A voice AI agent rebuilt this finance broker's lead qualification flow end to end. Cost per qualified call dropped from $42 to $1.20. Here is exactly what changed, what it cost, and what did not work the first time.
What was making lead qualification so expensive in the first place?
The broker was running a standard setup for an Aussie SMB: a comparison site sending inbound enquiries, a BDC staff member calling those leads within a few hours, and a CRM entry made after each call. Clean enough on paper. Brutal in practice.
The maths looked like this. Each staff member cost roughly $65,000 a year fully loaded. They were making around 25 to 30 dials a day, factoring in prep, CRM data entry, callbacks, voicemails, and the dozen leads who picked up but were not remotely ready to engage. Of those daily dials, maybe eight converted to a genuine qualification conversation. That is your $42 a call before anyone has even looked at whether the lead is worth a broker's time.
The second problem was timing. Lead response time is one of the strongest predictors of contact rate. Research published by Harvard Business Review found that firms responding within an hour were nearly seven times more likely to have a meaningful conversation than those who waited even 60 minutes. This broker's average first-touch time was sitting at four to six hours. By then half the leads had already called someone else.
Third problem: the leads arriving from comparison sites are not pre-qualified. You get someone who clicked a button at 11pm wondering if they could refinance. You do not know their loan size, their credit situation, their urgency, or whether they actually read what they were clicking. That context has to come from somewhere. It was coming from an expensive human making calls during business hours.
Lead qualification at this volume is a data-collection problem first. It just happens to look like a people problem.
What does the rebuilt lead qualification flow look like in production?
The replacement system runs on Retell AI for the voice layer, N8N for orchestration, and Pipedrive as the CRM. No exotic stack. The same tools described in the Retell vs Vapi vs Bland comparison we ran on 200 Australian broker leads, where Retell won on latency and call stability for this kind of outbound scenario.
When a new lead lands in the CRM, N8N fires within 90 seconds. The voice agent calls the lead, introduces itself as an assistant calling on behalf of the broker, and runs through a structured lead qualification script. Five questions. Loan purpose, estimated loan size, employment status, whether they own property, and their timing. The call averages three minutes and twenty seconds.
At the end of the call, the agent scores the lead against a qualification matrix defined by the broker. Hot leads get an SMS immediately telling them a broker will call within two hours. Warm leads get an SMS and a 24-hour callback scheduled in Pipedream. Cold leads get a follow-up email sequence and are parked. The broker never sees a cold lead until they warm up.
Every call is recorded, transcribed, and the summary pushed into the Pipedrive contact record before the agent has even hung up. The broker opens a CRM record and already knows the loan size, the urgency, and whether the person is employed. That is the context the staff member used to spend 42 dollars assembling.
How does the cost per lead qualification call actually break down?
This is the number people want. Here it is, plainly.
Retell AI charges per minute of call time. At roughly three and a half minutes per call, and using the current Retell pricing tier appropriate for this volume, the telephony and AI layer costs around 60 to 70 cents per completed call. Add N8N cloud execution costs, which are negligible at this volume, maybe two cents a run. Add the Twilio number rental and per-minute termination rates for Australian mobile numbers, which adds another 30 to 40 cents per call. The SMS notification costs under five cents.
Total: approximately $1.15 to $1.25 per lead qualification call. We use $1.20 as the working figure.
The human cost did not disappear entirely. Someone still needs to review the qualification matrix monthly, listen to a sample of calls, and handle the edge cases the agent escalates. That overhead is roughly two hours a month at this lead volume. It is not zero. It is just not $42 per call anymore.
The broker's staff shifted from making 25 to 30 qualification calls a day to spending their time exclusively on pre-qualified conversations with people who already know they are going to be called by a broker. Close rates on those conversations went up, because the broker was no longer burning energy on people who clicked a comparison site by accident at midnight.
What compliance guardrails did we build into the lead qualification system?
This is not optional. Australia has real teeth on this.
The ACMA's Do Not Call Register applies to AI-initiated calls the same way it applies to human-initiated calls. Before any outbound dial, the number is checked against the DNCR. That check is automated inside the N8N workflow. If a number is registered, the workflow routes to an email-only follow-up and logs the suppression in Pipedrive. No call is made.
Call timing is restricted to ACMA-compliant windows: 9am to 8pm Monday to Friday, and 9am to 5pm Saturday. No Sunday calls. These windows are hard-coded into the N8N scheduler, not left to chance.
The agent's opening disclosure is explicit. It identifies itself as an AI assistant, names the broker it is calling on behalf of, and gives the person an immediate out: they can say not interested and the call ends cleanly. We trained the agent to handle this without friction, which is covered in more detail in how we train AI voice agents to handle difficult callers.
Call recordings are stored in accordance with the Privacy Act 1988. Consent to record is disclosed at the start of each call. Recordings are retained for the period the broker's PI insurance requires, then deleted on a scheduled purge.
Building this without the compliance layer is not a cost saving. It is a liability.
What did not work the first time we built this?
The first version of the lead qualification agent had three problems worth naming.
First, the qualification questions were written for a human conversation. They assumed the person on the other end would volunteer information, ask clarifying questions back, and generally behave like someone who had agreed to be interviewed. They had not. They clicked a button. The script needed to be tighter, more directive, and forgiving of one-word answers. We rewrote it twice before the completion rate lifted above 70 percent.
Second, the escalation logic was wrong. The original version tried to handle too many edge cases inside the agent itself: people who were not the applicant, people calling about a different product, people who wanted to speak to someone immediately. The agent got tangled. The fix was simple: anything outside the five core lead qualification questions gets a warm transfer flag and a human callback within 30 minutes. The agent does not try to be a broker. It qualifies or it escalates.
Third, the CRM integration was writing incomplete data on calls where the person dropped early. Those partial records were causing confusion in Pipedrive. N8N now writes a status field on every call regardless of completion: qualified, partial, declined, or no answer. The broker can filter on these instantly.
None of these were catastrophic. They were the normal cost of building something real instead of demoing something pretty.
What kind of finance broker does this work best for?
This lead qualification setup suits brokers who are receiving inbound comparison site leads, aggregator leads, or paid search leads at volume. Roughly speaking, that means 20 or more new enquiries a week. Below that volume, the setup cost and ongoing maintenance may not return enough to justify it over just hiring a part-time BDC.
It also suits brokers who are currently using a BDC staff member or a shared admin person to do first-touch calls. If that person is doing anything more than qualification, which is to say if they are also doing broker support, compliance prep, or client liaison, then pulling them off qualification calls frees time that has clear dollar value.
Brokers who receive referral-only leads are a different conversation. Referral leads carry context and relationship. An AI handling the first call with someone a financial planner personally referred is a different risk calculus, and usually not the right fit without careful agent design.
Mortgage brokers, asset finance brokers, and commercial finance brokers all have slightly different qualification matrices. The loan size thresholds, employment questions, and asset questions change. The N8N workflow is built to accept a different qualification config per lead source, so a broker running both residential mortgage and asset finance enquiries can run different agents against each pipeline without rebuilding anything.
The economics at $1.20 per lead qualification call become particularly obvious when you are paying $15 to $30 per lead from a comparison site. The cost to find out if that lead is worth a broker's time should not exceed the cost of the lead itself.
Want to see if voice AI fits your business? DM me on LinkedIn with the word AUDIT and I'll send 5 questions. Honest answers get you a 30-min Calendly slot or a 'no, here's why' email.
Originally published at theautomate.io.
Top comments (0)