Buying guide · AI call answering

    Ten questions to ask an AI answering service before you buy

    Written for trades and service businesses. Several of these are questions we would rather you did not ask us cold, which is why they are the ones worth asking. We build these agents, and our own answer to question two is fifteen years in the HVAC trade across customer support, technical support, service, projects and emergency response.

    Updated 2026-08-29

    An AI answering service is easy to demo and hard to run. The demo always works: clean audio, a cooperative caller, a question the system was built to answer. The failure modes show up in month two, at 2am, on the call that was actually worth money.

    These are the questions that surface those failure modes before you sign. They apply to any vendor, including us.

    1. What happens when it does not know the answer?

    The only acceptable answer is: it says so and escalates on a defined path. A system that guesses is worse than a voicemail, because a voicemail does not confidently tell your customer the wrong thing under your brand.

    Ask to see the escalation rules written down. If they do not exist in writing before launch, they do not exist.

    The short versionAsk for the escalation path in writing. Vagueness here predicts everything else.

    2. Who writes the triage rules, and are they from your industry?

    Triage is the whole product. Deciding that "no heat, elderly occupant, minus twenty outside" is an emergency while "noisy fan" is a Tuesday appointment is trade knowledge, not software.

    Ask who wrote the urgency rules and what they did before they wrote them.

    The short versionA generic agent triages by keyword. A useful one triages by consequence.

    3. Can it see live availability, or does it promise a callback?

    "Someone will call you back" is a voicemail with extra steps. The commercial value of an answering agent is almost entirely in converting the call into a booked slot while the customer is still on the phone.

    That requires a real, two-way integration with your scheduling or dispatch system, which is an integration project rather than a setting.

    The short versionIf it cannot book, it is deflecting calls rather than capturing jobs.

    4. What does it do with a safety issue?

    Gas, carbon monoxide, an electrical burning smell, water where it should not be. The correct behaviour is a scripted safety response and an immediate human escalation, never a diagnosis.

    In regulated trades this is not a service-quality question. It is a liability question, and the answer should already be written into the contract.

    The short versionIt should never attempt a diagnosis on anything safety-related.

    5. Which of my systems does it actually integrate with?

    Ask for names, and ask whether that integration has been built before or would be built for you. Those are very different risks and very different timelines.

    Ask specifically what happens when the third-party API is down at 2am: a good integration has error handling, logging and alerting, so a failure is visible immediately rather than discovered weeks later as missing jobs.

    The short versionAsk what happens when the integration fails, not whether it can be built.

    6. Can I read the transcripts?

    You should be able to read every conversation the system had with your customers, without asking. If transcripts are unavailable, summarised, or behind a support request, you cannot audit what is being said in your name.

    Transcript review is also how the system gets better. A vendor that will not show you them is a vendor who is not tuning it.

    The short versionNo transcripts, no deal. This is the audit trail for your own brand.

    7. What happens to my data, and who owns the configuration?

    Ask where customer data is stored, under whose accounts the system runs, and what you receive if you leave: the configuration, the knowledge base, the call records.

    Plenty of vendors will hand you nothing, which turns a service problem into a migration problem later.

    The short versionOwnership on exit is much cheaper to negotiate before you sign than after.

    8. How is it priced as we grow?

    Per-minute, per-call, per-seat and flat-rate pricing behave very differently in a seasonal trade. A per-call model is comfortable in October and painful during the first cold snap, which is precisely when the system earns its keep.

    Model the cost against your busiest week, not your average one.

    The short versionPrice it against your peak week. That is the week it exists for.

    9. What does it sound like to a customer who is annoyed?

    Every demo uses a patient caller. Ask to hear a recording of an impatient one, or an accent the system was not trained on, or someone who interrupts.

    Ask what happens when a caller says "I want a human". The correct answer is that it hands over immediately, without a negotiation.

    The short version"Let me speak to someone" should end the automation instantly.

    10. Will it replace my office staff?

    If the vendor says yes, they are selling to the wrong problem and will disappoint you. These systems cover the hours nobody is at the desk, absorb the seasonal spike, and take the repetitive questions off a team so it can handle the work that needs judgement.

    A vendor pitching staff replacement is describing a product they have not operated.

    The short versionThe honest pitch is coverage and triage, not headcount reduction.

    If a vendor answers all ten well, they have probably run one of these in production. If they answer the pricing question well and the safety question badly, they are a software company that found a vertical. Our own answers to all ten are on the AI call answering page, and where the answer is "we have not built that integration yet", it says so.

    Frequently asked questions

    How much does an AI answering service cost?
    Pricing models vary enough that the shape matters more than the number: per-minute, per-call, per-seat and flat-rate behave very differently in a seasonal business. Off-the-shelf products are subscriptions, commonly advertised anywhere from a few tens to several hundred a month depending on volume and features, while a custom build carries a one-off cost and then a flat running fee. For reference, and so this page is not the only one on the subject without a figure in it, our own agents start at US$3,500 to build plus US$300 a month. Model any quote against your busiest week rather than your average one, because peak season is when the system does the work you bought it for.
    Will customers know they are talking to an AI?
    Many will, and that matters less than most buyers expect. The measure that decides whether it was worth buying is whether the caller got a useful answer and a booked appointment instead of a voicemail.
    What is the difference between an AI answering service and a chatbot?
    A chatbot answers questions. An answering agent triages urgency, captures the diagnostic detail a technician needs, and books into live availability. The second one requires integration with your scheduling system, which is where most of the engineering sits, and most of the value with it.