- Termite and pest control
- Southern California, multiple offices
- $28K+ phased build
- English and Spanish
The problem
Every call into a termite and pest company is a booking, a question about an inspection, or a customer who needs a person. When the phones ring past what the CSR team can pick up, the caller hangs up and calls the next company on the list.
This client runs several offices with a shared customer service team. They wanted inbound and after-hours calls answered, intake captured correctly, and anything complex handed to the right person, without changing how the offices already work.
What we built
- A voice receptionist that answers in English or Spanish and switches when the caller does
- Intake into the calendars and field-service CRM the team already runs, not a parallel system
- Handoffs to the right CSR when a caller needs a person
- An answer library the client’s own staff can edit, so the agent says what the company would say
- Per-office routing switches the client controls, with an alert every time one changes
- Back-office automation alongside the phones: inspection-report data extraction for staff review, and a technician license-expiration dashboard
Start narrow, then widen
The client’s team proposed the rollout plan, and it was the right one: one office and one call type first, booked into calendars the team already used, then the next office once the first was solid.
We rebuilt the call flow lighter to fit and kept a separate version per office, so each office goes live on its own schedule and nobody waits on the slowest integration.
Shipping without breaking the phones
Every release runs through a synthetic-caller harness. A realtime model places live test calls into the agent over the real phone path, drawn from a library of about 200 scenarios, and the release is gated on batch pass rates instead of a single lucky call.
After a change ships, real calls are checked and a regression rolls back automatically. Changes that touch compliance wording wait for a human sign-off.
Honest by default
Every call opens with a recording notice. The agent does not pretend to be a person: asked directly, it says it is an AI. Recordings and transcripts are kept for 180 days, not forever.
What we learned
The model was rarely what failed. Silence did, when latency defects showed up as dead air. Routing did, when calls in one metro area reached the wrong place. Staff readiness did, when an early rollout paused until the team was ready for software answering its phones.
Each of those is now a named check in how we scope the next build, and all three are written up in Paper 1 of our research series.
Where it stands
The agent is in live testing with the client’s team ahead of go-live, office by office. The full CRM integration is being built in phases alongside it.
What we are not claiming
No booking-rate, missed-call or revenue numbers appear here yet, because none have been measured on production traffic and signed off by the client. When they are, this page gets them.