What problem does it solve?
A health plan member trying to understand a bill, check whether a procedure is covered, or find an in network specialist usually has to navigate several separate systems: a benefits summary, a provider directory, a claims portal and a call center, often for a question that has a single, specific answer buried in their own plan documents. Large payers field many of these routine questions, and a member who cannot get a clear answer quickly either calls an advocate for something self service could resolve, or may put off care while the cost stays unclear.
Answering well is harder than it looks: every member has a different plan, different cost sharing and a different claims history, so a generic answer about "health insurance" is often wrong for the specific person asking, and a wrong coverage answer can leave a member with a real, unexpected bill.
How does it work?
- Authenticate the member. The assistant confirms identity and pulls the member's own plan, benefits, cost sharing and claims data, never a generic description of coverage.
- Answer in plain language. Questions like "is my surgery covered, and how much will it cost" are answered with the member's specific coverage details, an in network provider list and a personalized cost estimate.
- Act. Where the payer has connected the right systems, the assistant can check a claim's status or send a digital ID card, and some assistants book care: UnitedHealthcare's Avery calls network primary care providers to schedule an appointment on a member's behalf. The member does not repeat information across systems.
- Escalate with context. When a question involves an appeal, a complaint, a denial or anything needing clinical judgment, the assistant hands off to a live advocate with a full synopsis, so the member does not start over.
- Learn and widen coverage. Each new topic area is measured for accuracy before the assistant is allowed to answer it without a human in the loop, and coverage widens by plan type, language and topic over time.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Emerging
- Channels
- Mobile app, Web chat
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Containment rate | Too few to pool | 90% | 1 | 1 organization |
Value drivers: Lower cost to serve, Customer experience, Employee productivity, Inclusion and access.
Indicative value
A health plan with 5 million members
USD 3.6 million to USD 42 million
Advocate handled contact cost avoided per year
How this is calculated
Formula: members * contactsPerMember * eligibleShare * containment * costPerContact. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Members members, members | 5,000,000 | 5,000,000 | The reference plan. |
| Assisted contacts per member per year contactsPerMember, contacts per member per year | 1 | 2 | Editorial assumption for a health plan's member service volume; replace with your own. |
| Share of contacts suited to self service (coverage, claim status, provider search, cost estimate) eligibleShare, fraction of contacts | 0.3 | 0.5 | Editorial assumption, replace with your own contact reason breakdown. |
| Share of eligible contacts the assistant resolves without an advocate containment, fraction of eligible contacts | 0.4 | 0.7 | Conservative against the UnitedHealthcare figure: UnitedHealthcare says members using Avery needed no advocate 90% of the time, a company reported figure with no stated measurement period, member cohort or topic scope. Aetna's AI powered chat is the page's only other evidence record and publishes no resolution figure, so the 90% stands alone and the low end of the range stays well below it. |
| Cost of an advocate handled contact costPerContact, USD per contact | 6 | 12 | Editorial assumption for a blended phone and chat health plan member service contact; replace with your own fully loaded cost. |
What it leaves out: Gross avoided contact cost only. It leaves out the cost of building and running the assistant, the cost of a wrong or incomplete answer, and any effect on member satisfaction, retention or appropriate care seeking.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
UnitedHealthcare
United States · Healthcare · 2026
UnitedHealthcare introduced Avery, a generative AI companion available through the UnitedHealthcare app and myuhc.com, that answers members' questions about coverage, claims, costs, providers and rewards in natural language, gives customer advocates real time insights into a member's history so a handoff needs no repetition, and can call network primary care providers to schedule appointments on a member's behalf. At launch it was available to about 6.5 million members with employer sponsored plans and 160,000 Medicare Advantage members, with a stated plan to reach 20.5 million commercial, Medicare and Medicaid members by the end of 2026.
- Containment rate: 90%, reported in the March 26, 2026 launch release; measurement period not stated
"While 90% of the time, members who use Avery have not required assistance from an advocate, Avery will automatically transfer members to advocates when a live call is needed or requested."
Claimed by: organization
Aetna
United States · Healthcare · 2025
In a July 29, 2025 press release, Aetna, a CVS Health company, said members would find a chat icon on the Aetna Health app home screen and core landing pages by the end of July. Aetna describes the chat as AI powered, giving members immediate answers and connecting to its call centers when a higher touch is needed. The same release lists claims cost visualization, personalized cost tracking, provider search and vaccine scheduling as other app features, so the chat sits inside a wider self service app. No usage or resolution figure is given for the chat itself.
No outcome disclosed.
Elevance Health
United States · Healthcare · 2025
Elevance Health expanded its Virtual Assistant, available through the Sydney Health app and Elevance Health affiliated health plan websites, to about 22 million members with commercial plans. Elevance Health describes it as using conversational technology, not as an AI system; it answers plain language questions such as whether a surgery is covered and what it will cost, returning in network providers, a personalized cost estimate and a coverage breakdown, and it integrates with the company's Find Care platform and supports Spanish. Expansion to Medicare members was planned for 2026. Elevance Health says early pilot results showed nearly 9 out of 10 pilot users were able to use the Virtual Assistant to find answers to their questions.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Member specific benefits, eligibility and cost sharing data reachable through APIs
- Claims status and explanation of benefits data
- An accurate, network specific provider directory
Systems to integrate
- Core claims and benefits administration system
- Provider directory and network status system
- Appointment scheduling or provider call center system
- CRM or case management system for advocate handover
Complexity: Medium
Answering benefit and claims questions needs live, per member integration with the core claims administration, eligibility and provider directory systems. The accuracy bar is high: a wrong coverage answer creates a real, unexpected bill for the member, so grounding every answer in the member's own plan data matters more here than in most customer service use cases.
- 1
Start with the highest volume, lowest risk questions
Coverage lookups, claim status, ID card requests and provider search before cost estimates, appeals or anything involving a denial.
- 2
Ground every answer in the member's own plan data
Never let the model generalize about "health insurance"; retrieve the member's specific plan document, benefit summary and claims record for every answer, and refuse when retrieval is uncertain.
- 3
Design the advocate handover before launch
Build the escalation path and the context transfer to a live advocate first, so a question the assistant cannot answer well never becomes a dead end.
- 4
Pilot with a defined member segment and measure per topic
Launch with one plan type or member group, track containment and accuracy per topic, and only widen once accuracy holds on a reviewed sample.
- 5
Widen by plan type and language once accuracy holds
Add plan types, benefit designs and languages one at a time, each validated against its own plan documents before going live.
Guardrails
- Answers only from the member's own verified plan, benefit and claims data, never a general guess about coverage
- Automatic handover to a live advocate on appeals, complaints, denials and anything needing clinical judgment
- Multi factor authentication before any account specific information is shared
KPIs to instrument
- Containment rate per topic, counting repeat contacts within seven days as not contained
- Accuracy of benefit and cost answers against the member's source plan documents, on a reviewed sample
- Member satisfaction on assisted conversations versus advocate handled ones
- Handover rate and handover reasons
Human in the loop
Advocates own appeals, complaints, denials and anything requiring clinical judgment. They review a sample of contained conversations for accuracy every week, and each new topic area is approved before the assistant answers it unsupervised.
Common failure modes
- Confidently wrong benefit answers
- The assistant states a coverage or cost figure that is not accurate for the member's specific plan, creating a real bill or a denied expectation. Prevent by grounding every answer in the member's own plan document and refusing rather than guessing when retrieval is uncertain.
- Containment that hides frustration
- Members give up rather than escalate, which looks like resolution in the dashboard. Count repeat contacts on the same topic and track satisfaction on contained conversations, not only the containment rate.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Article 50(1): members must be told they are talking to an AI system. Answering benefit and claims status questions is not itself an Annex III use. It would become high risk under Annex III point 5(c) if the same system assessed or priced an individual's health insurance risk, or under point 5(a) if it decided eligibility for a public health scheme; keep those decisions in the plan's existing underwriting and adjudication process, which this use case only informs the member about.
Rules that apply
Guidance
- HIPAA Privacy Rule (US Department of Health and Human Services, North America). Governs the use and disclosure of protected health information the assistant handles, including the authentication needed before releasing claims or benefit detail tied to an individual member.
Controls to put in place
- Multi factor authentication before disclosing account specific benefit, cost or claims information
- Grounding every benefit and cost answer in the member's own plan documents, with a refusal when retrieval is uncertain
- Automatic handover on appeals, complaints, denials and anything needing clinical judgment
- Logged, auditable trail of every answer and the plan document it was grounded in
- Regular accuracy sampling of benefit and cost answers against the source plan documents
Frequently asked questions
- What can a payer virtual assistant actually resolve without a human?
It varies by company and by topic. UnitedHealthcare says that 90% of the time, members who use its Avery assistant have not required assistance from a live advocate; the company gives no measurement period, cohort or topic scope for that figure, and it is not a guarantee for every question. Aetna describes an AI powered chat in its member app that gives immediate answers and connects to its call centers when a member needs more help, but publishes no resolution figure.
- Is this the same as claims adjudication, prior authorization or a policy servicing assistant?
No. A member service assistant answers questions about coverage, cost and claim status using data the plan already holds; it does not decide whether a claim or a prior authorization request is approved. That belongs to the health insurance prior authorization and claims adjudication use case. A general insurance policy servicing assistant answers from one policy; a health plan assistant adds claims status, personalized cost estimates, provider search and, in the United States, protected health information under HIPAA (in the EU, health data is special category data under the GDPR).
- Is a payer virtual assistant high risk under the EU AI Act?
Usually not. It falls under the Article 50 transparency duty: members must know they are talking to AI. It would become high risk if it assessed or priced an individual's health insurance risk, or decided eligibility for a public health scheme, rather than only answering questions about an existing plan.
- What happens when the assistant cannot answer a question?
A well built assistant refuses to guess and hands off to a live advocate with the conversation context, rather than generating a plausible sounding but unverified answer about coverage or cost. UnitedHealthcare says Avery transfers members to advocates automatically when a live call is needed or requested.
How to cite this page
Blits.ai AI Use Case Library, "AI virtual assistant for health insurance member service", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/payer-member-service-virtual-assistant. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published