"Automated answering service" covers everything from a 1990s phone tree to something that books appointments. The words are identical and the experiences could not be more different, so it is worth being able to tell them apart before you buy one.
Three generations, all still on sale
The phone tree. Press one for bookings, press two for hours. Formally an IVR. It works if the caller's need fits a branch you anticipated and they are patient enough to listen. Most people press zero immediately, which is a reasonable summary of forty years of user research.
The recorded message with a mailbox. Your hours, read aloud, and a beep. Not really answering — a politer voicemail.
Conversational AI. The caller says what they want in their own words. No menu. It understands the request, checks the real calendar, and completes the booking on the call.
All three get described as automated answering. Only the third one does what the phrase suggests, and the first two are the reason people are sceptical of the third.
Why phone trees are worse than they look
The problem is not that callers dislike them, though they do.
It is that a phone tree can only route, never resolve. Every branch ends at a person or a mailbox. So the best possible outcome of a well-designed tree is that the caller reaches the right voicemail slightly faster. If nobody is there — after hours, mid-rush, a Sunday — the tree has spent the caller's patience and delivered them to the same beep.
And the branches never match real requests. "I need to move Thursday to next week and my daughter wants the same slot" does not fit a menu. The caller presses zero, and if zero goes nowhere they hang up and ring somebody else.
If you currently run a phone tree, the most useful thing you can do is call it yourself and try to accomplish something specific without pressing zero. It is usually a short and educational experience.
What "conversational" needs to mean to be worth anything
Vendors all claim it. Four things separate the real ones.
It handles a vague request. "Tomorrow afternoon", "whenever Maria's free", "same as last time". If the system needs a service selected and a time stated, it is a form read aloud.
It can be interrupted. A caller should be able to cut in mid-sentence and be heard within about half a second. Scripted systems finish their sentence first, and it is immediately obvious.
It reads the live calendar. Not a generic slot grid — the real diary, correct durations, only staff qualified for the service.
It knows when it does not know. The single best indicator. A system that invents a price, or reassures someone about an allergen, is more dangerous than one that admits ignorance, because it sounds right.
How voice AI understands a phone call goes through the machinery underneath those four.
The mistake in the middle
There is a tempting halfway house: keep the phone tree, and put AI on one branch. Press one to speak to our assistant.
It performs badly, for a boring reason. The people most likely to benefit — hurried, driving, older, not sure what they want — are exactly the people who press zero before the menu finishes. You have gated the useful thing behind the thing everybody skips.
If you are going to answer conversationally, answer conversationally from the first second.
What it should do with the calls it should not handle
An automated system earns trust by refusing work.
An allergy question. A complaint. A medical query. A colour correction that needs a consultation. A price that depends on looking at the client's hair. In each case the right behaviour is to say plainly that a person will deal with it, offer a callback, and transfer if someone is available.
And when the transfer does not connect — nobody free, after hours, a dropped handoff — the caller should be told and given your number, rather than left in silence. That specific failure is worth asking every vendor about, because it is common and rarely designed for.
Testing one honestly
Three calls, ten minutes.
Ask for a vague time. Listen for two or three genuinely spread-out options, not 12:00, 12:15 and 12:30.
Interrupt it. Does it stop, and does it do something sensible with what you said?
Ask it something it cannot know. Does it admit it, or improvise?
The longer version, with the specific failures to listen for, is in five calls to make before you trust an AI receptionist — all of which are failures we found in our own system before a customer met them.
Where to start
You do not have to replace anything to find out whether this works. Conditional call forwarding hands over only the calls you do not answer, which means the automated system gets the overflow and the after-hours calls and nothing else. Your number is unchanged and your existing setup keeps working. The codes are in the call forwarding guide.
Point the calls that are currently reaching a beep at something that can book, and see what comes back in a fortnight. That is a real test, and it costs you nothing that was not already being lost.