A new insurance lead is more useful when the adviser receives the property details and an agreed time to talk. The follow-up call needs to prepare that conversation while the enquiry is still fresh.

At Custom Voice Agent, we built an AI voice agent for a US home-insurance agency around that job: call the Meta ad lead, collect the quote information and book an available adviser callback. This historical build shows how the workflow connects and what early testing taught us.

The path from form to phone call

In the original build, the Meta lead form sent the enquiry to the agency by email. A workflow read the lead, created the contact in GoHighLevel and placed an outbound call through Retell. When the call ended, a second workflow sent the details to the CRM and a summary email to the agency.

YesNoSlots returnedLookup failsMeta lead formsubmittedLead email readby workflowContact createdin GoHighLevelOutbound callplacedGreeting, thenconfirm thecallback numberWhy are youshopping forinsurance?Household:owner or renter,spouse, otheradultsProperty: yearbuilt, roof,foundation, extrastructuresInterested inbundling auto?Vehicles anddriversClaims in thelast five yearsCheck advisercalendarOffer two orthree callbacktimesTake thepreferred timeas a staff taskBook the chosenslotTextconfirmation andend the callFields andsummary sent tothe CRM
Flow 1: The path from form to phone call. A numbered text explanation follows.
  1. The call opens with the reason for the follow-up and asks about the person's goal in their own words.
  2. The agent confirms the number it is speaking to before it collects anything else.
  3. Questions come in groups: household first, then the property, then the optional auto bundle, then claims.
  4. A callback is booked only after the calendar returns real slots.
  5. After the call, around twenty fields and a written summary go to the agency.

Why the questions are grouped

A home-insurance quote needs a lot of detail. Grouping related questions lets the caller stay with one subject at a time: household, property, then any auto bundle. The prompt also explains why a detail matters when the reason is not obvious.

Illustrative example — not a client transcript.

Agent: What year was the house built?

Caller: About 2000.

Agent: Got it. And how old is the roof? If it's as old as the house, that's fine.

Caller: Three years. It's metal.

Agent: Thanks, I'll include that for the adviser. And the foundation: crawl space, slab or basement?

Caller: Crawl space. Actually, no, it's a finished basement.

Agent: No problem, a finished basement.

The last line matters. Callers correct themselves mid-answer, and the agent has to change one field without restarting the section. Our December test calls included exactly this kind of correction.

The five tools behind the call

The agent had five tools: one to read back and validate the phone number, one to check the calendar, one to record the time the caller asked for, one to book and one to end the call. The prompt told the agent to say a time was booked only after the booking tool succeeded.

That rule exists because the calendar was the weak point of the build.

What went wrong

A booking tool pointed at the wrong calendar. In early testing, calls, lead creation and after-call emails were working, while appointment booking was failing. The cause was an agent version attached to a different calendar and webhook. The corrected version was relinked and tested before we treated it as fixed.

The calendar pause confused the agent. The availability lookup takes a moment. In one test call the agent filled the silence with an "are you still there?" even though the prompt told it never to say that. The lesson is to give the agent a short phrase such as "I'll check the calendar" and test what it says during the wait.

Early hang-ups produced half-empty summaries. If a caller left halfway through, the follow-up email still looked complete. A completeness check was added so staff could tell a finished intake from a partial one.

What the adviser receives

The adviser gets the callback slot, the answers that were given, and a summary. Fields the caller did not know stay empty rather than guessed. Useful outcomes are separate: booked, callback requested, no answer, declined and incomplete intake.

The adviser still owns everything that makes this insurance: advice, underwriting, the quote itself and the decision to bind cover.

What we would change today

  • Say it is an AI. The early version used a human first name and a human persona. A new rollout should identify the AI assistant at the start of the call.
  • Set a contact policy before the first dial. A form submission is not permission for an unlimited retry sequence. Decide how many attempts, at what times, and what a "no" does.
  • Leave unknowns unknown. If the caller does not know the roof age, the record says so.

Who this is for

This workflow suits an insurance agency with a steady flow of enquiries, an adviser calendar and staff who can take the booked callbacks. Decide which details a caller can reasonably provide by phone and leave the rest for the adviser.

Bring your lead form and adviser booking process to Custom Voice Agent. We'll map the call from new enquiry to a prepared quote callback.