I'm going to do something most founders won't. I'm going to tell you what blew up.
Over the last 18 months I shipped AI voice agents into 12 mortgage brokerages. Some are killing it. Some I had to rip out and rebuild. Here's the honest tape.
What worked beyond what I expected
The 2 AM call. Mortgage shoppers research at weird hours. They sit on Zillow at 11:30 PM thinking about their down payment. They scroll TikTok at 1 AM and see a "mortgage broker hack" video and click. Then they call. And every human broker in Canada is asleep.
Our AI agents answered 38% of all qualifying calls between 10 PM and 8 AM. Thirty-eight percent. Those leads previously hit voicemail and were never returned. Now they're booked into Tuesday-morning consultation slots before the broker has had coffee.
The AI doesn't compete with your broker. It competes with sleep, lunch, and the bathroom break. And it wins every single time.
What I got wrong on attempt #1
I tried to make the AI quote interest rates. Disaster. Rates change hourly. Compliance hated it. Applicants got quoted numbers that were 12 hours stale. We pulled it within a week.
Lesson: AI voice agents qualify and book. They do not advise. That's the line. Cross it and you make compliance officers want to murder you.
The integration problem nobody talks about
Every mortgage broker uses different software. Velocity. Filogix. Newton. Finmo. Lendesk. Some still use Excel and prayer. The AI is the easy part — making it write to your existing CRM without breaking your workflow is the hard part.
I built AICanadianSolutions specifically because every "AI voice agent" vendor I evaluated had one CRM connector and called it done. Brokerages don't run on one CRM. They run on five tools and three spreadsheets. The platform has to bend to the broker, not the other way around.
The numbers from production
Aggregated across the 12 brokerages, after 6 months in production:
- Inbound calls answered < 2s: 99.4%
- Calls that converted to a booked qualifying consult: 67%
- After-hours bookings (would have been lost): 38% of total bookings
- Average call handle time: 3:18 (faster than a human ISA)
- Cost per qualified appointment: $2.40 (vs $48 for outsourced ISA)
The thing nobody else will tell you
An AI voice agent doesn't replace your top producer. Your top producer was never going to answer the inbound phone anyway — they're closing deals. The AI replaces the receptionist, the after-hours voicemail, and the ISA you keep meaning to hire but can't quite afford. It's an org-chart compressor.
If you're a mortgage broker in 2026 still routing inbound calls to a human queue, you're running a 1998 business model with a 2026 logo.
What I'd do differently
Three things, in order:
1. Pick one CRM integration first, ship it, then add the next. I tried to support five on day one. Cost me 3 months.
2. Voice cloning is overrated for B2C qualifying. A professional generic voice converts as well as a cloned voice and saves you the legal headache.
3. Build the dashboard before you build the agent. Brokers don't trust what they can't see. The full transcript + audio playback + outcome tags per call was the feature that got every renewal.
The AI voice manager app I built lets a broker review every call from their phone in 30 seconds. That single screen — not the AI itself — is what made the product sticky.