AI that answers health insurance member eligibility and benefits questions can safely handle the factual lookups: whether coverage is active, effective dates, deductible and out-of-pocket progress, copays for a type of service, network status and claim status. It is safe only when it verifies the member first, reads every answer from the plan's own systems, and discloses no more than the question needs. Coverage determinations, appeals and anything clinical go to a person, with the full conversation attached.

This guide sets out seven rules for building that, grounded in HHS guidance on HIPAA and CMS rules for payers, plus the tools that publish support for it.

Key takeaways

• Health plans are HIPAA covered entities. HHS lists health, dental, vision and prescription drug insurers, HMOs, Medicare, Medicaid and Medicare supplement insurers as health plans, so a vendor whose AI touches member data needs a Business Associate Agreement.

• A member ID finds the account; it does not prove who is asking. Verify before anything member-specific is said, and make the check stronger as the action gets riskier.

• Scope the AI to the minimum necessary. HHS requires covered entities to take reasonable steps to limit use and disclosure of PHI to the minimum necessary, so the AI's tools should return accumulators and copays, not whole claim histories.

• Answer from systems of record, never from the model. Eligibility is one of the HIPAA standard transactions; the AI should read the same data your staff do.

• Escalate decisions, not lookups. From 2026, CMS requires impacted payers to give a specific reason for denied prior authorizations, which is a conversation for trained staff, not a chatbot.

Which tools answer eligibility and benefits questions?

Lorikeet lists benefits verification among the patient and member workflows it resolves across chat, email, voice and SMS, and its Trust page states it is HIPAA compliant, signs Business Associate Agreements and handles protected health information on a minimum-necessary basis. Three alternatives publish related healthcare work on their own pages:

• Sierra lists eligibility verification, proactive benefits guidance, claim status and EOBs on its healthcare page.

• Salesforce Health Cloud says Agentforce for Healthcare includes a benefits verification skill, and that its contact center tools surface eligibility, benefits and claims status (Salesforce).

• Infinitus describes an agent taking an inbound health plan call about a large hospital bill and opening a billing review (Infinitus).

How we chose and checked these

The seven rules come from the regulators that set the boundaries: HHS for HIPAA privacy and security, CMS for payer transactions and prior authorization, and NIST for authentication strength. Every regulatory statement links to the page we read on 23 Oct 2026. Vendors appear if their own public pages describe member eligibility, benefits or health plan support; each vendor claim links to that vendor's page. Lorikeet publishes this guide and is one of the tools listed. We have tried to describe every product, including our own, using only what each vendor publishes, and we link to the source for each claim.

1. Decide which questions are safe to automate

Automate questions with a factual answer in a system of record, and route anything that needs judgment. A practical split:

The general-definition questions are the easiest win. CMS describes the Summary of Benefits and Coverage as a concise plain-language document that insurers and group health plans provide, alongside a uniform glossary of terms such as "deductible" and "co-payment". An AI can explain those terms with no member data at all. Member-specific answers come next, and only after Rule 3.

2. Sign a Business Associate Agreement before member data flows

If the AI vendor creates, receives, maintains or transmits PHI for a health plan, it is a business associate and needs a BAA. HHS's Privacy Rule summary names health plans as covered entities, and its business associate guidance says the BAA must describe the permitted and required uses and disclosures of PHI and bar the business associate from using it otherwise.

The same page says a business associate must have a BAA with its own subcontractors before disclosing PHI to them, and those subcontractors are business associates too. For AI, that chain usually includes model providers and hosting. Ask each vendor which subcontractors see member data and on what terms. Ask whether models are trained on your data. The Security Rule's safeguards also apply to business associates, according to HHS's Security Rule summary, including audit controls that record and examine activity in systems that hold ePHI.

3. Verify the member before saying anything member-specific

A member ID, date of birth or phone number identifies an account; on its own it does not prove the caller is the member. Match the check to what the answer reveals: a general plan question needs no verification, a read-only question about one member's coverage needs at least one factor, and anything that changes contact details or reveals sensitive information should need two factors of different types.

Two design choices matter more than the exact factors. First, the AI should collect answers and pass them to your own verification service, which returns pass or fail; the AI should never hold the correct values, because a model that knows them can be talked into revealing them. Second, sensitive tools should run only with a short-lived token your systems issue after a successful check. For "have" factors, note that NIST's SP 800-63B treats one-time codes sent over the phone network, such as SMS, as a restricted option.

4. Apply the minimum necessary standard to what the AI can see

Give the AI access to the fields each answer needs and nothing more. HHS's minimum necessary guidance says covered entities must take reasonable steps to limit the use or disclosure of, and requests for, PHI to the minimum necessary for the purpose. It also says the standard does not apply to disclosures to the individual who is the subject of the information.

That exemption is narrower than it looks. Telling a verified member their own deductible is a disclosure to the individual, but the AI system reading the plan's data is a use, and HHS says policies must identify who needs access and the categories of PHI they need. In practice: an eligibility tool returns status, dates and accumulators; a benefits tool returns cost sharing for the service asked about; neither returns diagnosis codes or full claim lines unless the question requires them. Redact identifiers in transcripts that analysts and QA reviewers read.

5. Handle dependents, representatives and confidential communications

Treat a subscriber asking about a spouse or adult child as a separate person, not an extension of the subscriber's account. HHS's personal representative guidance says covered entities must treat a personal representative as the individual, with authority that derives from their authority under applicable law to make health care decisions; a health care power of attorney is one example. Your verification service should know who holds that status.

The Privacy Rule summary adds two rules the AI must respect. A covered entity may rely on informal permission to share PHI directly relevant to a family member's involvement in care or payment. And health plans must accommodate reasonable requests for confidential communications if the member indicates disclosure could endanger them, without questioning that statement. Before the AI discusses a dependent's claims with a subscriber, it should check for a confidential communications flag, and hand off if there is one.

6. Answer from systems of record and show the source

Every number the AI gives should come from a live call to your eligibility, benefits or claims system, never from the model's general knowledge. CMS lists eligibility, claim status, coordination of benefits and referrals and authorizations among the HIPAA standard transactions, so the data your staff and providers rely on already exists in structured form.

More of it is becoming available through APIs. CMS's Interoperability and Prior Authorization final rule (CMS-0057-F) applies to Medicare Advantage organizations, state Medicaid and CHIP programs, Medicaid managed care plans, CHIP managed care entities and QHP issuers on the federally facilitated exchanges. It adds prior authorization information to the Patient Access API, with API requirements generally due from January 1, 2027. When the AI answers, it should say which plan year and which document or system the figure came from, and offer the plan document when the member wants detail.

7. Escalate on clear triggers, with the full conversation

Write the handoff rules before launch: a failed or partial verification, any denial or appeal question, a request about medical necessity, any symptom or clinical question, a confidential communications flag, a complaint, or a member who is still confused after two answers. The person who takes over should see the conversation, what was verified and what the AI already looked up.

Prior authorization questions deserve their own lane. Under CMS-0057-F, impacted payers other than QHP issuers on the exchanges must send decisions within 72 hours for expedited requests and seven calendar days for standard ones, and from 2026 must give a specific reason for denials. An AI can report that a request is pending or decided; explaining a denial or what to submit next belongs with trained staff. Lorikeet's healthcare page describes a message check that hands a ticket to your team immediately when it detects a symptom, side effect or other red flag you have described.

Tools that help

These vendors publish health plan or member support on their own pages. None of them replaces your eligibility system; they sit in front of it.

A fair limit on Lorikeet: it works through your systems via APIs, so its answers are only as current and complete as the eligibility and claims data it can call, and it does not hold PCI DSS, so premium card payments should stay in your existing payment flow. Start ($2,100 a month) and Scale ($5,100 a month) are paid annually with a per-resolution rate, and its pricing page lists a DPA or BAA as standard on both, with US data residency.

Questions to ask before you pick

• Will you sign our BAA, and which subcontractors, including model providers, will see member PHI?

• Can the AI call our verification service and act only on a token it issues, rather than checking answers itself?

• Can each tool be limited to specific fields, so the AI never receives data the question does not need?

• How are dependents, personal representatives and confidential communications flags handled?

• Can we see, for any conversation, what was verified, which systems were called and what was said?

• Can we test escalation rules against simulated members before launch?

For a deeper checklist, see our guides to BAAs for AI support, what to verify on SOC 2 and HIPAA claims and voice AI account verification, plus our list of AI support platforms that sign BAAs.

What still needs a human

• Coverage determinations and appeals. The AI can report status; it should not decide or predict outcomes.

• Clinical questions. Whether a service is right for a member is a clinical matter. The AI hands off, every time.

• Authority questions. New power of attorney documents, deceased members and disputed access go to trained staff.

• Privacy incidents. If the AI may have disclosed PHI to the wrong person, your privacy officer decides what happens next, not the support queue.

• Compliance sign-off. Your compliance team approves the automated question list, the verification rules and the escalation triggers before launch, and reviews them when plan rules change.

A start that works: triage first. Health start-up easykind began by having Lorikeet read and tag every incoming patient email and route it to the right coordinator, and its customer story reports email response times falling 92%, from 24 hours to roughly two. To try it on your own member questions, start a 30-day free trial.