Why the Cheapest AI Receptionist Is Usually the Expensive One

The number everyone compares, and the number that matters

Every AI receptionist gets compared on one figure: what does a minute cost. It is the easiest number to put in a table, so it is the number vendors compete on.

For a business with genuinely high call volume, that is correct. If you take fifty thousand calls a month asking where a delivery is, a rupee per minute is real money and you should buy the cheapest thing that answers.

Almost no Indian SMB is that business. A busy single-location dental clinic takes somewhere between four hundred and a thousand calls a month. A design studio takes fewer. A diagnostic lab takes more but they are shorter. At those volumes the entire monthly bill is small relative to what one caller is worth - and that changes which number you should be optimising.

The arithmetic, on your own numbers

Work it out rather than taking anyone's word for it. Three figures, all of which you already know.

One: what is a converted caller worth to you? Not revenue - margin, and count the repeat business. For a dental clinic a single implant or ortho case runs to tens of thousands of rupees. For a salon a regular client is a few thousand a month for years. For a real-estate desk a site visit that closes is a different order of magnitude again.

Two: how many callers a month? Ask your operator for a call detail record - Jio, Airtel and Vi all provide one.

Three: what share of them currently become customers?

Now the comparison. Suppose the cheaper system saves you two thousand rupees a month. If a converted caller is worth twenty thousand to you, that system has to convert no worse than one-tenth of a caller per month worse than the alternative to break even. One lost booking a quarter wipes out the saving. One a month and you are paying several times over for the privilege of a lower invoice.

This is the whole argument. At low volume and high customer value, the per-minute rate is noise and the conversion rate is the signal. The trouble is that the per-minute rate is printed on the website and the conversion rate is not.

What cheap usually means in practice

Cheaper systems are not cheaper by magic. The cost comes out somewhere, and in voice AI it comes out of the same few places.

Language detection once per call instead of per utterance. The system decides what language you are speaking from your first few words and locks on. That is much cheaper to run and it is fine right until a caller opens in English, gives the date in Telugu and confirms in English - which in Hyderabad, Bengaluru or Kochi is simply how people talk. From the lock-on point the call degrades and nothing flags it.

Guessing instead of transferring. When a system does not understand, the safe behaviour is to say so and hand to a person. The cheap behaviour is to produce the most probable interpretation and carry on. You get a booking record that looks perfect and a patient who arrives on a day they never agreed to.

No read-back before committing. Confirming the appointment out loud and waiting for a yes costs an extra turn on every single call, which costs minutes, which costs money. Skipping it is invisible in a demo and expensive in production.

No transcript worth reading. Storing and indexing every call in the original script is a real cost. Storing a one-line outcome is nearly free. The second one means you cannot check any of the above.

None of these are visible when you are comparing price tables. All of them are visible in a two-minute test call, which is why we would rather you made one than read this.

What the difference is spent on here

To be specific about our own product, since that is the fair thing to do in an article like this.

The agent reads the appointment back and will not create it until the caller agrees. That is enforced in code, not asked for politely in a prompt, because we learned the hard way that a prompt rule is a suggestion. It costs a turn on every booking call and we think it is the single highest-value thing in the product.

Language is detected per utterance, across English, Hindi, Telugu, Tamil, Kannada, Malayalam, Marathi and Bengali, so a caller who switches mid-sentence is followed rather than slowly misunderstood. Where we genuinely cannot help - Gujarati, Punjabi, or any clinical question in a healthcare setting - the call goes to a person rather than to a guess. We publish those gaps rather than hiding them.

The agent answers one question at a time and will not restate something it has already said, which sounds like a small thing and is the difference between a call that feels like a conversation and one that feels like a form.

And the part you can actually check

All of the above is a claim. The reason to believe any of it is that you can audit it, and that is where most of the remaining difference goes.

• Every call as a written transcript, both sides, in the language and script it was held in, searchable - so you can ask how many people enquired about a treatment last month, not just read one call.
• The exact words the agent is permitted to say, on a screen, assembled from the live configuration rather than written by hand - so it cannot drift from what is actually running on your line.
• The knowledge base it quotes prices from, editable by you. It can only say what is in there, which is the structural reason it does not invent a price.
• Live calls, with takeover - pull a call in progress to your own phone without the caller being asked to ring back, and everything said so far comes with you.
• Per-call cost against your included minutes, with alerts before you reach the limit. Only connected talk time is billed.
• The calls you were missing, listed - the ones that ended before a conversation started.

And the screen that matters most: the booking funnel. How many calls were answered, how many reached a real conversation, how many asked about an appointment, and how many ended with one - plus the names of the people who dropped out between those stages, so you can ring them back this afternoon.

Every one of those screens is described in more detail on the dashboard page, and there is a separate guide to auditing what your AI receptionist says that applies to any vendor, not just this one.

That funnel is the honest test of everything above. If the conversion rate is not better than what you had, none of the engineering matters and you should stop paying us. We would rather sell you something you can measure than something you have to believe in.

It is also not a generic bot

One more thing the money buys, which is easy to miss when comparing voice quality.

A dental clinic gets clinical plans and patient records. A salon gets services, packages and per-stylist availability. A real-estate desk gets projects, units and a site-visit pipeline. A hotel gets rooms, reservations and housekeeping. A coaching institute gets courses and an admissions pipeline. These are not the same product with a different greeting - they are different operating screens, because the work a receptionist does in a clinic is not the work one does in a hotel.

A cheaper generic voice bot will answer your phone. It will not know what a follow-up visit is, or that a stylist is on leave, or that an enquiry for a two-BHK in one project should be routed differently from a site-visit request in another.

When you should buy the cheap one

We would rather say this plainly than have you find out in month two.

If your calls are high-volume and low-value - order status, delivery tracking, balance enquiries - buy the cheapest thing that answers. The arithmetic at the top of this article runs the other way and everything we have built is wasted on you.

If your front desk already answers essentially every call during working hours and your after-hours volume is negligible, buy neither. Spend the money elsewhere.

If most of your inbound calls need genuine judgement rather than repetitive administration, automation helps less than any brochure suggests, ours included.

What is left - a business taking a few hundred calls a month, where a single caller can be worth tens of thousands of rupees, losing some of them to a phone that rings out - is the case where paying more for conversion rather than less for minutes is simply the better trade.

The number itself, and what it does and does not include, is on the pricing page; there is a separate worksheet for working out the ROI on your own numbers.

And you do not have to take the argument on faith, which is the point of running a free seven-day pilot on your own number before you pay anything. Real calls from your own customers, every one of them readable afterwards, and the funnel showing what share of them ended in a booking. If the numbers do not move, you owe nothing and you have lost a week. That is a cheaper way to settle this than reading another comparison table.