How We Built an AI Voice Agent for Franchise Leads With Retell AI
Part of our complete guide: Franchise Lead Follow-Up With AI
We've been building AI automation since 2023 and franchise lead systems since 2024. This is how we built the AI voice agent behind our case study, including the parts that went wrong.
Disclosure: we're a certified Retell AI partner, so take that into account. You can see it on Zehaan's LinkedIn.
What the agent's job is
The agent has a narrow job. It reaches a new franchise lead, finds out if they're serious and books the next step. It doesn't sell the franchise.
- Investment range
In bands
- Credit
Self-reported
- Franchise of interest
The specific brand
- Veteran status
For discount programs
- Timeline
When they want to buy
- Next step
Book, call back or text
The stack
- Lead sourcesAds, forms, feeds
- GoHighLevelEvery lead lands here
- MakeConnects CRM to calling
- Retell AIRuns the voice agent
- GHL workflowsFollow-up after the call
How a call flows
- Lead landsFrom any source
- Workflow triggersThrough Make
- Retell callsThe agent talks
- Webhook firesWhen the call is analyzed
- Fields writtenEach answer in its own field
- Tag starts follow-upBased on what they said
If you're building this, trigger step 5 on Retell's call_analyzed event. At call_ended, the transcript and analysis may not be ready yet.
Pull answers into fields
The most useful thing we did was treat every call as data. A transcript sitting in a notes field helps nobody. An "Available Capital" field can be filtered, routed and reported on.
Define your post-call fields before you write the prompt. Then write the prompt to collect those answers.
What the data showed
calls placed by the AI
unique leads worked
reached a live person
went to voicemail
of live conversation
typical live call over a minute
response latency
The 41-second average covers every call, including the ones that didn't connect and the voicemails. Live calls of a minute or more typically ran about 2.4 minutes.
Retell lists its voice agents at $0.07 to $0.31 a minute, depending on the models you choose.
What went wrong
- We counted voicemails as connected
Our first number was a 77% connect rate. But 40% of calls had reached voicemail. The live rate was 37%.
- We counted calls as people
Our first reports said 1,385 people asked for a callback. 1,038 of those were voicemails, and on live calls the real number was 355.
- We claimed too much coverage
Early copy said the AI called every lead. It worked 1,733 unique leads, while roughly 750 arrived each month.
- A comparison flattered us
Our 2024 lead count included a one-time CRM import, so we stopped comparing lead volume and conversion.
- The platform changed under us
Retell retired
POST /v2/list-callson June 15, 2026. The v3 endpoint returns results insideitems, which can quietly break reports. - Seconds to first dial is still unmeasured
It needs each call matched to the lead's created time. Until that's done, we don't publish a number.
The rules we build around
An AI caller needs consent that covers AI voice, needs to say it's an AI, and can't talk earnings. Our compliance guide covers all of it, and you should have an attorney review your setup.
Is it worth building?
For this firm, booked appointments went from an average of 89 a month to 163, and show rate from 32% to 41%. This was consultant network leads, and other things changed too, so read the full case study before you draw conclusions.
If you want to know whether you need this at all, start by seeing how fast your team responds today with a free inquiry teardown.