AI answers the easy calls beautifully.
It is the other calls that decide whether you keep the customer. This is an honest account of what voice AI gets wrong on a business phone, why you will probably never see it happening, and how to work out which of your calls actually need a person.
First, the fair part
We are not against automation.
We use software on every call ourselves: the call plan is on screen, the details are typed straight into the client's system. The argument is not people versus machines.
If your calls are mostly opening hours, address, order status and simple bookings, meaning high volume, low value per call, and a mistake that annoys a customer rather than losing a job, then voice AI is good value and we will tell you so on the first phone call. What follows is about the calls that do not look like that.
The failure modes
Seven calls that go wrong, reliably.
None of these are edge cases. Every business that answers a phone gets all seven, most weeks.
The vague caller
Nobody describes a problem in your vocabulary
A customer does not ring and say they need a hot water system replacement. They say there is no hot water, then that the thing outside is making a noise, then that their daughter is visiting on Thursday. A person assembles a job from that. An automated system waits for a keyword that never arrives, then offers to take a message.
The upset caller
Tone is information, and it gets discarded
Someone whose ceiling is dripping, whose matter is going to court on Monday, or who is in pain does not want efficiency. They want to hear another human take the problem seriously. Being processed briskly at that moment reads as contempt, and it is remembered long after the job is done.
The Australian caller
Suburb names and trade slang are a known weak point
Place names here are unforgiving, and a system trained mostly on other accents mishears them. Getting a suburb wrong is not a cosmetic error: it routes the job to the wrong area, quotes the wrong call-out, or loses the address entirely.
The call that turns
It starts as one call and becomes another
A routine booking becomes a complaint. A price enquiry becomes an urgent job when they mention water is still running. People notice the change and re-route the call. Automation is usually still working the flow it started.
The policy question
It answers something it was never authorised to answer
Can you be here before three? Will this be covered? How much roughly? The safe answer is that nobody can promise that yet. The dangerous one is an approximation, delivered confidently, that your business then has to honour or walk back.
The silent failure
You never find out what you lost
This is the one that matters most. A mishandled call leaves no evidence: the caller simply hangs up and rings the next business. There is no voicemail, no missed-call log, no angry email. The reporting shows calls answered, and answered is not the same as handled.
The caller who hangs up
Some people will not talk to a machine at all
Older callers, people who are not confident with automated systems, and anyone in a genuine hurry. For a medical practice or a firm doing wills and estates, that is not a fringe group. It is a substantial share of the people you most want to reach you.
Why it is hard to see
The errors are asymmetric.
A person who mishandles a call usually leaves a trace: a note, an apology, a complaint you can act on. An automated system that mishandles a call produces a clean transcript and a satisfied-looking report.
The measurement problem
Answer rate is not the number you care about.
Every automated system will show you a near-perfect answer rate, because answering is the one thing it cannot fail at. The number that matters is how many calls reached the outcome they should have: booked, transferred, escalated, or properly qualified. Almost nobody reports that, because it requires knowing what should have happened.
The cost profile
The cheap calls are cheap to get wrong. The expensive ones are not.
If ninety percent of your calls are simple and ten percent are worth thousands, a system that is excellent at the ninety and poor at the ten is a bad trade, even at a fifth of the price. Work out the value of the calls in that ten percent before comparing monthly costs.
The reputation lag
You find out months later, in a review.
Nobody rings back to tell you their call went badly. It surfaces as a quiet decline in enquiries that convert, or a review mentioning that they could not get hold of anyone. By then the cause is months in the past and nearly impossible to attribute.
A straight test
How to work out which calls need a person.
Take last month's calls and sort them by one question: what does it cost if this call is handled badly?
Automate
Cost of error: an annoyed customer
Hours, address, order status, straightforward bookings, callers who already know exactly what they want. Predictable shape, low value, easy to recover from. Automate these without hesitation.
Use a person
Cost of error: a lost job, or worse
First-time enquiries, anything urgent, anything where the caller is distressed, anything with a compliance or duty-of-care dimension, and anything where the next step depends on judgement rather than a lookup.
The honest bit
If everything lands in the first box
Then you do not need us, and we will say so. A business whose calls are genuinely all predictable should buy the cheapest reliable automation it can find and spend the difference somewhere useful.
What we do instead
A person, following a written plan.
The reason most human answering services are also disappointing is not the people. It is that nobody wrote down what they were supposed to do.
Automation gets its consistency from being unable to deviate, which is precisely why it breaks when a caller says something unexpected. An unscripted person has the judgement but no discipline, so your intake quality depends on who picked up. Give a person a written decision tree and you get both: the plan handles the predictable ninety percent identically every time, and the person handles the ten percent that decides whether you keep the customer.
That decision tree is written with you, approved by you before a single call is answered, and corrected against real calls every month. You can build a draft of your own for free, or read how the method works in full. If you want the side-by-side, the AI versus human comparison sets both out plainly.
Questions
Common questions.
Is an AI receptionist bad?
No. For high volume, repetitive, low-stakes calls it is genuinely good value and we say so. The question is not whether it works, it is which of your calls it should be allowed to take.
What kinds of calls does voice AI struggle with?
Calls where the caller cannot describe the problem in the expected words, calls where the caller is upset or in pain, calls that change purpose halfway through, and calls where the right answer is a judgement your business has not written down yet.
Will callers hang up on an automated system?
Some will, and the ones most likely to hang up are often the ones worth the most: older callers, callers in a hurry, and callers with an urgent problem who want to hear a person take responsibility for it.
How do I know what an AI receptionist is costing me?
That is the hard part. A mishandled call usually leaves no trace: the caller simply rings someone else. Ask any provider how they report calls the system could not complete, and whether you can listen to them.
Can I use AI for some calls and people for others?
Yes, and for many businesses that is the sensible destination. Automate the predictable calls, put people on the ones where a mistake is expensive.
Work out which calls need a person.
Tell us what your phone does on a bad day. We will map the flow with you and tell you honestly which parts could be automated and which should not be.