The phone rings. Somebody has to pick it up, and the person who normally does is halfway through something else. That is the problem an AI phone agent is for.
Before you buy one, it is worth knowing exactly what it is good at and where it should be stopped from going anywhere near. Here is that list, written the way we would explain it on a call.
What it actually is
An AI phone agent answers your number, listens to a person talking normally, and replies in about a second. It is not a press-one-for-reservations menu. It is not the chat bubble on your website. When it works, the caller says what they want in their own words and gets a real answer, a booking, or a person.
It can also work the other way round and make calls: appointment reminders, confirmations, a follow-up after a no-show.
The calls it handles well
The same five questions, all day. Opening hours. Where you are, because the map pin is in the wrong place. Parking. Whether you are open on the public holiday. What to bring. Whether you take card. None of these need judgement, the answer never changes, and your team already recites them from memory.
Calls that arrive at the worst moment. The phone rings while the one person on the desk is checking someone in, taking a payment, or calming a waiting room. That is not really a staffing problem. It is two things happening at the same time, and one of them always loses.
Calls after you close. People ring when they are free, and free usually means evening. A line that goes to voicemail at six sends that caller to whoever answers next.
Routine bookings. A table, a room, a check-up, a viewing. This one comes with a condition: the agent has to be able to reach your booking system, read what is genuinely available, and write the booking back in. Without that connection you have bought an expensive answering machine that promises slots you may not have.
A caller speaking a language nobody on shift speaks. The agent answers in whatever language the caller opens with. For a hotel or a clinic with international customers this is often the single most useful thing it does.
A message taken properly. Name, number, what it is about, how urgent, written down, in the right place. Compare that with a voicemail nobody has played back since Tuesday.
Outbound reminders. Ringing tomorrow's appointments to confirm, with an easy way to move the time. Dull, repetitive, and exactly the sort of task that quietly stops getting done when the office is busy.
The calls it must never take
The rule we build to: the agent never guesses, and it never handles anything where being wrong costs money, health or trust.
Complaints. Somebody who is angry wants a human being who can do something about it. Handing them a machine confirms their suspicion that you do not care. Complaints go to a person immediately, with a summary of what was already said so nobody has to start again.
Anything about money. A refund, a disputed charge, a discount, a deposit that should have been released, a price that is not on your published list. The agent can gather the details and prepare everything. A person decides.
Anything clinical. Symptoms, results, medication, whether someone should come in today. There is no clever prompt that makes this acceptable. The agent handles the logistics around care -- times, directions, what to bring, moving an appointment -- and nothing inside it. We build this as a hard wall on every clinic project.
Anything contractual. Cancellation terms in a dispute, a promise about a delivery date, anything a customer could reasonably treat as your word.
Emergencies and anything about somebody's safety. These need to reach a human on the first sentence, without a menu in the way.
Anything where the answer is not written down. If two members of your team would answer a question differently, the agent should not have an answer either. It should hand over. Most bad agent behaviour traces back to a question the business never actually decided on.
The big, unusual enquiry. A wedding, a fifty-cover function, a corporate rate for a year. The instinct is to let the agent take these because they are worth the most. They are worth the most, which is precisely why they should reach the person who can win them.
Tell callers it is a machine, in the first sentence
Owners worry about this more than callers do. Hiding it backfires for three reasons.
People work it out. Not always at once, but they get there, and the moment they do they stop trusting everything the agent said before that point -- including the price and the availability. A small disclosure at the start costs you nothing. Being caught costs you the call.
It also makes the agent work better. When someone knows they are talking to software, they speak in short, clear sentences: "I want a table for four on Friday at eight." When they think it is a person, they ramble, interrupt and use hints. Honesty improves the accuracy.
And in a growing number of places, telling people is the law. The rules differ by country and are still moving, so check what applies where you operate -- but building on the assumption that you must disclose is the safer bet, because you never have to unpick it later.
Good disclosure sounds like this:
Good evening, you are through to the front desk. I am the automated assistant. I can check availability, take a booking, or put you through to someone. What can I do for you?
One sentence, no apology, and a way out offered before the caller has to ask for it.
Then honour it. When someone says they want a person, the next thing that happens is a person -- not a cheerful "I can help with that!" and another attempt. An agent that argues about being handed over is worse than no agent.
The handover is the real product
Most of the value is not in the calls the agent answers. It is in what happens to the calls it cannot.
A good handover rings a real extension, and when someone picks up, they already have a written summary of the call so far. A bad handover dumps the caller back at the start with a different voice asking the same questions.
Decide in advance what happens when nobody is free. The honest option is that the agent says so, takes a number, gives a specific time for the callback, and creates the task where somebody will see it. What it must not do is loop, or promise a callback that lands nowhere.
How to judge whether yours is any good
Ignore the demo. Demos are recorded by people who know exactly what to say.
Listen to twenty real calls yourself, end to end, including the awkward ones. You are checking four things: did it answer immediately, did it get the booking right, how quickly did it give up when it was out of its depth, and did anyone have to repeat themselves after a handover.
Then take three numbers from your own call log, which your phone provider can give you: how many calls went unanswered before and after, how many calls ended in a handover, and how many people rang back within the hour. That last one is the honest measure. A repeat call usually means the first one did not actually solve anything.
When we would tell you not to bother
If you already answer nearly every call. Get a month of call records from your phone provider first and count the ones that rang out and the ones that came in after hours. If that number is small, an agent is a solution to a problem you do not have.
If almost every call is a negotiation, a complaint, or a conversation that needs somebody who knows the customer. A specialist firm with thirty valuable clients does not need this. A hotel taking two hundred routine calls a week does.
If nobody in your business can write down the answers. The agent inherits whatever confusion exists. Fixing that is worth doing anyway, and it is cheaper than software.
The aim is not an agent that passes for human. It is one that is fast, says plainly what it is, and knows the edge of its own knowledge well enough to get out of the way.
Filed under
- ai voice
- phone
- customer service
- automation limits