Two different answers, and the cheaper one is right more often than we would like to admit.
Our own product does website chat, WhatsApp, multilingual answers, bookings and handover to a human, with setup and training done for you. If that is the shape of your problem, buying it is faster and cheaper than commissioning anything, and we will point you at it.
A bespoke build earns its cost when the assistant must read live data from your systems - stock, availability, a customer's account, a ticket history - or act on it. That is integration work, and no off-the-shelf product will do it for your systems.
Whichever route, somebody has to decide what the assistant is allowed to say and make sure the answers exist somewhere it can read. An assistant trained on a four-page site answers four pages of questions and guesses at the rest.
The fastest way to turn a good assistant into a complaint is to let it keep answering somebody who has asked for a person. Making that exit obvious is a design decision, not a setting.
Four things, in rough order of how much difference each makes.
Retrieval over your own pages and documents, so the answer is yours and can be sourced. Without that you have a general model guessing confidently about your business, which is worse than no assistant.
An assistant that says "I do not have that, let me get somebody" keeps the customer. One that invents a delivery time loses them twice - once on the wrong answer and again when they find out.
For most businesses here that means WhatsApp as well as the website, from the same training. A website-only assistant solves part of the problem and leaves the channel people actually use.
The log of what it could not answer is the most valuable output in the first month. It tells you what to write next, and it is the difference between an assistant that improves and one that is quietly stuck at launch quality.
What sits behind a chatbot, and what it becomes if it needs to act.