Inside UnitedHealthcare’s Avery: How a Generative AI Companion Is Rewiring Member Experience

Avery, UnitedHealthcare’s generative AI companion, shows how large language models are shifting from demo chatbots to core infrastructure in U.S. health insurance.[1][3] Instead of diagnosis, it tackles the hard work of navigating coverage, costs, and benefits.[2]

Embedded in the UnitedHealthcare app and myuhc.com, Avery now supports ~6.5 million employer-sponsored and 160,000 Medicare Advantage members, with plans to reach 20.5 million commercial, Medicare, and Medicaid members by year-end.[3] This puts a generative LLM in front of a population comparable to a major bank assistant, changing how people access health plans.[2][3]

💡 For benefits leaders, Avery is less a “chatbot” than a signal that health-plan portals may dissolve into conversational interfaces.

---

What UnitedHealthcare’s Avery Actually Does Today

Avery is an LLM-powered companion that learns from interactions and tailors responses to each member’s benefits and demographics, while staying in an administrative—not clinical—lane.[2][3]

Core self-service workflows include:[1][2][3]

• Coverage and personal benefits questions
• Provider search and appointment scheduling
• Cost estimates, plan balances, rewards
• ID cards, claims, explanations of benefits

LLMs turn dense plan documents and rules into stepwise, natural-language guidance.[1][2]

Example: if a member asks whether a cardiologist is in-network, Avery returns network status, contact info, location, and can offer to schedule—replacing phone trees, PDF booklets, and separate search tools.[1][2] One benefits manager reported employees no longer Slack HR screenshots of plan PDFs; they paste questions into Avery and get policy-grounded answers.[2][3]

When escalation is needed, Avery passes a structured summary and suggested next steps to human advocates so members avoid repeating their story.[3] UnitedHealthcare says Avery operates under a governance framework for safety, fairness, and privacy, with humans in the loop.[3]

⚠️ Still unknown: the base model, guardrails against hallucinated coverage advice, and the precise split between benefit navigation and any clinical guidance.[3]

---

How Avery Fits into the Healthcare AI Race—and Its Risks

Avery arrives as payers and providers wire LLMs into administrative workflows.[4] Anthropic’s Claude, for example, is used for coverage checks, coding, provider verification, and prior authorization, with connectors into CMS Coverage Database and claims systems.[4][5]

Claude for Healthcare focuses on clinicians and back offices—searching PubMed, drafting prior auths, pulling EHR and trial data—while Avery is member-facing.[4][5] Early Claude results are promising: Banner Health reports 85% of users work faster, and Novo Nordisk cut a documentation process from 12 weeks to 10 minutes.[4]

Vendors are racing to meet compliance expectations. Claude runs on HIPAA-ready stacks on AWS, Google Cloud, and Azure with BAAs and limits on using health data for training; OpenAI has announced ChatGPT Health and broader healthcare offerings.[4][5] UnitedHealthcare has not detailed Avery’s HIPAA architecture, leaving its exact risk posture unclear.[3][5]

Risks include:[3][6]

• Biased or confusing benefit navigation
• Broken appeal paths after bad coverage guidance
• Environmental costs from large-model compute, complicating “call-center deflection” ROI

---

Avery signals a shift from static portals to conversational, LLM-driven health-plan navigation at national scale, while accuracy, oversight, and equity remain unproven.[1][3]

Benefits leaders, payers, and providers should treat Avery as a live case study:

• Run controlled pilots
• Track member confusion and administrative load
• Benchmark experience, safety, and compliance against Claude for Healthcare and ChatGPT Health before committing to any single stack.[4][5][6]