AI Cold Calling Software: How It Works and What It Costs
AI cold calling software replaces human SDRs on outbound calls using a voice AI agent that dials numbers, speaks naturally, handles objections, and books meetings — automatically. The real cost is $0.07 per call. A human SDR runs $0.50–$1.00 per call. Here's the full breakdown of how the technology works and which platforms are actually worth using.
How AI Cold Calling Software Works
The technology stack has four layers. Understanding each one matters when you're comparing platforms or debugging why your calls sound robotic.
The lead picks up. Deepgram transcribes their speech in ~200ms. The LLM generates a response. ElevenLabs converts it to audio. The whole round-trip — from them speaking to the agent responding — runs in under 700ms on a good day. That's fast enough to feel like a real conversation.
The failure points are almost always in the STT layer (mishearing words) or the endpointing settings (the agent cutting in before the person finishes speaking). Both are tunable.
Real Cost Breakdown: AI vs Human
I've run this comparison on real campaigns. Here's what the numbers look like at 75 calls/day:
That's not a typo. A hundred-fold cost difference per call. The caveat: the AI caller doesn't close deals. A human still handles the actual sales conversation once the meeting is booked. The AI just fills the calendar.
Full cost breakdown per call: VAPI infrastructure ~$0.05/min · ElevenLabs voice ~$0.01/min · Twilio carrier cost ~$0.01/call · Average call duration 90 seconds = ~$0.065 per call. Round to $0.07 with buffer.
Platform Comparison: VAPI vs Bland.ai vs Retell
These are the three platforms I've evaluated. I run VAPI in production. Bland.ai and Retell I've tested on staging environments for client evaluations.
| Platform | Pricing | Voice options | Latency | Custom LLM | Best for |
|---|---|---|---|---|---|
| VAPI | ~$0.05/min + voice costs | ElevenLabs, PlayHT, OpenAI, Deepgram | 600–900ms | ✓ Full control | Developers who want full control over every layer |
| Bland.ai | $0.09/min flat | Built-in voices + custom cloning | 400–700ms | ~ Limited | Non-technical teams wanting fast setup |
| Retell AI | $0.07/min + usage fees | ElevenLabs, OpenAI, custom | 500–800ms | ✓ Via webhook | Mid-market, good dashboard and analytics |
My take: VAPI wins on flexibility and cost at scale. Bland.ai is the easiest to get live in an afternoon. Retell has the best out-of-the-box analytics dashboard if you're managing this for clients who want reporting without digging into logs.
How I Built Amy: A Real Production Setup
Amy is the AI caller I built for an Australian tradie lead generation campaign. Here's the actual configuration:
- Platform: VAPI (assistant ID:
- Voice: ElevenLabs Matilda — warm, professional, Australian-friendly accent
- STT: Deepgram Nova-2 — best accuracy on Australian English, especially tradesperson vocabulary
- LLM: GPT-4o with a system prompt that includes 4 script variants (A/B/C/D), objection handling trees, and explicit "do not say" rules
- Call window: 7–9 AM Sydney time — tradies are driving to jobs, phone in hand
- Daily cap: 75 calls — stays within carrier thresholds that trigger spam flags
- Booking: Calendly link sent via SMS after the prospect agrees to a screen share
The script variants matter. Amy-A works for cold contacts who've never heard of the service. Amy-D is for warm callbacks. Running A/B tests on scripts is how you move from 3% booking rate to 5%+.
Want the full technical deep-dive on VAPI? Read: VAPI Review: Honest Assessment After Running It in Production
What Actually Converts on AI Cold Calls
After running hundreds of calls, these are the patterns that move the needle:
- Opening under 10 seconds. State your name, company, and the one thing you're offering. Don't ask "how are you?" — it signals sales call immediately.
- One-sentence value proposition. "I help [job title] in [location] get [specific outcome] without [pain]." Specific beats vague every time.
- Acknowledge the awkwardness. Amy says: "I know this is an automated call — I'll keep it short." Open acknowledgment reduces hang-ups by roughly 20% in our tests.
- Single ask. Not a sale. Not a demo. A 15-minute screen share. Lower friction = higher yes rate.
- Concrete objection handling. "I'm busy" → "Totally, I can SMS you a link and you book whenever suits — takes 2 seconds." This recovers 15–20% of "I'm busy" responses.
The script mistake I see most: Trying to close too fast. The AI caller's job is to get a "yes, send me more info" or a booked slot. Trying to explain pricing, features, or case studies on a cold call tanks conversion. Save that for the human sales call.
Compliance: What You Need to Know
This is not optional. AI calling is regulated differently by country:
- US: FTC rules require disclosure that the call is AI-generated. TCPA governs autodialing to mobile numbers — you need prior express consent for most consumer contacts.
- Australia: ACMA's Do Not Call Register. Register your campaign, scrub your list against the DNC list before dialing. Tradies who are registered businesses are generally callable.
- UK / EU: ICO regulates automated calls. Legitimate interest basis is narrow for cold calls. Get legal advice before launching in the EU.
When AI Cold Calling Doesn't Apply
- Your audience won't answer unknown numbers. If your prospects screen all calls, the answer rate drops to 10% or less and the economics fall apart. Check this before building.
- Your sales cycle is complex. AI can book a discovery call, but it can't replace a 45-minute consultative conversation. If your deal requires deep qualification before booking, a human should handle initial outreach.
- You're targeting C-suite executives. Cold calling the CEO of a 500-person company with an AI agent is a brand risk. This channel works best for SMB, local businesses, and sole traders.
- You don't have a phone list. Email verification is cheap; phone number verification is significantly harder and more expensive. If your data is email-only, start with cold email.
FAQ
How does AI cold calling software work?
It dials numbers automatically, then uses a voice AI agent (STT → LLM → TTS pipeline) to hold a real conversation. The agent follows a script, handles objections via pre-configured responses, and books appointments by sending a calendar link via SMS after the prospect agrees.
How much does AI cold calling cost compared to a human?
Running VAPI with ElevenLabs costs roughly $0.07 per call. A human SDR in the US costs $15–25/hour handling 30–40 calls/hour — that's $0.50–$0.83 per call. At 75 calls/day, the AI costs $5.25 vs $562 for a human. The AI is 100x cheaper per call.
What is VAPI and why is it used for AI cold calling?
VAPI is a voice AI infrastructure platform handling the full real-time audio pipeline: STT, LLM inference, TTS, and call management in one API. Developers use it because it abstracts the hardest parts — low-latency audio streaming, interruption handling, and carrier integration.
What conversion rate should I expect?
For targeted lists with a clear offer, 3–7% of answered calls result in a booked meeting. Answer rates for tradies and local businesses run 40–60%. Combined: expect roughly 1.5–4% of all dialed numbers to convert to a booked meeting.
Can AI cold callers handle objections?
Yes, for pre-configured objections. The LLM can improvise somewhat, but the most reliable approach is explicitly scripting 5–8 common objections in the system prompt. "I'm busy" → schedule callback. "Not interested" → one clarifying question before accepting the no.
Want an AI Caller Built for Your Business?
I build VAPI-based voice agents from scratch — script, voice, objection handling, CRM integration. Apply to work with Roman and I'll tell you exactly what it would take for your use case.
Apply to Work 1-on-1 with RomanOr join my free community — AI Mastery Genesis on Skool — where I drop the templates I use to build these agents.