How to Audit What Your AI Receptionist Is Saying to Customers

The question nobody asks until month three

Every vendor demo answers the same question: can it book an appointment? Almost nobody asks the one that matters once the thing is live on your main line - what did it actually say to my customers yesterday?

You have probably never known the answer for your human front desk either. Nobody has. The difference is that with software there is no excuse: every word is already text before it becomes speech, so if you cannot read it back, that is a decision the vendor made, not a technical limit.

This article is about what to demand. It is written from building one of these, so discount it accordingly - but the checklist works on any vendor.

One: the full transcript, both sides, kept

Not a summary. Not "call outcome: booked". The actual words, in order, from both the caller and the agent.

Ask three things about it. How long is it kept? Twelve months is a reasonable answer; thirty days is not, because the disputes that need a transcript surface late. Is it searchable? The value is not reading one call, it is asking "how many people asked about implants last month" and getting a number. Is it in the original script? A Kannada call transliterated into English letters is not a transcript, it is a guess about a guess.

The thing this unlocks is not compliance. It is the ordinary Tuesday where you notice eleven people asked for a treatment you do not advertise, or that half your callers ask about parking.

Two: the script - the exact words it may say

This is the screen almost nobody publishes, and it is the one that decides whether you can responsibly put software on your main line.

An AI receptionist is working from instructions. You should be able to read them - not a marketing summary of them, the actual operative text, for every stage of a call. And critically, it should be assembled from the live configuration rather than written by hand, so it cannot drift from what is really running. A document describing what the agent is supposed to say is worthless three weeks after someone changes a setting.

Ask the vendor to show you this screen during the demo. If the answer is "we can send you a document", that is a no.

Three: the answer sheet - where its facts come from

When the agent quotes a price on a call, you need to be able to open a screen and see exactly where that number came from.

This matters more than it sounds. The dangerous failure of a voice agent is not mishearing - it is a confident invention. A price that was never loaded, a doctor's name it constructed, a slot that does not exist. The structural fix is that the agent can only say what is in a knowledge base you control, and you can read that knowledge base.

The test during a demo: ask it something deliberately outside what it was told. At a clinic that never loaded it, "do you do teeth whitening for children?" The good outcome is "I do not have that to hand, let me put you through". A confident answer is a reason to walk away, however smooth it sounded.

Four: watching a call while it happens

Recordings are retrospective. What changes how a business runs is seeing a call arrive live, with the transcript filling in line by line, and being able to take it over mid-conversation - pulling it to your own phone without the caller being asked to ring back, and with everything said so far going with you.

That capability is also the honest answer to "what if it gets something wrong". It will, occasionally. The question is whether you find out in three weeks or in three seconds.

Five: what it cost, to the paisa

Ask to see the per-call cost, not just the monthly invoice.

Two things to check. What counts as a billed minute - some vendors bill ringing and hold time, which is charging you for silence. And are there alerts before you hit the limit, at something like 80% and 95%, with a guarantee that a call in progress is never cut off mid-sentence.

A vendor who shows you per-call cost is telling you they are comfortable with you doing the maths. That is worth something on its own.

Six: what you were losing, made visible

The most useful screen is often the one showing calls that ended before a conversation started - the list that was invisible before. Pair it with a view of where callers drop off between being answered and being booked, and you stop guessing about the funnel.

And one India-specific detail worth checking: does the dashboard show times in IST, or in UTC? It sounds trivial until "today" on your report ends at half past five in the morning and your Sunday evening bookings land on the wrong day.

The six questions, together

1. Can I read the full transcript of every call, both sides, in the original script - and for how long?
2. Can I read the exact words the agent is allowed to say, generated from what is actually running?
3. Can I see the answer sheet it quotes prices from?
4. Can I watch a call live and take it over mid-conversation?
5. What did that specific call cost, and what counts as a billed minute?
6. Can I see the calls that were missed, and where bookings drop off?

A vendor who answers all six is not necessarily the right one for you. But a vendor who cannot answer four of them is asking you to put something on your main business line that you have no way to inspect, and that is a bad trade at any price.