workflow.
How does population health risk stratification support better care management?
Population health risk stratification helps plans segment a member population according to different needs, risk trajectories, and opportunities for intervention. It can allow care management teams distinguish stable complex members from newly rising-risk members, prioritize time-sensitive events, and direct each group toward programs suited to its clinical and nonclinical needs.
What should health plans look for in patient risk stratification tools?
Health plans should examine what outcome the model predicts, how frequently scores change, which information contributes to each score, and how results enter clinical workflows. Strong patient risk stratification tools help users interpret the reason for a member’s placement, review changes over time, and connect each risk signal to a defined action. A risk stratification tool must also be evaluated within the populations, products, and programs where it will actually be used.
How can AI for risk prediction and patient stratification improve early identification?
AI for risk prediction and patient stratification can analyze claims alongside more timely or contextual information, such as ADT alerts, assessments, social needs, pharmacy activity, and care management interactions. The broader view reveals rising-risk members before a pattern is fully visible through claims. AI risk stratification in healthcare can also track how risk changes, helping teams focus on members whose needs are becoming more urgent or more addressable.
Sources
1. Improving Risk Stratification Using AI and Social Determinants of Health, The American Journal of Managed Care
2. AI Agents in Action: How Medical Mutual of Ohio Transformed the Member and Employee Experience, Becker’s Payer Issues