Routine member calls, handled safely
An AI phone assistant answered routine member questions using only the plan's own records. It could look information up but never change it.
- Client
- U.S. Medicare health plan
- What the AI does
- Answers member and provider calls
- Calls per month
- ~16,600
- Connected to
- 4 plan systems + plan documents
- Time to launch
- ~10 weeks
524 calls fully answered by AI, safely
Calls answered by the AI
Fully answered — no staff needed
Share of all calls the AI finished on its own
Callers who passed the identity check
Understood what the caller needed
Calls handed to a person when a rule required it
Routine calls, high stakes
- About 16,600 calls a month from members and providers.
- Many are the same questions again and again — coverage, claims, approvals, benefits, appeals — and the answers already sit in the plan's systems.
- The plan needed more help on the phones before the busy Annual Enrollment Period.
- It couldn't risk leaking private health information, giving medical opinions, or making short hold times longer.
Looks it up. Never guesses.
- Confirms who the caller is, using enrollment records, before sharing anything private.
- Pulls answers from the plan's enrollment, claims, approvals, and appeals systems — or plan documents for benefits.
- When one of ten clear rules applies, hands the caller to a person — with their details passed along.
- Can look things up and explain them, but can't change any records or give medical opinions.
Strict limits are what made this possible. Every answer had to come from a specific plan record or document. If data was missing, the AI wasn't sure, or a judgment call was needed, it passed the caller to a person instead of guessing.
Results cover the gradual launch and first days live, through September 2, 2026. Performance was still being tracked through enrollment season. Client name withheld for confidentiality.
Every answer comes from the plan's own records
- 01
Verify the caller
Confirms identity before sharing anything private.
- 02
Understand
Figures out what the caller is asking for.
- 03
Look it up
Pulls the answer from a plan system, official document, or the plan's website.
- 04
Answer
Explains the result — no record changes, no medical advice.
- 05
Or hand off
Ten clear rules send callers to a person, with their details passed along.
What the AI looked up for callers
Claim status checks
Pre-approval (prior authorization) checks
Coverage (eligibility) checks
Benefit and copay answers
Appeal or complaint status checks
Caller identity checks completed (HIPAA)
Typical call length
Calls passed to a person
It fully handled 41% of the calls it was built for
Calls fully handled by the AI, by type
Verified members with supported questions
All calls
Calls from providers
Provider calls ran low because there was no data yet to verify providers, so most went to a person before the AI could try to help.
Why 1,416 calls went to a person
No data to verify a provider or representative
Caller asked for a person
Caller couldn't be verified
No caller had to verify their identity again after being passed to a person.
About $2,181 saved in one early week
In one early week, with only part of the call volume going to the AI, 524 calls were fully answered without staff. Using the plan's own estimate of the full cost of a staffed call, that saved about $2,181.
An early estimate from partial call volume — not a yearly savings projection.
Live in about ten weeks
- Step 1
Legal and planning
NDA, HIPAA agreement (BAA), security review, data mapping, and hand-off rules.
- Step 2
Connect systems
Linked four plan systems, alongside the legal work.
- Step 3
Test
Updated plan info, set up reports, and ran supervised tests on real call examples.
- Step 4
Launch gradually
Grew from about 10% to 35% of the calls it could take.
Five key design choices
- 01
Data first. Map each question to the system or document that answers it before writing any AI instructions.
- 02
Look, don't touch. The AI can read records but never change them — less security and operational risk.
- 03
Verify once. Callers prove who they are one time, and it carries over if they're passed to a person.
- 04
Clear hand-off rules. Ten specific rules decide when a caller goes to a person — the AI doesn't improvise.
- 05
People over percentages. Member experience and correct hand-offs come before keeping more calls with the AI.
What it doesn't do yet — on purpose
- Verifying providers
- No provider data to check against yet, so these calls go to a person for now.
- Someone calling for a member
- Those records may be out of date, so these calls go to a person.
- Benefits and copays
- Answered from official plan documents until a benefits system is connected.
- Filing a complaint
- Handled by staff; the AI only gives status updates.
What the team learned
- In a regulated industry, mapping where every answer comes from is the project plan.
- Look-but-don't-touch is a security plus, not just a limit.
- Better measures than “calls the AI kept”: calls it handled correctly, and calls it passed on with full context.
- Test identity checks on real speech, not just clean test data.
- Plan information needs an agreed update schedule — this one updates within 24 hours of any change.
Where it goes from here
- Keep close watch through the Annual Enrollment Period.
- Get the provider data file — it affects 39% of calls.
- Get current records of who can call for a member.
- Explore a benefits system for procedure copays.
- Set up regular reports tied to relevant Star Ratings measures.
More from financial services and insurance
Renewal calls that end with a price
A broker handed its renewal calling to an AI agent that asks permission, confirms the vehicle, pulls live quotes from multiple insurers mid-call, and reads back the best offer before hanging up.
Read the case studyFinancial & Insurance
How Dottie handles regulated, high-volume phone work for carriers, brokers, health plans, and lenders.
See the overview