How a Custom Chocolate Maker Got Per-Product Artwork Approval and Live Inventory in Airtable
A small custom chocolate maker replaced a crashed order system and an MRP app with Airtable and Softr: per-product artwork approval plus live inventory.
How a 40-Person Pest Control Company Turned Every Recorded Call into a Nightly AI Report
A 40-person pest control company replaced manual ChatGPT call reviews with an automated pipeline that classifies every call and emails a nightly digest.
TABLE OF CONTENTS
TL;DR: A family-owned pest control company with about 40 employees and several branch offices replaced a manual routine (downloading call recordings and pasting them into ChatGPT one by one) with an AI call analysis pipeline that sends every transcribed call through OpenAI, sorts it into sold, not sold or customer service, logs it in Airtable and emails management a digest every night at 9 PM. The AI bill for analyzing those calls runs about $64 a month.
The situation
The company does residential and commercial pest control, from recurring quarterly plans to termite work and a home protection plan it wants every rep to offer. Its office team handles a lot of calls: on one day it later checked, the phone system showed 276 calls, 133 inbound and 143 outbound. RingCentral already recorded and transcribed them. Nobody had a way to use those transcripts at that volume.
The problem
The owner found Connex through Zapier's expert directory and described the routine he wanted gone:
"I have been manually taking WAV files from RingCentral call logs and uploading them to two separate threads in ChatGPT for sold leads and missed opportunities. Using ChatGPT as a coach for our call takers."
He wanted every call sorted into sold, not sold and current-customer calls, plus a daily digest he didn't have to build himself. He also wanted to know where each lead came from: paid search leads cost about $70 each, and he couldn't see which source actually produced sales.
What they'd already tried
The owner wasn't starting from zero. He had already wired RingCentral transcripts to OpenAI through Zapier and had working prompts. In his words, "I already have most of this developed, but need help fine-tuning this."
The hard part was turning a working prompt into something the whole team could rely on: results routed to the right people, no duplicates, calls that break the prompt handled, and numbers management would trust.
What we built: AI call analysis with OpenAI, Zapier and Airtable
Over about 20 one-hour ZoomFlow sessions from April 2025 to June 2026, a Connex consultant built and then tuned the system with the owner and his operations manager. It went live in pieces, so the team was using it within weeks while later parts were still being built.
1. Every transcript goes to OpenAI
When RingCentral finishes a transcript, Zapier sends it to OpenAI with a dedicated set of instructions written for this company. OpenAI classifies the call (sold, not sold or customer service) and pulls out structured fields: marketing source, whether a plan was offered or declined, praise or frustration, missed appointments or cancellations, and which rep took the call, matched against a maintained roster.
2. Results go where people actually look
At first the results posted to three channels in RingCentral's team chat, one each for sales, customer service and billing, so reps saw feedback on their own calls. Calls under a minute were dropped so the channels weren't flooded with hang-ups.
3. A Google Sheet, then Airtable
The first log was a Google Sheet with a tab per call type. Within about a month it moved to Airtable: separate tables for sold, not sold and customer service calls, a duplicate check, date-grouped views and a manager review flag for cancellations and missed appointments. Later came sold and not-sold totals, a closing percentage and an All Calls table that pulls every category into one place for reporting.
4. The 9 PM digest
Every night Zapier builds a sold-calls digest and a not-sold digest as a PDF and emails it to a set distribution list. This is the piece that replaced the owner's manual routine.
5. Structured fields for the owner's own AI agent
By 2026 the owner was running an AI agent tool on the Airtable data to write an executive daily report. Free text made it guess, so the tables got structured fields: a dropdown issue type, resolution status, the appointment window given, a same-day-miss flag and a quoted amount. New calls fill these in automatically.
What broke, and how it was fixed
These are the parts a DIY build usually gets stuck on:
- "Frustrated" customers who weren't. The AI flagged a lot of callers as frustrated. The owner's diagnosis: "it's showing a ton of frustration when the client may not be truly frustrated. It's that they have ants, bees or spiders and we are going out to correct the problem." The prompts were changed to tell the difference between someone upset about a pest and someone upset with the company.
- The word "kill" stopped the pipeline. One afternoon the Zap started failing with content-moderation errors. The voicemail check ran through Zapier's built-in AI step, which applies content moderation by default when it uses Zapier-provided models (as of November 2026). It was flagging pest control transcripts for words like "kill." Moving that check to a direct OpenAI call fixed it. The first fix briefly filtered out every call, which the owner spotted the next day, and it was corrected within hours.
- "Sold" was too generous. In one night's log, the owner found 15 calls tagged sold and only 5 real sales. The rest were existing customers booking their regular quarterly visit, payment updates and an internal call. The fix added two gates: the caller must be new or buying a new service, and a price must be discussed and accepted on that call.
- Calls going missing. One day's phone analytics showed 276 calls; Airtable had 130 entries. The causes were outdated duration filters, Zapier errors four to seven times a day from OpenAI timeouts, and a new rep missing from the roster. Now only voicemails are excluded, the Zap retries automatically, and a call direction field lets the team reconcile Airtable against the phone system.
The results
- The manual routine is gone. Before, the owner downloaded recordings and pasted them into ChatGPT himself, so only the calls he had time for got reviewed. Now every transcribed call is classified and logged without anyone touching it, and the digests arrive at 9 PM.
- Management sees the whole day. One executive report the owner forwarded broke a roughly 200-call day into about 167 scheduling calls, 21 customer service calls, 6 sales, 3 not sold and 4 cancellations, and flagged an at-risk termite job and an upsell that would otherwise have been buried.
- The numbers got more accurate. Real sales are now separated from routine rebookings, so the sold count and closing percentage reflect actual new revenue.
- Lead sources and plan offers are tracked per call. That includes whether the home protection plan was offered and declined, which is the coaching signal the owner originally wanted.
- Low running cost. Twelve OpenAI API top-ups of about $21 each between mid-March and mid-July 2026 come to about $254, or about $64 a month. At roughly 130 logged calls a day across about 22 weekdays, that's about 2,900 calls a month, or around 2 cents per call analyzed.
*Estimated impact, based on the client's OpenAI billing records and one day's logged call count. We haven't put a time-saved figure on the old manual routine because the client never measured how long it took.*
Client voice
A few weeks after the nightly digest went live, the owner told the Connex consultant:
"I love the new automation. Every night at 9 o'clock I get the email, and I'm like, this is awesome. I was doing it all manually for the longest time."
The operations manager, reviewing the reports a month later: "It's been really good. I've loved the report." Nearly a year in, the owner summed up where the system had landed: "A lot of these items we have built into Airtable where we deployed agentic AI to be sales coaches, customer service managers, and more."
FAQ
Does this need a special phone system?
It needs a phone system that produces call transcripts and can trigger an automation when one is ready. Here it was RingCentral. The same pattern works with other providers that expose transcripts through Zapier or an API.
How fast was the first measurable win?
Categorized calls were posting to team chat channels within the first two weeks. The nightly digest, the part that replaced the manual routine, was running about two months in.
Won't the AI get things wrong?
Sometimes, and you should plan for it. This build has a manager review flag, owner spot checks that caught the sold-versus-rebooking problem, and explicit rules the AI has to meet before it can call something a sale. Tuning those rules is ongoing work.
Sitting on call recordings nobody has time to review?
If your phone system already records and transcribes calls but the insight still depends on someone listening after hours, that can run on its own every night. Book a free call with Connex and we'll map out what that looks like for your team.
Related Posts
How a Restaurant-Tech Company Built a Photo-Reading AI SMS Support Agent with n8n
A restaurant-tech company built an AI SMS support agent that reads customer photos, answers from its Notion knowledge base, and hands unclear cases to humans.
How a Mid-Sized Travel Call Center Cut Manual Call Review from 8 Hours a Day to Under 1 with AI + Zoho CRM
A travel call center cut manual call review from 8 hrs to under 1 hr/day using AI + Zoho CRM automation. Here's how the system was built and the results.
