AI inside every module. A person inside every decision.
In Medicare Advantage, the compliance officer reads this page before the CEO does. It states how the AI layer works, what it cannot do, and what is logged. No marketing language past this sentence.
Rules compute. The model rewrites. Nothing runs on a page render.
Every AI feature across the Vernier applications is built the same way. A deterministic rules implementation computes the complete answer from the member's or the plan's own records. If a language model is configured, it rewrites that answer in clearer prose, grounded in the same records. Numbers, severities, confidences and determinations are copied back from the rules result unconditionally. With no model configured, every application is fully functional and nothing is degraded.
Member portal
Claim, denial, lab, medication and approval explanations; appeal triage and letter drafting; natural-language provider search; a grounded assistant with safety interrupts and human handoff. The denial dictionary is a maintained data table in English and Spanish.
Analytics portal
Dashboard findings carry metric, observed, expected and window. MLR narrative and churn decomposition are deterministic. The model may rephrase a finding; it cannot change a number.
Admin portal
Authorization intake extraction and criteria triage, correspondence drafting with a CMS marketing pre-check, task clustering, workflow automations written from a sentence and always drafted before running, reports that show their interpretation back before the result.
Provider portal
Coding suspects with the evidence that raised them, care gaps ranked by closability, denial clustering, submission pre-flight.
Vernier HRA
A rules engine scores risk tier, red flags, care gaps and recommendations. The model drafts the narrative summary. Everything is a draft until a nurse accepts, modifies or rejects each recommendation and signs.
Credentialing
No generative AI in the decision path. Automated primary-source verification gathers; a match is flagged for the committee; the committee decides.
Guardrails, enforced in code rather than only in prompts
- Human-in-the-loop on consequential decisions. Coverage determinations, agent rejections and eligibility denials require a human action. The AI can prepare, evidence and draft. It cannot execute a denial.
- The prior-authorization endpoint never denies. A request is accepted and pended with the review action and the determination due date. A qualified person decides.
- Per-member scope isolation. Member-facing AI assembles its context from one authenticated member's records. Cross-member context is architecturally impossible in the assistant path, not just filtered.
- Safety interrupts. Certain topics in the member assistant hand off to a person. A positive suicide screen in Vernier HRA triggers an on-screen crisis protocol.
- Auditability. Every AI suggestion, the human action taken on it, and the model version that produced it. Every consent grant and revocation on the interoperability APIs. Logs are immutable and time-stamped.
- Minimum necessary. Controlled substances are masked until revealed. Bulk member exports are confirmed and logged with user, scope and timestamp.
- Business Associate Agreement before PHI. Any language-model call on real member data runs under a BAA with the model provider, on HIPAA-eligible infrastructure. Demos run on fictional data with no model key.
- No training on client data. Client data is never used to train or fine-tune models.
What is in place.
- Business Associate Agreements with every covered-entity client and with subcontractors that touch PHI.
- Encryption in transit and at rest; documents stored with AES-256 encryption at rest in production.
- Role-based access with least-privilege defaults, enforced in middleware. Dormant accounts surfaced by the data-health layer.
- Immutable audit trails in every application.
- Deployment on your infrastructure or ours, behind your identity and network controls.
- This marketing site collects no protected health information.
Security questionnaires (SIG, CAIQ or your own) go to info@halossystems.com. We answer in writing and run the review in parallel with the demo.
Bring your compliance officer.
The walkthrough shows every AI feature with the model switched off first, then on. The numbers do not change. The sentences get clearer.