Choosing an AI Phone Answering Service: What Matters After Hello

You have probably seen the ads. A calm voice picks up: "Hi, how can I help you today?" It sounds almost human, and the demo ends with a booked appointment. If you run a clinic, a pharmacy or any business where the phone rings all day, your honest reaction is likely: "Yes, I've seen those. Then what?" This guide is about choosing an AI phone answering service for a small business by what happens after the hello, because that is where real calls succeed or fall apart.

Phone calls break in ordinary situations, not in rare edge cases. Knowing where they break tells you which questions to ask before you sign.

The greeting is not the product

Every phone assistant sounds good for the first ten seconds. The question that matters is different: what share of callers finish the task with no human? In the industry this is called the completion rate, or the containment rate when a vendor counts calls handled end to end. Ask for the number, and then ask how it is measured. A caller who gave up is not a contained call.

A clinic's callers often ring from a car, a kitchen or a waiting room. Many demos are recorded in a quiet studio on a good microphone. The gap between those two places is where most of the trouble lives.

What breaks on a real phone line

Phone audio is narrow. A phone line carries far less sound detail than a voice memo. Standard phone audio is sampled at about 8 kHz, which engineers call narrowband. Short answers suffer most. A one-word reply about half a second long, such as the name of a language, can easily be heard as different words. A name, a "yes" or a single digit is exactly where recognition fails, and a laptop demo hides it.

Interruptions need a policy. Barge-in means the caller speaks while the assistant is still talking. There are two basic choices. The assistant can stop at the first sound and listen, or it can finish its sentence and then listen. If your callers are usually somewhere quiet, such as an office or home, always-on barge-in works well. If they may be in a car, on a street or in a busy home, a better design hears speech during its own sentence but does not stop for it automatically. It first judges whether that speech relates to the current step. If it does, it acts on it. If not, it drops it and carries on until it is the caller's turn.

Not all noise is equal. Steady noise, like a fan, traffic or a hum, can be handled well by noise suppression. Other people's voices are much harder, because to the system a TV or a person nearby is speech. This is the job of voice activity detection (VAD) and endpointing: deciding when someone started talking and when they finished. In plain energy-based detection, a sound burst of about 60 milliseconds can already count as "the caller started talking", so a laugh from the TV can cut the assistant off mid-sentence. Voice platforms publicly report that background voices and noise are the main cause of false interruptions, and that a short minimum-duration check before an interruption counts removes most of them, at the cost of a fraction of a second. How much this risk matters depends on how much your call flow relies on the caller speaking at all.

Why the fix is design, not more AI

A common first instinct is a "smarter" call: measure the room, and switch behaviour once the call sounds strained. In practice this needs a fragile web of counters, and the signal it relies on still cannot tell the caller from the person next to them. A sturdier approach changes the flow itself:

• Every closed question (yes or no, one of a few options, a date of birth, a number) also accepts the keypad, and plain logic reads the keypress. No AI is needed for that step.
• If a question has to be asked a second time, for any reason, the re-ask names the keys to press.
• Every confirmation is "press 1 for yes, 2 for no".
• Speech that arrives while the assistant is talking can interrupt only if it is about the step the call is on. A cough or a TV line is ignored.

The keypad works because of DTMF (dual-tone multi-frequency): the phone network sends each key as a defined tone, so it is never misheard. With this design the call can always move forward without the caller fighting the speech recognition. The unpredictable part, understanding free speech, shrinks to the few steps that truly need it, such as "what is the visit for?". Everything else gives the same result for the same input. Running cost tends to drop too, but the real gain is control.

A checklist for choosing an AI phone answering service

1. Completion rate. What share of calls finish the task with no human, and how is it counted?
2. Barge-in mode. Does the assistant stop at any sound, or judge relevance first? Can it be set per business, based on where your callers usually are?
3. Noise handling. What happens with a TV on or someone talking nearby, not just with traffic?
4. Keypad support. Can every clear, parametric answer (yes/no, options, dates, numbers) be entered on the keys? A flow where 80% of the steps are keypresses is fine if callers finish faster and more surely. Having AI does not mean every step must go through AI.
5. Narrow questions. Where speech is really needed, does the flow ask one narrow thing at a time, so it is easier to say and easier to understand?
6. The second try. If the caller is not understood, does the re-ask change strategy, by naming keys or offering numbered choices? Or does it repeat "Sorry, I didn't get that" until they hang up? How many tries before a person or a callback takes over?
7. Read-back before action. "Thursday 3 pm with the hygienist. Press 1 for yes, 2 for no." Nothing is booked or changed before that.
8. A human exit and your approval. A call the system was not built for should reach a person or a callback with a clear reason, never a made-up answer. What do you, the owner, see and approve?

Test it the way your callers will

Before you sign, call it yourself. Call from the car. Call with the TV on and someone talking next to you. Ask a friend with an accent to call on a weak signal. Give one-word answers. Interrupt it mid-sentence. Then count how many of those calls ended with the task actually done.

A phone assistant is not a voice actor. It is a process that has to finish the job without annoying the caller.

Bring this checklist to every vendor conversation. The answers will tell you far more than the demo.