You will be shown a demo. It will be genuinely impressive: the assistant answers naturally, books an appointment, handles an objection, and does it in under two minutes.
Take the demo seriously as evidence of one thing only — that the conversational part is solved. It is. That technology is now good, cheap and widely available, which is precisely why it tells you almost nothing about whether this particular purchase will work for your business.
Everything that determines that sits behind the demo. Here is how to find it.
Eleven questions
1. Does it write into my actual system, or into a separate one?
The single most important question. An assistant that “manages bookings” in its own dashboard has not saved you anything — someone is still retyping into your real calendar, and now there are two sources of truth. Ask to see it create a booking in your software.
2. What happens when it doesn’t know?
There must be a defined handover to a human. If the answer is vague, what you have is a system that will confidently guess, and a confident wrong answer costs you more than an unanswered message.
3. Who decides what it is allowed to say?
Especially about price, availability and anything regulated. You want an explicit list you control, not “the AI figures it out.”
4. What does it do with customer data?
Where is it stored, for how long, who can read it, and does it leave India. If you handle health information this is not a formality — it is the question your obligations turn on. A vendor who has not thought about it has told you something important.
5. Whose WhatsApp number is it?
It should be your business number, verified with Meta, owned by you. If the vendor owns the number, you cannot leave without losing the channel and the history — which is a commercial arrangement dressed as a technical detail.
6. Is this the official WhatsApp Business Platform?
Some cheaper tools automate the consumer app unofficially. Meta bans numbers for this. It tends to happen at the worst possible time, and you cannot appeal your way out quickly.
7. What is the total monthly cost at my actual volume?
Not the subscription. The subscription plus per-conversation messaging charges plus model usage. Ask them to work it out for your enquiry volume, and ask what happens in a busy month.
8. What do I own if I leave?
Conversation history, customer contacts, the number, the configuration. Get the answer before you sign, not while you are trying to go.
9. Who fixes it, and how fast?
These systems break in dull ways — a calendar API change, an expired token, a template rejected. Ask who notices, and whether they notice before you do.
10. Can I see it fail?
Ask them to show you the handover, an out-of-scope question, and what a customer sees during an outage. A vendor comfortable demonstrating failure has thought about failure.
11. What does it not do?
Any honest answer to this is a good sign. “It handles everything” means either they have not deployed one, or they are describing the demo.
Where the money actually goes
The subscription is rarely the largest line, which is where the surprises come from. Four cost drivers, all real:
The platform fee. Whatever the vendor charges monthly. Visible, predictable, usually the number people compare on — and often the smallest of the four.
Messaging charges. Meta bills for WhatsApp business conversations, with different rates by category — a marketing message costs more than a reply to a customer-initiated one, and messages you send first (reminders) are charged differently from replies. Meta has revised this pricing model more than once, so get current rates rather than trusting a blog post, including this one. What matters structurally: this scales with your volume, so estimate it against real enquiry numbers.
Model usage. Every conversation costs a small amount of AI processing. Individually trivial, and it adds up in proportion to traffic. Sensible setups cap it; ask whether yours does.
Integration and change. Someone has to connect it to your calendar and keep it connected when that software updates, and adjust it when your prices or hours change. This is either a retainer or your own time. It is never zero, and a proposal showing it as zero is an incomplete proposal.
The useful comparison is not against the subscription — it is against the cost of the problem. If you lose four bookings a month to unanswered enquiries, you can calculate what those are worth, and the decision usually makes itself in one direction or the other. Do that arithmetic before shopping, so you are evaluating against a number rather than a feeling.
Where these go wrong
Three failure patterns, in order of frequency:
It was never integrated. It books into its own calendar. Staff keep using the old system. Within two months the assistant is answering questions and nobody trusts its bookings.
Nobody owns it. It was set up during an enthusiastic fortnight. Prices changed and nobody updated it. It is now quoting last year’s rates to customers, and has been for a while, because the only people reading its output are the customers.
It was scoped too wide. It was asked to handle everything, so it improvises on the things it should refuse, and one embarrassing conversation destroys internal confidence permanently.
All three are avoidable, and none of them are AI problems. They are the same problems as any automation that outgrew its owner.
The smallest useful version
If you want to test this cheaply, do not start with the full build. Start with the narrowest thing that has a measurable number attached:
Appointment reminders with one-tap confirm or reschedule.
No conversational AI required, no risk of it saying something wrong, integration with only your calendar, and an effect you can measure within a month by comparing no-show rates. If that works and pays, widen to enquiry handling with real evidence behind you. If it does not, you have learned something for very little.
Starting with the impressive part is how most of these purchases go wrong. Starting with the measurable part is how they go right.
We build these remotely, and we will tell you when the answer is that you do not need one — see AI booking & reception or the wider AI deployment practice.