What information an AI receptionist needs before it goes live

Gathering this is not vendor paperwork. It is the same exercise as writing a training manual for a new hire who will never ask a follow up question, never read the room, and never quietly decide that this particular caller is an exception.

Budget half a day. The businesses that spend it launch once. The ones that do not spend it launch three times.

What does the agent need to know about your business?

The boring facts, written down in one place, in the form a caller would ask for them.

Hours, including holidays and the difference between when you answer and when you are open. Address, cross streets, parking, and which entrance. Service area, described the way locals describe it, not as a zip code list. Which services you offer and, more usefully, which you do not.

That last item saves more calls than anything else on the list. A plumber who does not do septic wants an agent that says so in five seconds and offers a referral, not one that books a visit somebody has to cancel.

Add the twenty questions your front desk answers most often, with the exact answer you want spoken. The single best input to any AI receptionist buildout is a list of real questions from real callers, written in the words those callers used. Have whoever answers the phone keep that list for a week before you start.

Include how to pronounce your business name, your neighborhood, and any provider names. Callers notice when a system mangles the name of the town.

Why can you not just buy a call intent list for your industry?

Because for most industries nobody has published one, and the pie chart in your vendor's deck is decoration.

There is exactly one solid public taxonomy of what people call local businesses about. Maple published it from 1.2 million calls across more than 1,000 US restaurant and local business locations: reservations 38%, hours and location 22%, menu questions 15%, takeout and delivery orders 12%, event and large party inquiries 7%, modifications 4%, everything else 2%. That is a real dataset with a disclosed sample, and it is about restaurants.

For med spas, no published call composition dataset exists anywhere. Boulevard, Zenoti and Mindbody all sit on the data and none of them publish a call intent taxonomy. Every page claiming one either extrapolates from industry spend figures or simply asserts percentages. If a vendor shows you a med spa call mix chart, ask for the sample size and the method, and watch what happens.

Dental is only slightly better. Patient Prism's 11,552,668 call dataset is the best in any vertical, and it publishes why patients did not book, not why they called. Different question.

"Nobody has measured this, so let us measure yours" is a stronger position than a borrowed chart. Five days of your own calls beats an industry average that does not exist.

What does it need to know about your calendar and pricing?

Rules, not preferences.

For the calendar: appointment types and durations, buffer time, how far ahead it may book, how far out it may book, blackout periods, who can be double booked and who cannot, and what happens when the requested slot is gone.

For pricing: exactly what the agent may state. A published range is usually fine. A firm number for unseen work is not, because the caller will hold you to it and your technician will eat the difference.

Write your pricing policy as three lists. What it may quote as a fixed price, what it may quote as a range, and what it must never quote at all. Anything not on a list defaults to the third.

The money question is more load bearing than most owners expect. Patient Prism categorized 1,163,398 dental booking opportunities and found the reasons patients did not book cluster as consideration and timing 45.2%, financial barrier 34.4%, service fit 10.3%, and connection failure 9.3%. Inside that financial cluster, insurance outweighs price by more than eleven to one.

Read that twice. The single most common hard question your agent will face is the one it is least allowed to answer authoritatively. A dental agent can confirm which plans you are in network with and collect a plan name and member ID. Telling a caller that a crown is covered, or what they will owe, is a different act entirely. Write that boundary before you write the greeting.

Then the routing rules for emergencies, which are their own category. Same day, after hours, and genuine urgency each need a defined path, and the paths are usually different.

What does it need to know about what it must not do?

This is the document nobody writes and everybody needs.

It contains the never say list, the specific replacement wording for each item, and the trigger words that force an immediate handoff regardless of context.

For a dental or medical practice, the boundary has a published source. The Doctors Company, the largest physician owned malpractice insurer in the US, states in its telephone triage guidance that unlicensed personnel should never be allowed to triage by telephone or provide telephone advice, and that failure to follow written advice protocols may be treated as practicing beyond scope. Its closed claims analysis attributes over 30% of adverse patient events in the office setting to miscommunication.

The same guidance contains the distinction that makes an AI agent viable at all. Don Balasa, chief executive and legal counsel of the American Association of Medical Assistants, separates triage, which requires independent clinical judgment and may not be delegated, from non triage communication, which is the verbatim receiving and verbatim conveying of information under provider approved protocols and may be. An AI agent may operate a provider approved decision tree. The moment it generates novel clinical language, which is what a language model does by default, it has crossed the line.

California adds two statutes for anything health care adjacent. AB 3030, effective 1 January 2025, requires a prominent AI disclaimer and instructions for contacting a human on generative patient communications, and its text says written or verbal, so it reaches voice. AB 489, in force since 1 January 2026, prohibits a persona that implies a licensed professional and makes each use of a prohibited term a separate violation. That reaches names, greetings, SMS signatures and your marketing pages. Put both in front of your own attorney.

For a restaurant, an agent must never answer whether a dish is safe for an allergy. The FDA Food Code puts allergen knowledge with the designated person in charge and satisfies its notification duty through written notice, which a voice agent is not.

Every business has a version of this even without a license involved. Guarantees, arrival promises, warranty interpretations, and anything a caller could reasonably treat as a commitment.

Alongside it goes the compliance section. Whether calls are recorded, where the consent notice sits, what is retained and for how long, and what happens if a caller declines to be recorded. California requires the consent of all parties before recording begins, so that notice belongs in the first line of the script rather than the fourth. Have your own counsel review the wording rather than accepting a vendor template.

Get the never say list signed by whoever carries the license or the liability. Not the vendor and not the office manager.

How do you know when the document is finished?

Where does the lead go after the call ends?

Decide this before launch, because a perfectly handled call that lands nowhere is still a lost job.

Name the destination system, the exact fields it needs, and the person responsible for acting on each outcome. Booked appointments, callback requests, escalations, and spam should each have a defined path.

Then decide the notification and the response time you are committing to, because the agent will be promising it to callers on your behalf.

One caution on the callback path. If your plan is for an automated voice to ring people back rather than a human, that is outbound calling with an artificial voice, which the FCC has held requires the prior express consent of the called party. It is a different legal object from answering an inbound call and it belongs in the same conversation with counsel as the recording notice.

The phone is one channel of several, and leads leak at the seams between them. Routing leads from four channels without dropping any is the wider version of this problem, and it is worth solving once rather than per channel.

The escalation rules deserve their own attention too. Who takes a transfer, at what hours, on what number, and what happens when they do not pick up. How good implementations handle escalation to a human is the difference between a system that protects the relationship and one that strands people.

What to do this week

Ask whoever answers the phone to log every question they get for five business days. Exact wording. That log is your script.

Run the five step coverage arithmetic and stop at the top twenty.

Write the three pricing lists and the calendar rules on one page each.

Draft the never say list with replacement wording and get it signed by whoever holds the liability.

Write the routing table: outcome, destination, owner, response time. Four columns, one row per outcome.

Then, and only then, hand it to a vendor. Testing an AI phone agent before it takes real calls comes next, and it goes much faster when the documents exist.

While you are gathering business facts, the same material usually improves the website. How licensing and insurance are presented converts, and it is often sitting in this intake document already. Page speed and what it does to conversion is a separate question with more myth attached than most.

Questions about your intake documents?

Email me at eric@seod.com with the list of what you have gathered so far for your AI buildout, even if it is three bullet points. I will send back exactly what is missing, in the order I would collect it, and flag which items need a signature rather than a decision.

I answer these personally and it takes about fifteen minutes. If your vendor has already sent you an intake form, send that too and I will tell you what it is not asking for.

The rest of the AI phone and lead response library covers what happens around it.