Built a multi-agent AI system for an HVAC company in Tucson, using voice and text agents to qualify leads and automate bookings, reducing dispatcher time to zero.
built this for a Tucson HVAC company last month. it's a few agents working as a team. 2 that talk on the phone, 4 that text. sharing the full setup since this sub likes real agent builds here is how it runs. someone fills out a web form. that goes to GHL. GHL sends it to n8n. n8n cleans up the phone number and sends a text right away. if they don't text back in 5 minutes, n8n starts a Retell voice agent (Kate) and she calls them. if the lead is a good fit, she books them. that booking has to pass through 5 tools in a row: Retell, n8n, GHL, Zapier, ServiceTitan. it all has to finish in under 2 minutes while she is still on the call. if one step is slow, the booking dies. so each step sends back a pass or fail so she knows right away ServiceTitan was the tricky bit. it does not use the same words the agent hears on the call. so I added a step that fixes the words (the agent hears "package unit", ServiceTitan wants "Package Unit"). then a small Zapier code step sets the date, writes a short summary and splits the address. after that it makes the booking. it shows up in ServiceTitan as an open job for the dispatcher. rest of the build: * the inbound voice agent has no saved contact, so it looks the person up, makes a new one midcall, then books * every call, picked up or not, goes back to n8n and gets saved to postgres. about two dozen things: call length, mood, result, where it went, booking status, recording link and more * the text side is one flow that picks one of 4 agents (website, engagement, followup, KB). it reads the last 60 messages from postgres, runs on gemini 2.5 flash through openrouter, has 6 booking tools and pulls answers from a supabase vector store * reactivation ran on 4000+ old contacts. it takes about 26 hours: text, wait 10 min, call, wait 2 hours, second text, wait a day, then email if still nothing. it stops the second they reply the dispatcher does no qualifying now. they just get readytogo bookings to assign and confirm I documented the whole thing up as a 150 page build doc: every flow, tool and call script. all the private info is swapped for fake values so you can copy it (you can find it in comments) happy to walk through any part
A Canadian real-estate agent with 60k+ stale CRM leads got a text and voice AI system built for $5k, which re-engaged old leads, flagged ready buyers, and produced 34 closed deals worth ~$490k in commission over seven months. The post details outreach numbers, consent handling, follow-up scheduling, and lessons learned.
The author built callitdone.today, an AI voice agent that makes real phone calls, navigates IVR menus, waits on hold, and speaks with humans, sharing key technical challenges and lessons learned.
A B2B sales rep built a full sales automation platform in 6 days using AI agents (Claude and open-source coding agents), integrating lead harvesting, enrichment, LLM-based drafting, and multi-touch sequences for under $274/month.
The author shares key lessons from building an outbound voice agent that has handled 50,000 calls, emphasizing that success lies in peripheral systems like scheduling and compliance rather than the AI model itself.