How We Built an AI Cold-Calling Bot That made 2K calls for $17K—6 Key Lessons You Can Use Without the Costly Learning Curve

Imagine replacing a portion of your outbound calling team with an AI bot that can talk to customers, book meetings, and even handle objections—all while following strict compliance rules. That’s exactly what we did for a solar panel installer. Our AI agent, “Jason,” dialled 2,000 customers, and from this experiment, we gained deep insights into what works and what doesn’t in the world of AI cold calling.

In this deep dive, I’ll walk you through the steps we took, how much it cost (spoiler: upwards of $17K), and how you can leverage our learnings to avoid the expensive trial and error.

The Hypotheses We Wanted to Test

Here are the key questions we wanted to answer:

1. Can an AI bot actually generate leads?
2. Is it better to use a local or mobile number for outbound calls?
3. Is latency low enough for AI to hold a natural conversation over the phone?
4. Are large language models (LLMs) capable of sustaining a customer call without sounding awkward or robotic?
5. If outbound calls work, could we also use the AI bot for inbound customer service?

Our Must-Have Requirements

To make this experiment work, here’s what we needed:

• Scalability: The ability to process a list of thousands of names and numbers for outbound calls.
• Local Presence: A Queensland-based phone number to establish trust with local customers.
• Privacy Compliance: Adherence to Australian privacy and telemarketing laws, including call recordings and an easy way to remove customers from future call lists.
• Minimal Latency: No awkward delays—calls had to feel natural.
• Conversational AI: The AI needed to sound genuinely Australian.
• Reporting: We needed detailed reports on call outcomes and metrics.

How We Set It Up (and What We Learned)

We used a platform called Synthflow, which integrates with Twilio for phone numbers. The setup wasn’t perfect—Synthflow felt like it was in an early beta stage and needed workarounds to function smoothly. Despite its immaturity, their team was responsive, and we managed to A/B test local vs. mobile numbers (hint: local numbers performed better and had cheaper rates from Twilio).

For the workflow automations we initially used Zapier, but its cost ballooned quickly, so we switched to ActivePieces, which was much more affordable. On average, we used around 3,000 tasks per month.

Key Learning: Voice Matters

The default AI voices worked well for the US market but were a disaster for Australian customers. We eventually trained our bot to use an Australian voice (shoutout to Oliver). This small change dramatically improved engagement because Australians are notoriously wary of foreign accents in telemarketing.

The Six Steps to Make an AI Cold-Calling Bot Work for You

Here are the six steps that helped us get our AI cold-calling bot, Jason, up and running:

1. Nail the AI configuration

Jason’s first iteration was too robotic (Telemarketing T-800 Terminator). He wouldn’t hang up until the customer booked a meeting. We realised that an AI bot has to be more human—asking the right questions and giving the customer an easy exit if they’re not interested. You’ll want to build and test your AI configuration to find the right balance between persistence and politeness.

2. Use the Right AI Model

We implemented Retrieval-Augmented Generation (RAG) to give Jason deep knowledge of domain-specific information, like a custom knowledge base about solar power in Australia. We also implemented custom real time actions like calculating government incentives for solar panels based on the customers postcode. This made conversations richer and prevented the AI from making things up.

3. Handle Objections with Precision

About 1 in 3 households on the Sunshine Coast already had solar panels. Jason needed to pivot quickly, upselling them to a larger system, battery installation, or complementary services like electric pool heaters.

4. Leverage Real-Time Data

One of our biggest breakthroughs was integrating real-time solar savings calculations based on customers’ postcodes and past electricity bills. This allowed Jason to deliver personalised insights on how much they could save by switching to solar, enhancing the value of the conversation.

5. Integrate Seamless Scheduling

Using Synthflow’s integration with cal.com, Jason could book follow-up meetings with solar installers in real-time by checking their calendar availability. This removed the friction of scheduling and increased our close rate.

6. Measure and Optimise

Our initial approach to managing call data through Google Sheets worked for the PoC, but scaling up requires integration with a CRM to track telemarketing data and post-call outcomes more robustly. You’ll need real-time insights to optimise your AI’s performance.

Privacy and Compliance (A Non-Negotiable Step)

Compliance was key for us. We strictly followed Australian privacy and telemarketing laws, including informing customers when calls were recorded, and we developed a smooth workflow for customers who wanted their data removed. For any AI project, ensuring privacy protection is not just best practice—it’s essential. Learn more about the Australian Privacy Principles we followed here.

Results and Metrics

Here’s what we found after 2,000 calls:

• Lead Conversions: Jason (AI) booked 20 meetings, a conversion rate of 1% giving a CAC of $50. About 50% of sales are closed giving a theoretical profit of $3K.
• Cost Comparison: The total operating cost of running the Jason for a month was around $2K, far lower than employing a full-time sales person of $6K.
• Customer Reaction: The vast majority of customers believed they were speaking to a human, which speaks volumes about the quality of our AI implementation.
• Are any sales people going to lose their jobs tomorrow to AI? I am doubtful, but I can see AI being used with agents in a co-pilot capacity perhaps doing the initial part of the call and handing off.

Aha Moments (What We Didn’t Expect)

One surprising result was how well Jason handled objections and upsells. We were also amazed at how few customers realised they were talking to an AI. Our efforts to ensure the voice was locally recognisable paid off in spades.

Key Takeaways for Your Business

1. Start Small, Scale Smart: The data you feed your AI agent is critical. If you can avoid buying lists and focus on more targeted, high-quality data, your results will improve dramatically. Our conversion rate on a cold call list was 1%
2. Be Prepared for Compliance Changes: Australian legislation is tightening around AI calls, so ensure your system is ready to identify itself as AI.
3. It’s All in the AI configuration: Fine-tune your AI’s responses. Over-aggressiveness kills conversions; a friendly, conversational, well-paced approach wins.
4. The Cost of Learning: We invested over $17K in this experiment—now you can use our learnings to avoid that steep price.

By following these steps, you can avoid the costly trial and error we faced and build an AI cold-calling system that enhances your outbound sales process without the hefty price tag. Ready to get started?