DEV Community

RyanCwynar
RyanCwynar

Posted on • Originally published at ryancwynar.com

I Built an AI Prospecting Pipeline That Runs Itself

Most people think about AI sales tools as fancy chatbots that send LinkedIn messages. I built something different: a fully autonomous prospecting system that finds leads, qualifies them, and queues them for outreach — without me touching anything.

Here's how it works and what I learned.

The Problem

I'm building AI-powered services for local businesses — think AI receptionists, automated review management, that kind of thing. The target market is doctors, lawyers, and dentists in South Florida. Great market, but finding prospects manually is soul-crushing work.

You search Google Maps, check if they have a website, find their phone number, figure out if they're a good fit, add them to a spreadsheet... multiply that by hundreds and you've got a full-time job that produces zero revenue.

The Architecture

The system runs on cron jobs through my personal AI assistant (yes, I have a Claude-powered assistant that runs 24/7 on my server). Here's the flow:

  1. Scheduled searches fire every few hours across different campaigns — receptionist, reviews, AI services
  2. Google Maps scraping finds businesses matching specific criteria in target cities
  3. Deduplication checks each prospect against the existing queue (this alone saves massive time)
  4. Qualification filters out businesses without phone numbers or that don't fit the campaign
  5. Queue insertion adds qualified prospects to a Convex database ready for outreach

Each campaign has different search parameters. The receptionist campaign targets medical practices and law firms. The reviews campaign focuses on businesses where online reputation matters most. The AI campaign goes broader — CPAs, financial advisors, anyone who answers phones all day.

What Actually Surprised Me

Duplicate rates are insane. On a typical run, 60-70% of results are already in the queue. This tells me two things: the system is thorough, and the local business market in any metro area is finite. After a few weeks you start saturating your target geography.

Campaign diversity matters. Running the same search query repeatedly yields diminishing returns fast. By rotating through different business types and cities, each run still finds 2-6 new prospects. That's 10-30 per day across all campaigns, completely hands-free.

The queue grows faster than you can call. I'm sitting at 166 prospects right now. Even with AI-powered outbound calls, working through that list takes time. This is a good problem to have, but it means the bottleneck shifted from finding leads to converting them.

The Tech Stack

  • OpenClaw (Claude-powered assistant) handles the orchestration
  • Cron jobs trigger prospect searches on schedule
  • Convex (self-hosted) stores the prospect database
  • Web scraping via search APIs for business discovery
  • Twilio + AI voice for the outbound call layer

The whole thing runs on a single VPS. No Lambda functions, no Kubernetes, no managed services beyond the APIs themselves. Total infrastructure cost is maybe $30/month.

Lessons for Builders

Start with the boring automation. Everyone wants to build the sexy AI chatbot. The real leverage is automating the boring stuff — data collection, deduplication, scheduling. These are the tasks that eat your time and produce no creative output.

Deduplication is a feature, not a bug. When I first saw 70% duplicate rates I thought the system was broken. It's not — it's working exactly right. Those are hours of manual checking I'm not doing.

Let it compound. The system has been running for about a week. It's added hundreds of qualified prospects without me doing anything. That's the power of automation that runs while you sleep, eat, and work on other things.

Keep the stack simple. I could have built this with a dozen microservices and a message queue. Instead it's cron jobs and a database. Simple systems are easier to debug, monitor, and extend.

What's Next

The prospecting pipeline feeds into an AI calling system that actually phones these businesses and pitches the service. That's a whole other article. But the key insight is this: the hardest part of sales isn't the pitch — it's having enough at-bats. This system gives me unlimited at-bats.

If you're building a service business and spending hours on manual prospecting, stop. Build the pipeline first. Even a simple version will pay for itself in the first week.

Top comments (0)