How to Customize an AI Receptionist Without Writing a Novel

Write the transfer first, keep the prompt short enough to test, and prove three calls from an outside line before any customer uses the number.

Published 2026-09-12

Customizing an AI receptionist is not picking a voice and writing a cheerful greeting. It is writing the rules a competent front-desk person already follows, in an order you can test. The businesses that get a good agent do this in an afternoon. The ones that get a vague one skip the test call and find out from a customer.

This is the self-serve job. If a call center is encoding an existing runbook, that is a designed engagement — different article, different door. What follows is for a business that will write its own instructions.

Do the transfer first

Every useful setup starts with the cases that must reach a person.

Write them as rules, not vibes: emergencies, callers in distress, anything clinical or legal, a complaint that is about to become a refund, a caller who asks for you by name. Then call the number from a phone that is not yours and try to trigger each one.

If the transfer is awkward for you, it is worse for a customer. Fix that path before you write hours and pricing. A system that handles the routine beautifully and fails the escalation is not a receptionist. It is a wall.

What belongs in the instructions

The standing prompt should hold things that are true for every caller:

• Hours, locations, and what you do not do.
• Prices you are willing to say out loud, and the ones you will not estimate.
• Who you serve and who you send elsewhere.
• The exact sentences for the five questions you already know you get.
• When to stop and conference you in.

It should not hold Mrs. Chen's dog's name, last Tuesday's special case, or a paragraph about brand personality. Those either belong on a contact record or they belong in the trash. Long prompts hide the one line that is wrong.

What belongs on a contact

Names, organizations, and the background that matters when that number calls. Import them from CSV or JSON. A known caller should not have to re-introduce themselves; an unknown caller should not inherit someone else's notes.

Caller memory — the agent remembering facts from prior conversations with that person — is off unless you turn it on for both the identity and the contact. Leave it off until you have a reason. Memory that is on by default becomes a privacy setting you did not mean to publish.

Guardrails that are not the prompt

Three controls sit next to the instructions and matter as much:

1. Handoff numbers. The verified phones Terri is allowed to conference into a live call. If this list is empty, there is no person at the end of the escalation you just wrote.
2. Outbound prefixes. The E.164 prefixes the agent may dial. The default is US and Canada. Narrow it. An outbound-capable receptionist with a wide prefix list is a liability.
3. Recording and retention. Default is transcripts and metadata, no audio, kept 90 days. Audio is a consent decision. Retention is a jurisdiction decision. Set both on purpose.

The three test calls

After the first draft, place three calls from an outside line:

1. Covered. Ask something the instructions answer. The reply should match the wording you wrote, not a paraphrase that changes a price.
2. Uncovered. Ask something you did not write down. The acceptable answer is "I do not know; I can take a message or transfer you." A confident guess is a fail.
3. Transfer. Trigger an escalation. You should end up on the call with the context, not a cold "someone is asking for you."

Read the transcripts. Add one line for each miss. Do not add a paragraph. Re-run the failing call. When all three pass, point after-hours traffic at the number and keep reading transcripts for a week.

Identities, if you have more than one job

One identity per job. A main line that books and an after-hours line that only takes messages and transfers should not share a prompt. Bind each identity to its number. Leave the unused one alone; a second personality you are not testing is a second way to be wrong.

Voice is optional. The platform voice is the default. Change it only when you have a reason a caller would notice.

What "done" looks like

You are done customizing when a stranger can call, get a correct routine answer, get an honest "I do not know," and reach you when the rules say they should — and you can point to the transcript that proves each one.

You are not done when the greeting sounds nice. That is the part every vendor already has.

Common questions

How do I customize an AI receptionist?

Write the cases that must reach a person, test those transfers from an outside line, then add hours, services, and the wording for common questions. Keep caller-specific notes on contacts, not in the prompt.

What should go in an AI receptionist script?

Rules that apply to every caller: hours, services, prices you will state, who you do not serve, and when to transfer. Not individual customer trivia and not a brand-personality essay.

How do I test an AI receptionist before putting it on my main line?

Three calls from a phone that is not yours: one question the instructions cover, one they do not, and one that should transfer. Read the transcripts, add one line per miss, and re-run the failing call.

When is an AI receptionist actually set up?

When a stranger gets a correct routine answer, an honest 'I do not know' on an uncovered question, and a working transfer when the rules require one — and you have transcripts that show all three.