Most AI receptionists fail inside the first two weeks -- not because the technology is bad, but because nobody tuned them. We have seen it happen to HVAC owners in Calhoun, plumbers in Rome, and roofers in Dalton who bought a voice bot, plugged it in, and watched it fumble caller names and mis-book service windows.
The fix is not a smarter model. The fix is a structured onboarding that treats the first 30 days as a calibration period, not a launch party. Here is exactly how we run it at Forge.
Why "Plug and Play" Is a Lie
Every vendor demo sounds clean. A friendly voice answers, collects the caller's address, confirms a time slot, and hangs up gracefully. That demo was recorded after weeks of tuning against a controlled script. Your real calls are not controlled.
Real callers in Cartersville say "my unit is making that noise again." Real callers in Canton call back three times in a row because they do not trust a machine. Real callers speak Spanish, mumble, or ask for "Mike" because Mike fixed their furnace two winters ago. An out-of-the-box voice agent will stumble on every one of those cases.
The vendors who sell you on five-minute setup are optimizing for the sale, not for your answer rate six months from now. We optimize for the six-month number. That requires 30 days of deliberate, engineer-led calibration before we call the system production-ready.
The first 30 days are not a launch -- they are a tuning period. An AI receptionist that skipped this step is just an expensive voicemail box.
Days 1 Through 7: Baseline and Business Logic
The first week is about capturing what your business actually does, not what a generic template assumes. We start with a structured intake call that usually runs 45 to 60 minutes. We are not selling you on features during that call. We are extracting signal.
We document the following before a single live call is routed to the agent:
- Your service area -- every zip code, city, and county you will and will not cover
- Your job types and how you categorize urgency (emergency heat-of-summer AC call versus routine seasonal tune-up)
- Your dispatch model -- whether you run ServiceTitan, Jobber, Housecall Pro, or a paper schedule
- After-hours rules, on-call technician routing, and hold escalation thresholds
- Your pricing posture -- do you quote ranges on the phone or send an estimator?
- Key staff names callers are likely to ask for by first name
We then build the first version of your agent's call flow from that documentation. This is not a drag-and-drop template. It is a prompt architecture written by a Forge engineer who understands how service-business calls actually unfold. By end of day seven, the agent is running in a sandbox, and we are making internal test calls to break it.
Days 8 Through 14: Live Traffic With a Safety Net
Week two is the riskiest week, which is why we do not leave you alone in it. We move from sandbox to a live call whisper setup. The agent answers every call, but your existing number still rings through to a human after the agent's greeting if the caller explicitly asks or if the agent detects an edge case it cannot handle confidently.
Every call during this week is logged, transcribed, and reviewed by a Forge engineer -- not an automated grader. We are listening for three things: missed intent, incorrect field capture, and dead-air moments where the agent paused too long or gave a non-answer.
A plumber in Rome went live with us in early spring. In week two, we caught that their agent was asking callers for a "service address" but a significant portion of callers said their neighborhood name instead of a street address. The agent was not mapping "Coosa Valley Estates" to a serviceable zip. We fixed that mapping before week three. That kind of catch does not happen if nobody is reviewing the transcripts.
By the end of day 14, we expect the agent to handle at least 80 percent of inbound call types without escalation. If we are not there, we do not move to the next phase -- we iterate until we are.
Days 15 Through 21: Integration Hardening
Week three is where the workflow integrations get stress-tested. If you are on ServiceTitan, we are confirming that job records are being created correctly, that the booking source tag is mapped, and that the agent is respecting your dispatch board's open slots rather than hallucinating availability. Same process for Jobber and Housecall Pro.
We also run the bilingual test suite this week. If you operate anywhere in Gordon, Floyd, or Whitfield County, there is a meaningful probability that a caller will open in Spanish. Our Forge Voice bilingual agent is built to detect language at the first utterance and switch without asking the caller to press a number. We confirm that detection is working reliably against a set of native-speaker test calls we run internally.
Edge cases we specifically test in week three:
- Caller hangs up before giving a phone number -- does the agent recover gracefully on callback?
- Caller asks for a specific technician by name -- does the agent acknowledge and route correctly?
- Caller calls from a blocked number -- does the agent handle the missing caller ID field without breaking?
- Caller disputes an appointment time the agent offered -- does the agent escalate or loop?
- Caller speaks heavily accented English -- does transcription accuracy hold above our threshold?
Each of these scenarios gets a pass/fail in our internal tracking sheet. Anything that fails gets a corrective prompt or flow update before we move to week four.
Days 22 Through 30: Performance Baselining and Handoff
The final week is about locking in the metrics that will govern ongoing performance. We pull the full call log from weeks two and three and compute the numbers that actually matter to a service business owner:
- Answer rate -- percentage of inbound calls the agent handled without transfer or voicemail
- Booking conversion rate -- calls that resulted in a confirmed appointment or job record
- Escalation rate -- calls that required a human to intervene
- Average call handle time -- a proxy for whether the agent is being efficient or getting stuck in loops
- Lead capture completeness -- percentage of calls where name, address, and job type were all captured
We set your 30-day benchmark numbers and schedule the first monthly review call. That review call is not a check-in for its own sake. It is the trigger point for the next round of tuning. Seasonal changes, new service offerings, updated pricing -- all of these require prompt updates. A dental office in Gainesville that adds Saturday hours needs that reflected in the agent's availability logic by the following Monday, not three weeks later.
At the end of day 30, you receive a one-page performance summary: your baseline metrics, the top three issues caught and resolved, and the tuning roadmap for the next 60 days. This is the document your operations manager can look at and understand without sitting through a product demo.
What This Looks Like in Practice for a Trades Business
Take an electrical contractor in Dalton running four trucks and a two-person office. Before Forge, the office was fielding 60 to 80 calls a day across two lines. After-hours calls went to voicemail. Roughly 15 percent of those voicemails were from callers who also called a competitor before the office opened the next morning.
After the 30-day onboarding, the agent was handling 74 percent of all inbound calls without a human touch. After-hours booking was live. The office staff went from triaging calls to managing jobs. The agent was routing Spanish-language calls to the appropriate technician notes automatically.
That outcome did not happen on day one. It happened because someone spent 30 days breaking the agent before customers could.
If you want to see how this applies to your specific field-service stack -- whether you are on ServiceTitan, Jobber, or something custom -- the process is the same. The variables change. The discipline does not.
Why Most Voice AI Vendors Skip This
Tuning takes time, and time costs money. A vendor optimizing for low customer acquisition cost cannot afford to put an engineer on your account for 30 days. They can afford a chatbot-style setup wizard and a help center article.
We are not built that way. Forge is a small shop in Calhoun, Georgia, and we treat every onboarding as a custom system deployment because it is one. The managed nature of our product means we carry the operational risk if the agent performs badly. That incentive structure is why we tune aggressively and why we do not hand you a login and disappear.
Ready to see the 30-day process against your specific call volume and service area? Talk to a Forge engineer and we will map out what onboarding looks like for your business before you commit to anything.