TL;DR: Do not launch an AI receptionist by pointing your main number at it on a Monday morning. Start with the calls nobody answers today. Phase one is after hours, where the current baseline is voicemail, so almost anything the agent does well is an improvement and its mistakes are cheap. Phase two is overflow: the line rings for your staff first and only forwards to the agent when nobody picks up. Phase three, if the first two earn it, is the agent as the front door during busy hours or all day. Each phase has a clear baseline to measure against, a narrow job, and a way back. The work that makes it safe is mostly unglamorous: one source of truth for business hours, a greeting that matches how the caller arrived, and an escalation path that never forwards a caller back into the same busy line.
At Fortell AI I built production voice and chat agents for hospitals, real estate, vet clinics and car garages. The businesses were very different, but the most useful question in every early conversation was the same: which calls are you missing right now? The answer almost always pointed at evenings, weekends and the busiest hour of the day. That is where an agent should start.
Why "replace the front desk" is the risky default
The pitch for an AI receptionist tends to be all or nothing: it answers every call, 24/7, so staff can stop picking up the phone. The trouble is everything around the technology.
When the agent takes every call from day one, three things happen at once:
- Every mistake lands on your best callers. The regular client who calls at 10am on a Tuesday now meets the agent before anyone has seen how it handles real traffic.
- There is no baseline to compare against. If bookings dip in week one, you cannot tell whether the agent caused it or it was just a slow week.
- Staff feel replaced rather than helped. The people who know the business best, and who will be reading the post-call summaries, start the project hoping it fails.
A phased rollout fixes all three. It is slower by a few weeks and much safer, and in my experience it is also the easier sale, because the first phase asks the business to change almost nothing about how it works today.
Phase 1: after hours
After-hours calls are the ideal starting point because the bar is so low. Today those callers get a voicemail greeting, and a large share of them hang up without leaving a message. An agent that answers, understands why they are calling and captures their details is already better than that, even before it books anything.
What the agent owns in this phase:
- Capture. Name, number, reason for calling, and the one or two details the business needs to call back usefully. I wrote up the hard parts in capturing caller details correctly.
- Booking, if the calendar allows it. If the business already takes bookings online, the agent can do the same through the same calendar. If it does not, phase one is capture only, and that is fine. See how I wire calendar booking.
- Urgent escalation. This is the part that needs the most thought, because after hours there is often nobody to transfer to.
That last point is where the verticals differ. For vet clinics, an after-hours caller with a sick animal needs to be told plainly where the nearest emergency service is, not offered a callback tomorrow. For auto repair shops, a breakdown at the roadside needs a recovery number, not an appointment. For hospitals, anything that sounds clinical goes to the proper emergency and clinical lines, and the agent stays strictly administrative. Each business gives me its own after-hours escalation script, and I test the emergency paths before anything else.
The agent also has to know it is after hours and say so honestly. "The office is closed right now, they open at 8am, and I'll make sure the team has your details first thing" sets the right expectation. An agent that implies someone will call back in ten minutes at 11pm creates a complaint the next morning.
Phase 2: overflow during business hours
Once after hours is stable, the next set of missed calls is the ones that ring out during the day because everyone is busy: the receptionist is with a client, two lines are already in use, it is lunchtime. Overflow routing catches exactly those.
The mechanics are simple. The main number rings the business's own phones first, and if nobody answers within a set number of rings, or the line is busy, the call forwards to the agent. Most phone systems and carriers support this as conditional or no-answer forwarding, or you can do it in a Twilio flow that dials staff first and falls back to the agent. The setup choices are covered in setting up the phone number for an AI receptionist.
Three details make overflow feel right rather than odd.
The greeting has to match how the caller arrived
An overflow caller has already listened to the phone ring for fifteen or twenty seconds. Opening with the same breezy greeting as an after-hours call sounds strange. Something like "Thanks for holding, the team's on other calls right now, I can help you or take a message" acknowledges what just happened. I pass the routing reason into the agent as a variable so it knows whether this is an overflow call or an after-hours one, and the opening line changes with it.
Never transfer back into the line that just failed
The most common overflow bug I look for is a loop. The agent decides the caller needs a human, transfers them to the main number, the main number is still busy, and the call forwards straight back to the agent. The caller hears the agent greet them a second time and hangs up angry.
The fix is to give the agent a different destination for warm transfers than the number that forwarded the call, such as a direct line or a manager's mobile, and to have a defined fallback when that does not answer either: take the details, promise a callback, and flag it as priority. My general rules for this are in voice agent human handoff.
Pick the ring timeout with the staff, not for them
Too short and the agent takes calls the team would have answered. Too long and callers hang up before the agent gets a chance. I set an initial value with the person who actually answers the phone, then adjust it after looking at a week of call logs.
Phase 3: the agent as the front door
Some businesses stop at phase two and are happy. Others want the agent answering first during peak hours, or all day, with staff handling transfers and callbacks. That is a legitimate end state, but I only recommend it when the first two phases have produced evidence:
- The summaries are good enough that staff act on them without opening the recording.
- The escalation and transfer paths have been exercised by real calls, not just test calls.
- The edge cases the business worried about at the start have come up and been handled.
- The team trusts it, which you can tell because they have stopped asking whether it is working.
If those are true, widening its hours is a small configuration change. If they are not, widening its hours just exposes more callers to the same gaps.
The technical spine: one source of truth for hours
Phased rollouts introduce one problem a single always-on agent does not have: the system has to know what time it is, in the business's timezone, and whether the business is open. That fact gets used in several places at once: the carrier or Twilio routing, the agent's greeting, whether it offers today's slots, and what the escalation path is.
The failure I design against is hours defined in three places that drift apart. The routing knows the office closes at 6pm, the prompt says 5:30pm, and nobody updated either for the bank holiday. So I keep business hours, holiday closures and the after-hours escalation numbers in one place the automation reads, usually alongside the client's other configuration in the CRM or an n8n data store, and I pass the current local time and an "open or closed" status into the agent at the start of each call. Holidays are where this pays for itself: one record changes, not three systems.
How to measure each phase
The best part of phasing is that each stage has an honest baseline.
- After hours: compare against the voicemail period before launch. How many after-hours calls ended with usable details or a booking, versus how many voicemails were left. Calls that were previously lost entirely are the headline number.
- Overflow: compare against missed calls during business hours before launch. The question is how many of those became captured leads or bookings, and whether the ring timeout is stealing calls the team would have taken.
- Front door: compare against the overflow phase. Did booking rate, transfer success and complaint rate stay steady as the agent took more calls?
Measure per call outcome, not just call volume. I go into which numbers matter in voice agent metrics that matter, and what to watch in the first days of each phase in the first two weeks after a voice agent goes live.
When I skip the phasing
Phasing is a default, not a rule. I skip it, or compress it to a few days, when:
- The business genuinely has nobody answering today. A solo operator who is on job sites all day is effectively "after hours" all the time, so the baseline is already voicemail.
- The agent is outbound, not inbound. A follow-up or reactivation campaign has its own ramp, based on list segments rather than hours, which I described in outbound voice agents vs inbound.
For everyone else, especially a first agent in a business where the phone is how money comes in, phasing is the cheapest insurance available.
The short version
Start where the baseline is voicemail. Let the agent earn wider hours by doing a narrow job well. Make the greeting match the caller's route, never transfer back into a line that is already failing, and keep business hours in one place. If you are still deciding whether an agent is the right tool at all, AI receptionist vs answering service vs hiring is the place to start.
I build production AI voice and chat agents on Retell AI, wired into n8n, GoHighLevel and Twilio, for businesses across the US, UK and Europe. If you want an agent that starts by catching the calls you miss today and grows from there, more about me is here, or book a call and we can map your missed calls first.