How long does it take to launch a good chatbot?

It depends almost entirely on how fast you have the content: the 20 questions with approved answers and the handoff rules. The technical part is small next to that. Without content, any timeline is an illusion.

Does it have to use generative AI?

Not necessarily. For 20 fixed questions, deterministic answers are enough and safer. The model helps with understanding varied phrasing and natural conversation — but the answer must come from your approved text.

What happens at night, when nobody can take over?

The chatbot says clearly when the team replies ("tomorrow by 10:00"), collects the contact and the question, and sends everything to the team so it is the first thing resolved in the morning. A concrete promise matters more than an instant answer.

A chatbot that actually helps: what it must do for a Romanian

The situation

It is 9:40 pm. A potential customer is on your site and wants to know whether you deliver to Cluj and how long it takes. Nobody answers until morning. By morning she has ordered from someone else. The next day an existing customer writes on WhatsApp "where is my order?" — for the third time this week — and someone on your team loses ten minutes looking it up.

A good chatbot solves exactly those two moments: the simple question at the wrong hour and the repetitive question at the right hour. It does not replace your salesperson. It gives them their time back.

Good vs. bad

What a good chatbot does

Answers the 20 real questions. Not everything — the ones that actually come in: hours, delivery, price, availability, how to book, where to park, what documents to bring. In most businesses, 20 questions cover most of the volume.
Books or qualifies. Since it is there anyway, it can reserve a slot in the calendar or collect the 3–4 details that make a lead useful: what they want, when, rough budget, how to reach them.
Hands off to a human fast. When it does not know, or the customer is upset, it says clearly "let me connect you with a colleague" and does it — a WhatsApp or email notification to the team, with the full history. No "please rephrase your question".
Lives where the customer is. In Romania that means WhatsApp and the website first, then Messenger and Instagram for certain sectors. One brain, several channels.
Sounds like you. Proper diacritics, no translated phrasing. Short, to the point, in the register you use with customers.

What a bad chatbot does

Menus. "Press 1 for hours, 2 for prices." The customer arrived with a question already phrased; forcing them through a menu is a small insult, every time.
Dead ends. It does not understand, asks you to rephrase, does not understand again, and there is no way out to a human.
Lies. It invents a price, promises "24-hour delivery" you do not offer. The customer considers that you promised it. A generic chatbot wired to your site without your data does this from day one.
The annoying bubble. It pops up three seconds after load, on every page. People close it and never come back.

The content you need before any code

A chatbot is exactly as good as the text behind it. Before you talk to anyone about technology, prepare:

The list of 20 questions — take them from the last two weeks of messages and calls, not from imagination. Count how often each appears.
The approved answer for each — written by someone allowed to make promises. In two versions: one sentence, and a detailed one.
Handoff rules — when it transfers to a human, to whom, on which channel, in which hours, what happens at night.
What it does not answer — topics you do not want it touching (legal complaints, price negotiation, medical data). Polite refusal + handoff.
The tone — three examples of "yes, like this" answers and three "never like this".

Where it lives

Channel	When it makes sense
Website	New visitors, pre-purchase questions, qualification
WhatsApp	Existing customers, statuses, bookings, confirmations
Messenger / Instagram	Sectors where customers already write there (beauty, hospitality, retail)

The rule: start on the channel where you already have the highest volume of repetitive questions. Usually WhatsApp for services, the website for e-commerce.

How you measure

Resolved rate: of the conversations started, how many ended without a human and the customer got what they wanted. Not "how many got an answer" — how many got the right answer.
Handoff rate: how many reached a human. A high number is fine at the start — it is a problem if the handoff is slow or loses context.
Leads and bookings: useful contact details captured, confirmed reservations. Compared with the same period before.
Team response time: after handoff, how long until the human replies. A chatbot that hands off to someone who answers the next day has solved nothing.

The two weeks of transcripts

The most important thing you do after launch: read the transcripts. All of them, for two weeks, 20 minutes a day. You will find questions you did not think of, phrasings it does not understand, moments where the right answer is too long. Every finding becomes a new answer or a correction. After two weeks you have a chatbot built on reality, not on assumptions.