The AI Receptionist Playbook: Scope, Build, and Measure.

The AI receptionist is usually the first voice agent a services business deploys, and it is the one most often scoped wrong. Owners either ask it to do too little (a glorified voicemail) or far too much (quote prices, diagnose problems, handle complaints). The version that works sits in between: it answers every call, captures the reason for the call, books what can be booked, and hands everything else to a person with a clean summary. This playbook covers how to scope it, how to build it, and how to measure it.

What does an AI receptionist actually do?

Strip a human receptionist's job down and most inbound calls fall into a short list of intents. An AI receptionist is built to handle the predictable ones end to end and route the rest:

• Answer and identify. Pick up on the first ring, greet the caller, and work out who they are (new prospect, existing customer, vendor, spam).
• Capture the reason for the call. In the caller's own words, written to the CRM record.
• Qualify. Ask the three or four questions your team already asks — service area, type of job, urgency, how they heard about you.
• Book. Put a confirmed appointment on the real calendar, not a request in a queue.
• Answer factual questions from an approved source — hours, service area, what to expect at the first visit.
• Route. Warm-transfer or create a callback task for anything outside its scope, with a summary attached.

It is the first use case in The AI Voice Agent Playbook for a reason: it sits on the highest-intent channel a business has, and it touches revenue on the very first call.

Why are businesses replacing or backing up the front desk?

It is rarely about eliminating a person. The U.S. Bureau of Labor Statistics puts the median receptionist wage at $18.27 per hour as of May 2025, and a single front-desk hire is not expensive relative to the revenue on the phone line. The problem is coverage. One person cannot answer two calls at once, take lunch, cover evenings and weekends, and also do the filing. In our engagements, the pattern that triggers an AI receptionist project is almost always one of these:

• Calls ringing out during peak hours while the desk is on another line.
• After-hours and weekend calls going to a voicemail nobody returns until Monday.
• A skilled office manager spending a large share of the day on routine scheduling calls.
• Inconsistent intake — every caller asked different questions, half the answers never reaching the CRM.

The best deployments treat the AI as the first line and overflow, and free the human team to handle the calls that need judgment.

What should an AI receptionist never do?

Scope is decided by what the agent is not allowed to do. We write these as Truth Boundaries before a single prompt is drafted:

• Quote a price it did not read from an approved source. If pricing is not in a validated price sheet, the answer is "a specialist will confirm that with you."
• Diagnose or advise. No medical, legal, financial, or technical diagnosis — ever. It gathers symptoms; a qualified person interprets them.
• Handle complaints or disputes. An unhappy customer goes to a human immediately, with the context.
• Promise what operations cannot guarantee. Arrival windows, same-day service, and warranty terms come from the system of record or not at all.
• Pretend to be human when asked. If a caller asks whether they are talking to an AI, the answer is yes.

The architecture behind those boundaries is covered in AI Hallucinations in B2B, and the audit trail that proves the agent stayed inside them is described in Decision Log and Truth Boundaries.

How is a production AI receptionist built?

The voice is the easy part. Modern voice models sound natural. What separates a demo from a production system is the plumbing around it. Six components:

1. Telephony. Call forwarding or a SIP trunk so the agent can answer, transfer, and record. Decide up front whether it answers every call or only overflow and after-hours.
2. An intent map. The list of call types, written from a sample of real recorded calls — not from what the owner assumes callers ask.
3. A knowledge source. One approved document for hours, service area, policies, and FAQs. The agent answers from it and nothing else.
4. Calendar write access. Direct booking into the dispatch or practice-management calendar, with rules for job types, durations, and buffers, and locking to prevent double-booking.
5. CRM write-back. Every call creates or updates a record with the caller, the reason, the outcome, and the transcript.
6. Escalation paths. Warm transfer during business hours; a callback task with an owner and a response-time target after hours. A transfer that rings into an empty office is a failure, not an escalation.

What do the consent and disclosure rules require?

Answering inbound calls is the lower-risk side of voice AI, but it is not rule-free. Three things to get right:

• Outbound callbacks are regulated. In February 2024 the FCC confirmed that the TCPA applies to AI-generated voices, which means AI calls placed to consumers fall under the same consent rules as prerecorded calls. If your receptionist ever calls back, get consent rules reviewed first. The FCC's telemarketing and robocalls guidance is the starting point.
• Call recording consent varies by state. Some states require every party to consent to recording. Announce recording at the start of the call.
• Disclose the AI. Tell callers they are speaking with an AI assistant and how to reach a person. It builds trust and keeps you on the right side of emerging state disclosure laws.

None of this is legal advice; have counsel review your scripts and consent language before go-live.

How long does it take to go live?

Timelines depend on how clean the calendar and CRM are, not on the AI. In our engagements, the sequence looks like this:

1. Week 1 — Discovery. Pull and listen to a sample of recorded calls, build the intent map, and write the Truth Boundaries.
2. Week 2 — Build. Telephony, knowledge source, calendar and CRM integrations, escalation paths.
3. Week 3 — Shadow. The agent runs on test calls and a limited share of live overflow while a person reviews every transcript.
4. Week 4 onward — Live and tuning. Full coverage for the agreed hours, with weekly transcript review for the first month.

If you want the full deployment discipline behind that schedule, see How to Deploy a Bespoke AI System: A 90-Day Framework.

Which metrics show whether it is working?

Do not grade an AI receptionist on how human it sounds. Grade it on the same outcomes you would grade a front desk on:

1. Answer rate. Share of inbound calls answered, split by business hours and after hours. This should move first.
2. Booking rate. Share of qualified callers who end the call with a confirmed appointment.
3. Escalation rate and quality. How often the agent hands off, and whether the human who receives the handoff had what they needed without calling back to re-ask.
4. Intake completeness. Share of calls where every required CRM field was captured.
5. Boundary violations. Any call where the agent quoted, promised, or advised outside its scope. The target is zero, and every instance gets a fix.

Measure a baseline from your phone system before go-live so the comparison is honest. The NIST AI Risk Management Framework asks for the same discipline — define, measure, and manage the risks of any AI system that faces customers.

What are the common failure modes?

• Booking into a calendar nobody trusts. If the team keeps a second, real schedule on a whiteboard, the agent will double-book. Fix the calendar first.
• An intent map written from memory. Owners underestimate how many calls are vendors, wrong numbers, and existing customers checking on a job. Listen to real calls.
• Dead-end escalations. Callback tasks with no owner. The caller was answered, then ignored — worse than voicemail.
• Scope creep after launch. "Can it also take payments?" Every new capability needs its own boundaries and review, not a prompt edit on a Friday.
• No one reviewing transcripts. Agents drift as your services, prices, and staff change. Weekly review in the first month, monthly after that.

The AI receptionist pairs naturally with Missed-Call Rescue for any call that still slips through, and the cost ranges for both are laid out in How Much Does an AI Agent Cost for a $5M–$50M Business?.

Where should you start?

Pull last month's call log and listen to twenty recorded calls. Count how many could have been answered and booked by someone following a written script, and how many needed real judgment. That ratio tells you how much of the front desk an AI receptionist can own. If you want help reading it, the HI into AI Assessment ranks the AI leverage points in your business and tells you whether the phone line is the right place to start.