An AI receptionist is only as good as the instructions it is given. The same technology can sound like a helpful member of your team or like a phone tree with a personality problem, and the difference is mostly in the script: what it asks, in what order, and what it promises. This is the structure we use when configuring agents for local service businesses.
Open with the business name and one useful sentence
The first eight seconds decide whether a caller relaxes or starts pressing zero. Use the business name, a short offer of help, and nothing else. Something close to: thanks for calling Northside Plumbing, this is the assistant, I can help you get scheduled or get a message to the team. No jokes, no long welcome, no menu of options unless the caller later asks for one.
Test the greeting on your own phone in a noisy room. If you cannot understand it the first time while a truck idles nearby, it is too long. Every extra clause in the opening is a chance for a caller to hang up before they explain what they need.
Ask one question at a time, in priority order
Write down what your best front-desk person asks, in the order they ask it. For most trades that is the problem, the address, the urgency, and a name and number. Four questions, one at a time, with a short acknowledgement between them. Callers will not recite answers to a list, but they will happily answer a single question.
Put urgency early, never last. If the caller has water coming through a ceiling, knowing that in the first twenty seconds changes everything about the call. A script that collects a full name and email before asking whether the situation is an emergency has the priorities backwards.
Capture the details your team cannot work without
A lead with no address is a phone number and a hope. Decide the minimum record for each service: name, callback number, service address or location, a one-line description of the problem, and whether the caller is a new or returning customer. Everything else is optional and can be added later.
Say the phone number back to confirm it. Read it in groups of three or four digits, never as one run of numbers, and ask the caller to correct it if you heard wrong. This single confirmation step removes the most common reason a captured lead cannot be called back.
Be exact about what the agent can promise
The agent must know whether it is booking real appointments or collecting requests. If it writes into a schedule you control, it can offer the times that schedule actually offers and nothing more. If it cannot write to a schedule, it must say a team member will confirm, then send a booking link rather than imply a confirmed visit.
Write the limits into the script explicitly: no price quotes unless you have published a fixed price, no promises about arrival windows, no medical or legal advice, no estimates over the phone. An agent that invents a number to sound helpful creates a problem your team has to solve on site, in front of a customer who was told something else.
Set escalation rules with names attached
Decide what triggers a live transfer or an immediate alert: emergencies, high-value commercial work, an angry caller, or any request the agent cannot handle twice in a row. Give each trigger a destination that a person actually watches, not a shared inbox nobody reads until Thursday.
During hours, transfer to the office. After hours, follow your own policy: an on-call technician for genuine emergencies, a text alert to the owner otherwise. Test each path before launch by making the call yourself, and confirm that the alert arrives with the caller's number and a summary attached.
Review transcripts for the first two weeks
Read every transcript from the first week and skim the second. You are looking for three things: calls where the caller repeated themselves, calls where the agent guessed at an answer, and calls that ended without a name. Each of those is a script edit waiting to happen, and most take one sentence to fix.
After that, review a sample each month and after any change to your services, hours, or pricing. The script is not a document you write once. It is a living description of how your best receptionist behaves, and it improves fastest when someone reads what callers actually said.
Three lines to remove from the script
Most awkward calls trace back to a small set of phrases. The first is a long explanation of what the caller is speaking to. Nobody needs the architecture of the system; they need help with a furnace. Let the agent behave like an assistant rather than a product demonstration.
The second is the vague promise that someone will call back, with no window attached. Say when, even if the answer is tomorrow morning. The third is a string of apologies. One acknowledgement is enough. More than that sounds like the business has already given up, and it invites the caller to hang up and start over somewhere else.
What the agent should never be asked to do
It should not quote a price that is not published and fixed, negotiate a discount, or estimate a job over the phone. Those promises travel straight into the customer's memory and become an argument on site, with a technician who never said them standing in a hallway.
It should never take card details. Payment links exist for a reason, and a caller reading a card number to a scripted system has no way of knowing who is listening. Capture the intent, send the link, and let the payment happen somewhere secure.
It should not try to sell. The job of an answering agent is to capture the details correctly and get the right appointment on the calendar. Pitching to a caller who rang about a leak makes the business sound like a call centre, which is exactly the impression most owners are trying to avoid.
