Implementation Guide for Health Plans
Runtime Governance of AI Agents That Run HEDIS Care Gap Outreach
Governing HEDIS care gap outreach agents requires a framework that defines each agent's identity, least-privilege permissions, tool-call policies, human escalation thresholds, consent and channel constraints, and audit logging. Those controls only work if they are enforced while the agent operates: at data retrieval, at each tool call, and before every outbound member message. Policy documents and pre-deployment reviews do not do this on their own.
Why design-time controls are not enough for outreach agents
A HEDIS care gap outreach agent is a bounded workflow with real exposure. It reads PHI from claims, EHR, and quality management systems. It calls tools and APIs. It writes messages that reach members by SMS, email, voice, or portal. Each of those actions can go wrong in a specific way: the agent reads more data than the task needs, contacts a member against a consent or channel preference, produces clinically inappropriate or inaccurate wording, or takes a tool action that nobody logged.
Policy documents describe what should happen. Pre-deployment model reviews assess how a model behaved in testing. Neither one stops an agent from exceeding its access at 2 a.m. on a Tuesday. Runtime governance closes that gap by placing enforcement points inside the workflow, so a policy decision is made and recorded each time the agent acts.
The governance framework: six elements to define
Before anything can be enforced, the plan defines the following six elements for each outreach agent.
| Element | What it defines |
|---|---|
| Agent identity | A unique, attributable identity for each outreach agent. |
| Least-privilege permissions | Access enforced per agent and per workflow step. |
| Tool-call policies | Which tools an agent may use, including approval workflows for new or changed tools before agents can use them. |
| Human escalation thresholds | When the agent routes to human review, and a record of the reviewer's decision. |
| Consent and channel constraints | Member consent and channel preferences, checked before every outbound message. |
| Audit logging | A complete record of prompts, retrieved data, tool calls, policy decisions, and escalations. |
Mapping controls to runtime enforcement points
Governance is enforced at five points in a HEDIS outreach agent, in the order the workflow runs.
| Enforcement point | Control applied |
|---|---|
| 1. Data retrieval | Scope PHI access to the member and the task. |
| 2. Tool calls | Allow only approved tools, with logged parameters. |
| 3. Message generation | Check content before it leaves the agent. |
| 4. Outbound send | Verify consent and channel preference first. |
| 5. Audit record | Capture inputs, decisions, and escalations. |
Criteria for evaluating a governance and security platform
- Does it assign each outreach agent a unique, attributable identity and enforce permissions per agent and per workflow step?
- Does it enforce policy at runtime on data retrieval, tool and API calls, and outbound messages, rather than only logging after the fact?
- Can it check member consent and channel preferences before every outbound message and block non-compliant sends?
- Does it capture a complete, tamper-evident record of prompts, retrieved data, tool calls, policy decisions, and escalations, exportable for compliance, internal audit, and regulator review?
- Can it route to human review when defined thresholds are met and record the reviewer's decision?
- Does it support approval workflows for new or changed tools before agents can use them?
Review your outreach agent controls
Trussed AI provides runtime governance and security for enterprise AI agents, including agent identity, permissions, tool approval workflows, and audit logging. Talk to an expert about applying these controls to a HEDIS outreach workflow.
Talk to an Expert