What an AI Voice Agent Should and Should Not Answer for a Small Business
A practical boundary list for owner-operators: what the phone agent can say, what it must hand off, and how to keep it from inventing answers.
TL;DR
An AI voice agent should handle public, repeatable facts and simple intake: hours, location, services, booking, and status it can look up. It should not give licensed advice, quote a job it has not inspected, share private account details without proving who is calling, or stay on the line for emergencies, threats, or complaints that need an owner. Train short refusals and a working human handoff. A wrong, confident answer costs more than a transfer.
An AI voice agent should answer public, repeatable facts: hours, location, services, booking windows, and status checks it can look up. It should not give licensed advice, invent a price, discuss private account details without proving who is calling, or stay on the line when the caller is angry, hurt, or asking the business to commit to something a person has not approved. How natural the voice sounds is the wrong test. Ask whether you would let a new hire say the same words on day one, with no manager in the room.
What should an AI voice agent actually answer?
Give it the questions already on your website, voicemail greeting, and the one-page sheet by the phone.
Hours, holiday closures, address, parking, service area, and how to get on the calendar. Those answers are the same for every caller. When they change, you update one source and the agent reads from that. That is AI integration in practice: the voice is only as current as the calendar and the hours file behind it.
It can take a name, a callback number, a short description of the work, and a preferred time window. Intake, not judgment. A plumber's agent can ask if there is standing water. A salon agent can ask if the visit is a cut or a color. Neither question is a diagnosis. Neither one locks you into a price.
Connect it to live availability and it can offer open slots and book them. Booking is clean because the system either has the slot or it does not. Nothing to invent.
Published policies belong here if you would stand behind them in writing: cancellation windows, deposit rules, service area limits, what you do not offer. "We do not work on commercial roofs" is a boundary. "We can do that cheaper than the last quote you got" is a negotiation. Only the first is an agent line.
Status questions work when two things are true. The agent can look the record up, and the caller has already proven who they are. "Your appointment is Tuesday afternoon" is a lookup. "The part is in, and we can schedule the install" is a lookup. If the lookup fails, say so and offer a person. Guessing a status is how you double-book a slot, or leave a caller thinking a job is done when it is not.
Simple routing is in bounds. "I can take a message for the owner." "I can put you through to billing." "I can text you the new-patient form." Define those actions in advance. They do not ask the model to decide what the business will promise.
What should it refuse to answer?
Anything that needs a license, an inspection, or a signature.
Do not let it give medical, legal, tax, or financial advice. A dental office agent can confirm the office is open and that pain is a reason to come in. It cannot say which tooth is the problem or whether a medication is needed. A bookkeeping firm can confirm a return was filed. It cannot tell a caller how to treat a specific deduction.
Do not let it quote a firm price when the real number depends on what someone sees on site. A published diagnostic fee or a menu item with no variables can be read aloud. "This is what it will cost to fix your unit" is a commitment. Wrong quotes turn into fights at the door and on review sites. If pricing is not a menu, collect details and promise a human estimate.
Do not let it recite account balances, medical notes, Social Security numbers, or home addresses until it has authenticated the caller. A booking code, an account PIN, or the last four of a phone number already on file is a start. Sounding like the customer is not identity. If authentication fails, take a message. Do not keep guessing until something matches.
Do not let it collect a full payment card number unless you have a payment process built for that channel and you know how card data is stored. A note that the caller wants to pay is fine. Digits in a call transcript are not. If you need payments tied into the same systems that run the rest of the shop, that is a custom software problem, not a prompt-writing problem.
Do not let it handle a caller who says someone is hurt, a pipe is flooding the house, or there is a fire. Those calls go to a person, or the agent tells the caller to hang up and contact emergency services. An agent that keeps chatting through an emergency is not saving anyone time.
Do not let it negotiate, waive fees, match a competitor, or "make an exception just this once." Those are owner decisions. If the caller asks the agent to agree to something off-script, transfer. Do not invent a yes.
Complaints that need an apology from a person should not stay with the machine. The agent can log the issue, repeat back the facts, and move the call. Trying to fix a bad experience in real time often adds a second grievance: the caller wanted a human and got a script.
Employee and vendor issues are out. Payroll questions, "is so-and-so working today," and disputes with a contractor are not public facts. Take a name and a callback number. Do not confirm who is on the job.
Want this working on your numbers?
Viewmedia makes marketing you can prove, matched to real, closed sales.
How do you draw the line without a 40-page policy?
Use three questions on every topic you might add.
Is this answer the same for every customer, or does it change by job, by person, or by mood? Same for everyone: in bounds, as long as the source is current. Changes by job or by person: human, or a lookup after authentication. Never a best guess.
If the answer is wrong, do you lose money, a license, or a customer you cannot replace? If yes, the agent may collect facts. It may not decide. It may not tell the caller "it will be fine."
Would you let a new employee say this on a recorded line with no one listening? If you hesitate, it is not an agent question. That test is stricter than "could the model probably get it right," and it should be. You are on the hook for the words, not the model vendor.
Write the out-of-bounds list as short refusal lines, not speeches. "I cannot quote that without a visit. I can take your details and have someone call you." "I cannot give medical advice. If this is urgent, hang up and call emergency services. Otherwise I can get you on the schedule." Short refusals are easier to train than long explanations, and they leave less room for the model to improvise.
Keep a human path that actually works. If "press zero" dumps into a full mailbox, callers will keep pressing the agent. After hours, a recorded promise of a callback by a named window beats a fake conversation that cannot dispatch anyone.
Review the list when you add a service, change a price, or hire a specialist. The boundary is operations, not a one-time setup.
How should the agent talk when it does not know?
It should say it does not know.
Silence is not the expensive failure. A confident wrong answer is: a booked slot that does not exist, a "yes we do that" for a service you dropped, a price nobody approved. Instruct the agent to stop, admit the gap, and transfer or take a message. Put that instruction next to your hours, not in a slide deck you read once.
Log every handoff and every refusal. The log shows new questions worth adding, and questions that should stay off limits. If callers keep asking for a phone quote on a job you never price without a visit, that is a training note. It is not a reason to let the model guess.
Do not hide that the caller is speaking with an automated assistant. Say it at the start. That sets the right expectation, and it makes the handoff to a person feel normal instead of like a trick. If a caller asks for a human, give them a human, or a clear time when a human will call back. Arguing with that request teaches people to work around you.
Repeat back the facts you captured before you hang up: name, number, reason, time window. Intake only helps if the next person can act on it. A transcript full of polite filler and a wrong callback number is worse than a short voicemail.
The voice agent is a front door with a script, a lookup, and a stop rule. Used that way, it can take the routine calls that used to eat your morning. Used without a stop rule, it becomes an employee you cannot fire for making things up. Draw the line in writing, train the refusals, and keep a person on the other side of the transfer.
FAQ
Can the agent take a deposit or a card over the phone? Only if you already have a payment process meant for that channel and you know where the digits go. Otherwise offer a secure link, a callback from billing, or an in-person payment. Wanting to pay is a message worth taking. A card number in a transcript is not.
Should it tell people they are not speaking to a person? Yes. Say it at the start, in plain language. Callers who know they are in an automated flow tend to ask cleaner questions and accept a handoff. Callers who find out later may feel misled, even if the answers were right.
What if the same question is both a fact and a judgment call? Split it. "Do you repair water heaters?" is a service-list fact. "Should I replace mine today?" is a judgment call. The agent can answer the first and book the second as a visit. Do not let one question drag the other into scope.
Can it handle after-hours calls without a person on backup? It can take messages, read hours, and book slots that already exist on the calendar. It cannot dispatch a crew it cannot reach, and it cannot promise a same-night arrival you have not staffed. If you offer true after-hours emergency service, a person still owns the yes.
Founder, Viewmedia
Brian Wroblewski is the founder of Viewmedia. For more than two decades he has helped local and regional businesses turn marketing spend into provable, closed sales.
