Common Questions

What Happens When an AI Receptionist Gets It Wrong

It will, occasionally. A name misheard, a slot booked wrongly, a question answered from information that was out of date. What matters is that these are visible, correctable and rare — and that you know in advance how each one is handled.

The four failure types How each is caught Who is accountable

Why this page exists

Most suppliers discuss errors only if you push. We would rather set out what goes wrong, how often it is caught, and who is responsible when it is not — because that conversation is far more useful before you buy than after an awkward Monday morning.

Worth holding one comparison in mind throughout. A tired receptionist at 4pm also mishears numbers, and a voicemail nobody played back loses the enquiry entirely. The standard is not perfection; it is whether errors are visible and fixable.

Four ways it goes wrong

FailureWhat it looks likeHow it is contained
MishearingA digit, a name or a street wrongDetails are read back and confirmed on the call; the recording and transcript let you check
MisunderstandingThe wrong intent, so the wrong path is takenLow confidence should trigger a clarifying question, then a handover rather than a guess
Wrong informationA confident answer drawn from outdated contentScope is limited to what you approved; anything outside it routes to a person
Wrong actionA booking made in the wrong slot or under the wrong serviceWritten to your calendar immediately, so you see it rather than discovering it later

The third is the one to watch. A system that has been given incomplete information will still answer, and it will sound certain. That is a configuration failure rather than a technology failure, and it is why the review process matters.

Where errors are cheap and where they are not

A misheard surname on a haircut booking is an apology. The same error on a clinical enquiry is a different matter entirely, which is why scope is decided by consequence rather than by difficulty.

The categories where the cost of error is unacceptable are simply excluded. Matters involving a child, clinical triage and symptoms, security incidents, allergen confirmation, callers in distress, grievances and regulated legal, medical, financial or immigration advice reach a person immediately. Not because they are hard to attempt, but because a confident wrong answer in those areas can cause real harm.

That boundary is fixed in every deployment and cannot be configured away, however convenient it might occasionally be. The reasoning is set out in handling sensitive enquiries.

How you find out

The worst kind of error is one nobody notices. Three layers make that unlikely.

On the call

Numbers, dates and names are read back for the caller to correct. Most mishearings are fixed by the person who noticed, seconds later.

Immediately after

A summary reaches your inbox or systems with the outcome and the details, so a wrong booking is visible the same hour rather than on the day.

In review

Transcripts and recordings let you sample calls weekly at first. Patterns show up quickly, and most are fixed by improving the information rather than the technology.

See call transcripts and summaries and measuring call handling performance. If you record calls, the notification and retention obligations in PDPA and call recording apply.

Who is accountable

Plainly: the business remains responsible to its customers for what is said on its phone line, exactly as it would be for anything an employee said. That is the honest position, and any supplier suggesting otherwise is worth questioning.

What sits with us is the service performing as configured, correcting faults, and telling you when something has gone wrong rather than waiting for you to find it. The contractual detail is in our terms of service, and your own advisers should review it.

  • Errors that recur are treated as configuration faults and fixed at our end
  • Escalation categories are reviewed rather than narrowed when something slips through
  • Incidents affecting call data follow the process in incident response for call data
  • You keep the records, so nothing depends on us to reconstruct what happened

Frequently asked questions

How often does it get things wrong?

We will not quote an accuracy figure, because it varies with line quality, accents, vocabulary and how well the information was prepared. Measure it on your own calls during testing.

What if a customer is upset by a mistake?

Handle it as you would any service failure, with the advantage that you have a transcript showing exactly what was said. Complaints and grievances reach a person rather than being handled automatically.

Can we correct a wrong booking easily?

Yes. It is written to your calendar like any other, so your normal process applies.

What stops it inventing an answer?

Scope. It answers from information you approved, and questions outside that should route to a person. Test this deliberately, as covered in testing before go-live.

Will we know if it hands over more often than expected?

Yes, and a rising handover rate is useful information rather than a fault. It usually means your information needs extending.

What happens if the service is unavailable?

Calls follow your fallback routing, which is decided during setup. Agree that before go-live rather than discovering it during an outage.

Try to break it before you buy it

Bring the calls most likely to go wrong. Watching how it fails tells you more than watching how it succeeds.