The AI answered the phone. Everything around that was the hard part.

Before writing a line of code, we spent months on the phone with service businesses around Charlotte asking one question: what happens when the phone rings and nobody's free to pick it up? The answers barely varied. Calls go to voicemail after five, during jobs, and whenever two people call at once — and the callers don't leave messages, they dial the next company. Industry call data says roughly a quarter of inbound calls to home-services businesses go unanswered [Invoca, 2024]. The owners we talked to weren't outliers. They were the rule.

The existing fixes were a full-time desk hire most couldn't justify, or an answering service that takes a message somebody still has to call back. Both leave the after-hours emergency — the most valuable call a service business gets — underserved.

A 24/7 voice agent that answers in English and Spanish, works out what the caller needs, and books the job — plus everything a real product needs around it: call triage with emergency routing to an on-call tech, real calendar booking against live availability, per-call transcription and cost accounting, prepaid metered billing with a hard spend cap, a self-serve customer portal, carrier call-forwarding onboarding with a self-verifying line test, and connectors into the field-service software the shops already run.

Every part of that list exists because a real call, a real customer, or a real failure demanded it.

Not the AI. The telephone system. A partial list of things nobody's strategy deck warns you about:

• ·Carrier registration. US carriers now block business text messages from unregistered numbers — silently. The API accepts the send; the carrier drops it. Registering legitimately requires a legal entity, a tax ID, and a review process.
• ·Phone number inventory runs dry. Some area codes simply have no local numbers available to buy on a given day. Provisioning needed a fallback ladder — exact area code, then same-metro, then any local — with honest failure states when all three miss.
• ·Call forwarding is carrier folklore. Conditional vs. unconditional forwarding uses different dial codes on every carrier, some of it undocumented. We ended up building carrier auto-detection and a self-verifying line test — the system calls the customer's line and confirms the forward actually works before anyone relies on it.
• ·Transcripts across a conference bridge. Keeping one coherent transcript alive while a call gets redirected into a live three-way demo took a second parallel recording and stitching the halves back together afterward.

These are the kinds of problems forward deployment exists for: nobody scopes them in advance, because nobody knows they exist until the system is real.

Three lessons, now the method: instrument first, because you can't improve a phone line you aren't measuring. Ship into the tools the business already runs — the calendar, the phone, the CRM — because a system nobody opens is a system nobody uses. And put a number on it monthly, because a system that can't say what it returned will eventually be turned off, deservedly.

PlumbDesk runs today, on real customer calls. If you want to see it, it lives at plumbdesk.ai.