Questions to Ask Before You Buy an AI Voice Agent

Every vendor selling you an AI phone agent has the same demo, the same three slides, and roughly the same promise. None of that tells you which one will still be answering your phone correctly in March.

What separates them is a short list of questions most buyers never ask, because the questions sound boring. This is that list. For each one, what a good answer sounds like and what a bad one sounds like. Use it on us too.

One thing worth settling first: if you are not yet clear on what an agent actually does in the ninety seconds a call takes, read one waterproofing call walked through step by step before you start evaluating anyone. Buying before you know the mechanics is how people end up comparing price only.

Can I call it right now, before I sign anything

This is the first question and it filters more vendors than the other nine combined.

A good answer is a phone number, said out loud, on the spot. You hang up, you call it, you hear the thing you are being sold. Ours is (844) 493-2745 and it answers on the first ring, which is the only reason it belongs in this article. Any vendor should be able to do the same.

A bad answer is a recorded demo video, a scheduled “demo call” where a salesperson plays you audio, or a sandbox agent trained on a fictional plumbing company. A recording proves nothing. It is one take out of however many they needed. If the product genuinely works, listening to it costs them nothing, so a refusal is information.

When you do call, interrupt it. Ask a question out of order. Ask what a bowing wall costs. Cough over its second sentence. The demo script survives a polite caller. Your callers are not polite.

Who trains it on my services, area and pricing

The agent is only as good as what it knows about your business, and somebody has to put that in.

A good answer names the person and the process. They interview you, they write the knowledge base, they test it against your real calls, and they own it after launch. When you drop a service, add three towns, or raise the price of a full interior drain tile system, you send one message and they change it. Ask how long that change takes and get a number in hours or days.

A bad answer is “you’ll set that up in the dashboard.” That is not a service, that is software with a subscription. You are on a jobsite. The dashboard will be edited once during onboarding and never again, and eleven months later the agent is still quoting last year’s crack injection range and booking inspections in a town you stopped driving to.

What does it do when it does not know the answer

Every agent hits questions it cannot answer. The difference is what happens in that second.

A good answer describes a specific fallback, and the vendor can show it. It says it does not have that information, offers to have you call back, and creates the callback with the question attached. Better, it transfers. Best, the vendor tells you which categories of question it is deliberately built to fail on.

A bad answer is “our model handles that.” Models fill gaps. That is what they do. An agent with no defined fallback will tell a homeowner her insurance covers groundwater seepage, and you will find out when she repeats it back to you in her basement.

Can I read transcripts of real calls it handled for another contractor

Not a case study. Transcripts.

A good answer is a set of redacted transcripts from a business in your trade or close to it, including calls that went badly. A vendor confident in the product will show you a call where the agent lost the thread, because they want you to see the handoff working.

A bad answer is a PDF with a percentage on it, or “we can’t share those, privacy.” Privacy is a real constraint and redaction is a solved problem. If nobody will show you a single unpolished transcript, assume the polished ones are the only ones that exist.

Read for two things: whether the caller sounds like your callers, and whether the agent’s questions would have changed what you loaded on the truck.

What happens to a call it cannot handle

A missed call routed into an apology is still a missed call.

A good answer distinguishes between a warm transfer and a message. The agent tries your cell, tries the second number, and if nobody picks up it takes the details, texts you immediately, and tells the caller when to expect you. You set the triggers: an adjuster on the line, a commercial building, someone saying a wall has moved, active water.

A bad answer is “it takes a message.” That is a machine doing what a human answering service already does, for the same outcome. If you are weighing the two against each other, we walked both through the same flooded-basement call.

Who owns the phone number and the call data if I leave

Ask this before you ask about price. It decides what your options are later.

A good answer: the number is yours, it ports out on request, and they will say how long porting takes. Recordings, transcripts and contact details export in a format you can open. No fee attached to leaving.

A bad answer is any version of “the number is provisioned on our platform.” A number you cannot port is a number on your truck, your yard signs and your Google profile that belongs to someone else. That is not a vendor relationship, that is a hostage situation with a monthly invoice. Same with call data. Two years of transcripts is a record of what your market asks for, and it should leave with you.

What is the contract, and what does leaving look like

A good answer is a short initial term or none, monthly after that, and a stated notice period. Ask directly what happens on day one after you cancel: does the number forward back to your line automatically, or does your phone go dark while someone processes a ticket?

A bad answer is twelve months locked, an early termination fee, or a setup fee large enough that leaving feels wasteful. Long contracts in this category exist because the product’s value is hardest to see in month two, when the novelty is gone and you are counting booked jobs.

What happens when it breaks at 2am

It will break. Every phone system does, and yours breaks on the Sunday it rains.

A good answer has a failover that does not depend on anyone answering a support ticket: if the agent does not pick up within a set number of rings, the call falls through to your cell or your existing line automatically. Ask whether that is on by default. Ask for the emergency contact and whether it reaches a person.

A bad answer is a support email address and business hours. Ask what their last outage was and how long it lasted. A vendor who says they have never had one is either new or not telling you.

How is it priced, and what is not included

A good answer is one number and a clear boundary. Whether it is per minute, per call, per month, what happens if February is busy, and specifically what costs extra: telephony charges, extra numbers, spam calls that get answered, changes to the script, additional agents for a second location.

A bad answer is an unlimited plan with a fair-use clause nobody will define, or a price that only appears after two calls with a closer. Ask them to price your actual volume. Pull last month’s call log and give them the real number.

What does it refuse to do

The best signal in the whole conversation. A vendor who has run this on real contractors knows exactly where an agent should stop.

A good answer is a list, without prompting: it will not quote a firm price on a foundation it cannot see, will not say whether a crack is structural, will not comment on insurance coverage, will not discuss a job in progress it has no record of, will not argue with an angry customer. That list is the difference between a tool and a liability, and it is the core of what AI voice agents for contractors have to get right in a trade where one wrong sentence about mold exposure becomes a legal problem.

A bad answer is “it can handle anything you want.” Nothing that answers your phone should be able to handle anything.

The questions that sound smart but tell you nothing

Three of these come up constantly and none of them predict whether the thing works.

• “Which AI model does it use?” The model changes every few months and every serious vendor uses one of the same handful. It tells you nothing about whether the agent knows your service area, your pricing rules, or when to transfer. Those are configuration and process, not model choice.
• “What is your accuracy percentage?” Accuracy at what, measured by whom, on whose calls? There is no shared definition and no audit, so the number is whatever sounded good in the deck. Any figure quoted without a stated methodology is decoration. Replace this question with “show me a transcript of a call that went wrong.”
• “Is it custom built or off the shelf?” Almost everything in this category is built on the same underlying components. What is custom is your knowledge base, your call flow and your escalation rules. Ask about those instead. A vendor who says “fully custom” is usually describing the invoice.

When the honest answer is to buy nothing yet

Three situations where the right move is to close the tab.

Your call volume is genuinely low. If you miss two calls a week and you are back to them within the hour, the math on any monthly system is thin. Fix the forwarding on your existing line first and see what is left.

Nobody in the business can answer the setup questions. If you cannot state your service area in zip codes, your inspection length, and which jobs you refuse, the agent cannot either. That work is yours regardless of who you hire, and it is worth doing before you pay anyone.

You are hoping it will close jobs. It books appointments and collects answers. The homeowner ringing four companies for the cheapest number to seal a cold joint is still yours to win, in the basement, with a flashlight.

Take this list to whoever is selling you one, including us. The vendor who answers all ten without flinching is a different proposition from the one who sends a video. That gap is the entire decision.