An AI powered insurance call center is four jobs on the dialer, not a chatbot essay glued to a phone tree. The jobs are answer, qualify, follow up, and train. The voice greets the campaign, asks the questions you saved, books a time, takes a message, ends a call that does not qualify, or warm-transfers with the transcript on the agent's screen. It does not enroll, quote, bind, or recommend a product. A licensed person still has the sales conversation. The $50 seat unlocks the profile. AI voice bills per minute from the prepaid wallet, and training role-play bills at $0.15 a minute. Name the one job this campaign needs before you turn on a second.
Facts that bound the four jobs
What an AI powered insurance call center actually runs
Map the four jobs to agent types before you name anything. Answer is a Receptionist. Qualify is a Qualifier, or a Transfer Agent when the only decision is which human queue gets the bridge. Follow up is an outbound agent you start with Trigger Outbound Call on that agent's card, inside the campaign calling window and the dialer's do-not-call suppression. Train is a Training agent that calls your own producers and scores the role-play. Custom is a fifth type for a flow you can draw step by step. Pick one type for one campaign number. A profile that tries to answer, sell, and coach in the same greeting will say something your compliance review cannot defend.
The phone path is a queue whose routing strategy is AI Agent. You select the profile before the queue will save. After the AI bridges, the human queue still applies its own strategy, required skills, accepted states, agent priority, and overflow. That second hop is where a license and a cap still matter. Nothing in the AI profile overrides a state the human queue does not accept. If you need the role-by-role choice, read which role to turn on first before you build a second queue. The label on the website is not the configuration.
Keep the marketing rules in the same picture. CMS Medicare communications and marketing guidance still governs what a Medicare voice may say, whether the speaker is a person or a profile you saved. A greeting that names a plan as the right choice, quotes a premium, or confirms an enrollment has already left the receptionist job. Final expense and life lines have the same split even when CMS is not the regulator: the AI may identify, ask, book, message, or transfer. It may not bind a policy. Write that sentence into the profile notes so a new supervisor does not 'improve' the greeting into a pitch.
Answer and qualify, then stop
Answering means the campaign number is picked up, the opening greeting you wrote is spoken, and the caller is sorted. Sorting means a short list of questions whose answers are saved as variables: state, intent, a yes or no, a number, or a selection you defined. Qualification rules then say which combination is qualified and which is not. Qualified can transfer to a queue, transfer to a phone or SIP URI, or end. Not qualified can end with a sentence you wrote. Incomplete and timeout have their own outcome, so a silent caller is not left in an open-ended conversation. The default opening greeting is a placeholder. Replace it with the campaign name, the agency name, and the recording notice you already use.
The warm transfer is the product moment. The licensed agent sees the transcript and the captured answers before they talk. They should not re-ask the state and the reason for the call. If the human queue is empty, at cap, or outside hours, booking or a message is the configured outcome, not a hold with no one behind it. Disqualification is also a complete outcome when the caller is outside the states you sell or will not answer the one field the route requires. A polite end is not a failed call. A transfer of an unqualified caller is the failure, because it spends a licensed seat on a conversation you already decided not to have.
Follow up and train are separate switches
Follow up is outbound, so it is a different legal act from answering a call the person placed. Trigger Outbound Call sits on the outbound AI agent in the dialer. It has to run inside the queue's calling window and the internal do-not-call list. Do not describe the automation builder as the thing that places these voice calls. That surface is not the switch you turn for this job. An outbound profile still does not quote, bind, or recommend. It can remind, confirm, or ask a question you wrote, and only where your consent record supports an artificial voice. The dollar figure if you get that wrong is statutory: $500 per violation under the TCPA, up to $1,500 if a court finds the violation willful or knowing.
Training never shares a campaign number with live leads. A Training agent has its own personas, competency checks, and a training dashboard. Producers take mock calls at $0.15 per minute from the wallet. The AI plays the prospect. Your agent practices the disclosure, the question order, and the moment they should stop and book rather than invent a benefit. Supervisors read scores. They do not paste the training persona onto the inbound receptionist to 'make it sound more real.' Role-play language and live-campaign language are different documents. Mix them and the next real caller hears a practice objection instead of your greeting.
What you review in the first week
Turn on one job on one number. Read every transcript the same day for the first week. You are looking for a greeting that drifted into product talk, a transfer into a state the human queue should have blocked, a qualified caller who was ended because a rule was too tight, and a follow-up that fired outside the window. Change the saved greeting and the rules. Do not tell the floor to ignore the script while the profile stays wrong. Count transfers, bookings, messages, polite ends, and disqualifications. If polite ends spike, the rules are rejecting people you meant to reach. If transfers spike and the agents re-interview everyone, the AI is bridging too early.
Pair the week with the deployment rules you already published. How to deploy AI agents safely is the guardrail list: what the voice may say, what it must not sell, and who reviews the transcript. Which role to turn on first is the next decision if this campaign is still trying to do two jobs. Owners who want a single badge will skip both and then explain a call the profile was never supposed to make. The badge is not a configuration. The four jobs are. Write the job name on the queue so the next supervisor does not add a second type just to see what happens.
How AgentTech holds the four jobs apart
AI voice agents are the profiles. AI receptionist answers and books. AI lead qualification asks the questions you saved and warm-transfers with the transcript. AI mock calls is the training job at $0.15 a minute. Each one is a type on an agent you attach to a queue with routing strategy AI Agent. The $50 seat unlocks them. AI voice on live calls bills per minute from the prepaid wallet. Recording and the transcript stay with the contact so a supervisor can review the call in the same place as a human conversation.
compliance scoring scores the recorded call after the fact. It does not grant a TCPA safe harbor and it does not make a Medicare greeting compliant. CMS Medicare communications and marketing guidance still applies to marketing speech. 47 U.S.C. § 227 still applies when an artificial voice is used to telemarket. The operating rule is narrower than the slogan: one job, one campaign, a licensed person for the sale, and a counsel review before any outbound profile is allowed to dial. If the week of transcripts is clean, then you add the next job. You do not add it on the same greeting.
Frequently Asked Questions
Does an AI powered insurance call center enroll or bind the policy?
No. The voice can answer, qualify, book, take a message, disqualify, or warm-transfer with the transcript. Enroll, quote, bind, and recommend stay with a licensed person under your existing sales process.
Can one profile answer, follow up, and train?
It should not. Receptionist, outbound, and Training are different types. Outbound starts from Trigger Outbound Call inside the calling window and do-not-call suppression. Training uses mock calls and never shares the live campaign number.
Is the setup automatically TCPA compliant because it is inbound?
No. Answering a call the person placed is not the same act as dialing them. Outbound use of an artificial or prerecorded voice for telemarketing generally needs prior express written consent. Read the FCC guide to robocalls and robotexts guide and 47 U.S.C. § 227. $500 per violation, up to $1,500 if willful. A product setting is not a legal opinion.
Key Takeaways for Agency Operators
- Four jobs, named on purpose. Answer, qualify, follow up, and train.
- One type per campaign. Do not merge the receptionist and the closer.
- The sale stays human. No enroll, quote, bind, or recommend on the AI leg.
- Outbound is a dialer trigger. Calling window and internal DNC still apply.
- Read the first week. Fix the saved profile, not a side instruction.
Next, decide which AI voice role to turn on first and keep the guardrails in how agencies deploy AI agents safely.
