AI Voice Agents for Insurance: What They Can and Can't Do

I've sat through a lot of AI voice agent demos this year. The good ones are hard to tell from a person for the first thirty seconds. The bad ones still talk over the caller, mispronounce "Humana," and answer a question about Part B premiums with something that would get a licensed agent written up. The technology isn't the problem anymore. The problem is that most pitches skip the two questions that matter to anyone selling Medicare or ACA: what is this thing allowed to say, and who gave it permission to call.

This is the conversation I'd want before signing anything. Where LeadGPT fits, I'll say so and move on.

What can an AI voice agent actually do for an insurance agent?

Four jobs, all of them structured, repetitive, and free of any opinion about coverage.

Answering the inbound call you would have missed

You're at a kitchen table in Pearland walking a couple through a Part D formulary and your phone buzzes twice. That's a 64-year-old in Katy whose employer plan ends in March, calling the number on the postcard you mailed. If she gets voicemail, she calls the next agent in the stack. An AI receptionist picks up, says it's an assistant for your office, takes her name, number, and ZIP, and offers her a slot on your calendar. The Lead Response Management study from 2007 found that reaching a web lead within five minutes instead of thirty made qualifying it about 21 times more likely. That's the whole case for speed to lead.

Confirming and rescheduling appointments

Day-before confirmation calls are the least glamorous work in the office and some of the highest-return. A voice agent confirms Tuesday at 10, asks the client to have their Medicare card and drug list handy, and if Tuesday no longer works, reads back two open slots and books one. No pitch, no plan talk.

Returning the callbacks you promised

Someone leaves a message asking you to call after 6. At 6 you're still on the phone. A voice agent makes that callback, explains you're tied up, and locks in a time. Because they asked for the call, the consent picture is simple.

Capturing after-hours and overflow volume

Open enrollment weekends. The last week of AEP. The Monday after a carrier mails a rate-change letter. You can't staff for those spikes. The agent takes every call and you start the day with a sorted list instead of a full voicemail box.

What should an AI voice agent never do on an insurance call?

Here's where I get opinionated, because I've heard demos cheerfully cross every one of these lines.

• Recommend, compare, or characterize a plan. "Plan G is usually better for someone in your situation" is advice. Advice comes from a licensed agent who has done a needs analysis. The AI's job is to get that person on your calendar.
• Quote premiums, copays, or drug tiers. Even when the number is right, it's plan-specific information delivered without the disclosures CMS expects. Sometimes the number will be wrong.
• Discuss Medicare Advantage or Part D specifics before a Scope of Appointment is on file. The SOA rule has no exception for software.
• Enroll anyone in anything. Enrollment is a licensed act, recorded end to end, with the agent of record on the line.
• Pretend to be human. If someone asks "am I talking to a robot," the answer is yes, every time. California has required bot disclosure in sales contexts since 2019, and a prospect who catches the lie won't trust anything that follows.
• Collect a Medicare number or Social Security number. An appointment-setter doesn't need either, and you don't want it in a vendor's transcript.
• Keep going after a "no." A "please don't call me again" said to an AI is a revocation of consent. It needs to land on your suppression list the same hour, not after someone reads the transcript next week.

Every vendor calls these "guardrails." Ask to see them. If the guardrail is a sentence in the prompt that says "don't give plan advice," that's a hope, not a control.

Does the TCPA apply to AI-generated voices?

Yes, and the FCC removed the ambiguity. In a declaratory ruling adopted in February 2024, the Commission confirmed that calls made with AI-generated voices are "artificial or prerecorded voice" calls under the Telephone Consumer Protection Act. So the TCPA treats an AI-voice call to a cell phone like any robocall: you need the called party's prior express consent, and if the call is marketing, prior express written consent. A lead form that says "I agree to receive calls, including from an automated system or artificial voice, at the number provided" is language that holds up. A number on a purchased list is not consent at all.

In practice:

• Outbound AI calls to a purchased or aged list without documented written consent for automated or artificial-voice calls are exactly the calls the ruling makes illegal. Don't.
• Outbound AI calls to people who submitted your form with proper consent language are permissible, provided you scrub against the National Do Not Call Registry and your internal list, stay inside 8 a.m. to 9 p.m. local time (8 to 8 in Florida and Oklahoma), and identify who is calling at the start of the message with a callback number.
• Inbound calls the consumer places to you are a different posture: they initiated contact and the AI is answering the phone. Still record, still disclose, still honor any "don't call me" you hear.

State law stacks on top. Texas SB 140, effective September 1, 2025, folds marketing texts into the state's telephone solicitation rules. Keep the consent record (what they agreed to, when, where, and the exact language) on the lead's file. The FCC's announcement is on fcc.gov; the rules themselves are in 47 CFR 64.1200.

I'm a sales coach, not a lawyer. If your outbound plan involves an AI voice, pay a TCPA attorney for an hour before you pay a vendor for a month.

What do CMS rules require when AI is on a Medicare call?

CMS has no separate rulebook for AI. It has rules for marketing and sales calls, and an AI voice on a Medicare Advantage or Part D call is on one.

CMS requirement	What it means when the voice is AI
Record marketing and sales calls in their entirety	The AI's portion is part of the call. Record it, store it with the human portion, and retain it for the period in 42 CFR 422.2274, currently ten years.
TPMO disclaimer within the first minute of a sales call	Keep the AI to booking; the licensed agent reads the current CMS wording verbatim when the real conversation starts. A bot that opens a sales call has to read the disclaimer in the first minute, and you own whatever it says next.
Scope of Appointment before plan discussion	The AI can tell the caller you'll need a Scope of Appointment on file before you go over plans, and that you'll handle it. It cannot take plan questions in the meantime.
No marketing of next-year plans before October 1	An AI confirmation call in September doesn't mention next year's benefits. Not even "there are some exciting changes coming."

One more item that lands on the same desk: HIPAA. A vendor that stores recordings or transcripts containing health information is a business associate, and you need a signed BAA before the first call. Our HIPAA guide for insurance CRMs covers what to check.

How do you evaluate an AI voice vendor before it talks to a prospect?

Take this into the demo. Any answer that starts with "our model is really good at" instead of a mechanism is a dodge.

Ask	What a good answer sounds like	Walk away if
How does it disclose that it's an AI?	In the greeting, on every call, with a fixed line you can review.	"Only if they ask" or "that's configurable."
What happens when a caller asks for plan advice?	A hard stop, a scripted deflection, and an offer to book or transfer.	It "does its best to help."
Show me the transfer path to a human.	Warm transfer during hours, booked callback after hours, and you hear it work in the demo.	Transfers are "on the roadmap."
Where do recordings and transcripts live, and for how long?	On the lead record in your CRM, with retention you control.	Only in the vendor's dashboard, or purged after 30 days.
How does it handle "stop calling me"?	Marks the lead opted out from the transcript automatically and suppresses future dials.	You're expected to catch it by reading transcripts.
Does it check DNC and quiet hours before every outbound dial?	Yes, against the federal registry, your internal list, and the called party's time zone.	"That's the customer's responsibility."

One test beats the demo. Call it yourself at 11 p.m. from a number it doesn't know, and be difficult. Ask which plan is best. Ask whether it's a person. Interrupt it twice. Then read the transcript in the morning and decide whether you'd have been comfortable if that caller had been a 71-year-old widow in Sugar Land.

Which scripts can you hand an AI voice agent on day one?

These are the three I'd let run first. Change the names, keep the boundaries.

Inbound greeting

"Thanks for calling [Agency]. I'm an AI assistant for [Agent name]'s office. I can take a message, answer questions about office hours, or set up a time for [Agent name] to call you back. What can I help with?"

If asked about plans, coverage, or prices: "That's a question for [Agent name], who's licensed to walk you through it. Can I get you on the calendar? I have [slot] and [slot] open."

Before ending: "I have you as [name] at [number], and [Agent name] will call you [day and time]. Anything else I should pass along?"

Appointment confirmation, the day before

"Hi, this is the AI assistant for [Agent name]'s office. I'm calling to confirm your appointment [day] at [time]. Does that still work?"

If yes: "Great. [Agent name] asked me to remind you to have your Medicare card and a list of your current prescriptions handy. See you [day]."

If no: "No problem. I can move you to [slot] or [slot]. Which is better?"

If they ask about the plan they're considering: "I'll leave that for [Agent name] tomorrow. I'll note that you asked."

Requested callback, when the lead asked to be called

"Hi [name], this is the AI assistant for [Agent name] at [Agency]. You asked us to call about [topic]. [Agent name] is still with a client, so I'm calling to find a time that works for you. Would [slot] or [slot] work?"

If they say don't call again: "Understood. I've removed your number and you won't hear from us again."

Notice what's missing: no plan names, no dollar figures, no "we have great options this year." Every plan question gets the same answer, which is a slot on your calendar.

How do LeadGPT's AI receptionist and Sofia split the work?

LeadGPT has two AI pieces that touch the phone, and both are deliberately narrow.

The AI receptionist answers inbound calls on your local numbers when you can't. It finds out what the caller needs, captures contact details, and books against your Google Calendar. The call is recorded and transcribed onto the lead record and runs through the same call analytics your own calls do, so a manager can review it the way they'd review a rep.

Sofia works for you rather than for the caller. When a lead arrives, from the receptionist, a web form, your API, or a MeetBrokers inquiry, Sofia texts you right away. If you've allowed it for that lead, she can hold the first text conversation on your behalf, book the appointment, and send reminders, all inside quiet hours, and you can mute her on any lead. She also answers by SMS, by voice through her callback number, and in the app.

Neither piece recommends a plan, quotes a premium, or enrolls anyone. Voice campaigns cover outbound on lists where you hold documented consent, and Number Health warns you when a local number's reputation is about to cost you answered calls. Consent records travel with the lead: every MeetBrokers inquiry arrives with the consumer's coverage interest, ZIP, and the consent they gave when they asked to be contacted. The full calling stack is on the features page.

Where should you start if you're new to AI calling?

Inbound first, and only inbound, for at least a month. The downside is close to zero because every call the AI answers used to go to voicemail. Read the transcripts weekly. Once you trust it, add day-before confirmations, then requested callbacks. Only after all three are boring should you consider outbound, and only on leads whose written consent for automated calls you could hand to a judge.

That's the whole division of labor. The machine takes the parts of the phone that never needed a license: the greeting, the booking, the day-before reminder. You keep the kitchen table in Pearland.