Decide what the phone agent is allowed to quote

The first question a lot of callers ask a service business is some version of "how much?" How much to unclog a drain. How much for a cleaning. How much to come out on a Saturday. When an AI receptionist answers that line, it will get this question constantly, and it will answer it one way or another. If nobody decided ahead of time what it is allowed to say, the agent will improvise, and improvised pricing is one of the fastest ways to start a fight between a customer and the person who shows up to do the work.

We treat pricing as its own part of every voice reception build. Before the agent goes live, we sit down with the owner and write out, in plain sentences, what the agent can quote, what it can describe as a range, and what it has to hand off. It is usually a short conversation. Skipping it is what causes trouble.

Why improvised quotes go wrong

A language model is good at sounding helpful. Ask it what a service costs and it wants to give you a number. If the number is not in its instructions, it may pull one from a pricing page it was given months ago, from an old promotion in the knowledge base, or from a general sense of what that kind of job usually costs. All three are bad sources.

The pricing page might be out of date. The promotion might have ended. The general sense of the market has nothing to do with this owner's rates, their service area, or what they charge after hours. And the caller does not know any of that. They heard a number from the business, so to them it is the business's number.

Then the tech arrives, looks at the job, and gives a real price. If it is higher than what the phone said, the customer feels misled before any work starts. The tech now has to argue about something they never said. The owner either eats the difference or loses the job and maybe gets a bad review on top of it. None of this shows up as an error in any log. The call looked fine.

There is a quieter version too. The agent avoids numbers completely and says "someone will call you with pricing" to every question. That is safe, but it frustrates callers who wanted a simple answer to a simple question, like whether there is a fee just to come out. Some of them hang up and call the next shop on the list. So the goal is not to block pricing talk. The goal is to decide exactly which answers are safe and give the agent those answers word for word.

Sort every price into three buckets

When we work through pricing with an owner, we sort what callers ask about into three groups.

Fixed prices the agent can state. These are charges that do not depend on what the tech finds. A diagnostic or trip fee is the most common example. A standard cleaning, a flat rate inspection, or a membership plan price can fit here too. If the owner is comfortable saying the number to anyone who asks, it goes in this bucket, and the agent gets the exact figure and the exact wording.

Ranges the agent can describe with a condition. Some jobs have a typical range that the owner is willing to share, as long as it comes with a clear caveat. "Most water heater replacements we do fall in a certain range, but the tech will confirm the price after seeing the unit and the space." The agent says the range and the condition together, every time. If the owner does not want a range said out loud, the job moves to the third bucket.

Things only a person can price. Anything that depends on inspection, parts, access, or judgment goes here. So do discounts, price matching, financing terms, insurance questions, and any request to negotiate. For these, the agent explains that pricing depends on the job, offers to book the visit or take a message, and says who will follow up and roughly when.

Writing these down has a side benefit. Owners often find that their own staff give different answers to the same question. Agreeing on what the phone agent says tends to fix what the front desk says too.

A few rules we apply across all three buckets:

• The agent never does math. It does not add fees, apply discounts, or estimate a total for multiple items. Those are easy places for a model to be confidently wrong.
• The agent never promises a price will be honored later. It can say what the current fee is. It does not say "we'll lock that in for you."
• After-hours, weekend, and emergency pricing are written out separately, because those are where surprises cost the most.
• Old promotions get removed from anything the agent can read. If a deal is live, it gets added on purpose, with an end date.
• If the caller pushes for a number the agent is not allowed to give, the agent holds the line politely and offers the handoff. It does not guess to make the caller happy.

Keep the price list in one place and test it

Prices change. Fees go up at the start of a year, a new service gets added, a promotion runs for a month. The pricing answers need to live somewhere the owner or office manager can update without asking a developer to edit a prompt.

In our builds, that usually means a short, structured price list the agent reads from at the start of each call. Each line has the service, the approved wording, the bucket it belongs to, and when it was last confirmed. When the office changes a fee, they change one line, and the next call uses it. Nobody has to remember that the agent was told about the old fee somewhere else.

We also test pricing the same way we test any other part of the agent. Before launch, and after any price change, we run a set of calls through it: a caller asking about the trip fee, a caller asking for a ballpark on a big job, a caller asking for a discount, a caller asking about weekend rates, a caller who keeps pushing for a number. We check that each answer matches the approved wording and that the agent hands off when it should. This takes less than an hour and it catches the problems that would otherwise surface on a real customer's doorstep.

Finally, we log what the agent said about price on each call, right in the call summary the team gets. If a customer later says "your phone told me it would be less," the office can look up the exact words instead of arguing from memory. Most of the time the log shows the agent said the right thing with the right caveat, and that settles it quickly. When it shows a mistake, it points straight at the line in the price list that needs fixing.

Start with the question callers ask most

If you run an AI receptionist today, or you are about to, pull a handful of recent call transcripts and search for the word "cost" or "price." Read what the agent actually said. If any of those answers would make your techs wince, that is the place to start. Sort the questions into the three buckets, give the agent exact wording for the first two, and make the handoff for the third one clean.

It is a small piece of work compared with the rest of a voice build. It also decides whether the first conversation a customer has with your crew starts with trust or with an argument about a number the phone made up.