Chatbot project readiness: what to settle before you start

Projects stall on unanswered internal questions, not on technology

A chatbot project usually starts like this: someone says «let us build a bot», a platform is chosen, documents are uploaded, and two weeks later the team reads the test conversations and gets nervous. The bot quotes the wrong price, promises a discount, answers a complaint coldly. The root cause is rarely technical. It is that nobody answered five simple questions before the project began: who owns these answers, what the bot must not say, whose authority it speaks with, where to start, and who decides it is ready.

This article is about that preparation — not about loading material into the bot. Choosing and preparing the material is a separate subject, covered in training an AI chatbot on your business data. Here we deal with the decisions that come before it.

Why preparation gets skipped

There are three understandable reasons the preparation stage is skipped, and each one costs more later.

• The chatbot is bought as a tool, not as a process. Buying a tool raises «how does it connect?»; changing a process raises «who is responsible for what?».
• The information «already exists»: there is an FAQ on the site, a price list, a sales script. But these were written for people and may contradict each other.
• No owner is named. Marketing sees the bot as the brand's voice, support as a way to cut workload, sales as a lead source — and each can overrule the others.

So the preparation stage still happens after launch — only in front of customers.

Decision one: who owns the information

Every sentence the bot says needs a source behind it, and every source needs a person. Ownership works when it is split by subject, not by document.

1. 1Prices and promotionsUsually the head of sales or marketing. They own the change that reaches the bot when a promotion starts and ends.
2. 2Product and service descriptionsThe product manager or the unit that delivers the service. A new product's details are approved by them.
3. 3Rules: delivery, returns, warrantyOperations or legal. These texts are given word for word, not paraphrased.
4. 4Support answers and escalationThe head of support. Which questions go to a person is their call.

When each subject's owner is written in one table, the project's hardest question — «why did the bot say that?» — is answered in a day, not a week.

Decision two: what the bot must not say

A list of forbidden answers matters more than the knowledge base itself: a gap in the knowledge base produces «I don't know», while a missing prohibition produces a wrong commitment.

• An unapproved discount, individual price or «special offer for you»
• A guaranteed delivery or completion date
• A refund, compensation or exception decision
• Legal, medical or financial advice
• Comparative claims about competitors
• Internal information: staff personal numbers, internal price lists, margins

Decision three: authority and tone

The bot speaks for the company, which raises two questions: how far it may go, and how it sounds.

• Can the bot change an order, cancel an appointment or assign a lead to a salesperson — or does it only take the request?
• Formal or informal address? Emoji or none? Short replies or detailed ones?
• Does the bot introduce itself as a bot? Our recommendation: yes, especially in the first message.
• Which languages does it answer in, and which language hands the customer to a person?

These are not technical settings; they are brand and risk decisions. Writing them on one page before the project is cheaper than correcting hundreds of conversations afterwards.

Decision four: the pilot's scope

The most common mistake is to start on every channel and every topic on the same day. A pilot has to be narrow so that its result can be read.

1. 1One channelThe one with the most messages — often Instagram or WhatsApp. A second channel is added once the first one's result is clear.
2. 2Five to ten topicsThe questions repeated most in last month's messages. Everything else starts with «hand over to an agent».
3. 3A clear success criterionFor example: most conversations on pilot topics close without a person, and wrong answers are read weekly and fall. The target figure should come from the company's own baseline.
4. 4A time frameUsually two to four weeks, with small fixes daily and a decision on expanding at the end.

Decision five: who signs off «ready»

Going live needs one person's signature — and that person should read the test conversations themselves. A few practical rules:

• The test set is built from real messages, not questions the team made up
• Each subject owner approves the answers in their subject
• Every item on the forbidden list is checked separately — does the bot really stay silent?
• Handover to an agent is tested live at least once
• Daily review time for the first week after launch is booked in the calendar

How to build the pre-launch test itself — which families of scenarios to check — is covered in detail in the chatbot test checklist.

Example: a two-hour preparation session

The five decisions can be made in one meeting rather than weeks of email. One person each from sales, support, marketing and operations attends, and each brings at least twenty real customer messages from the last month.

1. 10–20 minutes: read real messagesEveryone reads five of the messages they brought. The goal is a list of topics: price, delivery, returns, appointments, complaints and so on.
2. 220–45 minutes: the ownership tableA name is written next to each topic. A topic nobody takes is dropped from the pilot — an answer without an owner does not go to the bot.
3. 345–70 minutes: the forbidden listEach function writes two or three sentences the bot must not say in its area, and agrees what the bot does instead.
4. 470–95 minutes: authority, tone and the pilotWhat the bot may change, how it speaks, which channel and which five to ten topics it starts with.
5. 595–120 minutes: success criterion and sign-offBaseline figures are written down, along with what will be checked at the end of the pilot and who approves going live.

The session produces a one-page document. Technical work with the platform starts from it, and every later question is settled by referring back to it.

Readiness checklist

Before the project starts, you should be able to say «yes» to every item.

• Information owners by subject are written in a table
• There is a forbidden-answer list with an alternative action for each item
• The bot's authority — what it may change, what it only records — is written down
• Tone, language and how the bot introduces itself are agreed on one page
• The pilot has one channel, five to ten topics and a success criterion
• It is known who answers handovers, and during which hours
• The person who approves going live is named
• A weekly review for the first month is in the calendar

Limits and common false expectations

Preparation does not make a project risk-free; it makes the risks visible. After launch, the first weeks will still bring new kinds of questions, unexpected phrasing and gaps in the material. The difference is that a prepared team knows which owner each mistake goes to, and the fix takes hours rather than days.

Another false expectation is handing preparation entirely to an outside party. A platform can do the connection and technical setup, but only the company can name the owner of prices, write the forbidden list and set the success criterion.

Conclusion: a good project starts with a one-page decision

Preparing for a chatbot project is not uploading documents. It is five decisions: who owns the information, what is forbidden, authority and tone, the pilot's scope, and sign-off. When they are written on one page, the project is measured in days rather than weeks, and there are no surprises in front of customers after launch.

In Vexvon, the material the bot answers from is managed in the knowledge base — with subject categories, an active/inactive state and a view of which fragment was used in each answer. If you are still choosing a platform, twelve questions for choosing a B2B chatbot will help. The rest of this section is under AI chatbot strategy. To run the preparation session together, get in touch.