Customer support chatbot setup: a plan for the first 90 days

Why support chatbots often disappoint

A customer support chatbot usually starts with a big promise: «the bot will close half the requests». Three months later the picture is different — the bot answers simple questions, but customers skip past it and type «agent», and agents restart from scratch the conversations the bot left half done. The problem is not the bot; it is the order it was built in: trying to automate everything at once, not defining what the bot may answer, and leaving handover to a person for later.

This article is a plan for building a support chatbot in stages: which requests to start with, what answering authority to give the bot, how to set up escalation, how to run the pilot and what to watch in the first 90 days. The quality of the bot's material and the mechanics of handover are covered in separate articles — here we put them into one system.

Step one: the top requests

A support chatbot's scope is set by real messages, not the team's guess. Take at least two hundred support messages from the last month or two and sort them into topics.

1. 1Topic listDelivery time, order status, returns, payment method, opening hours, how to use the product, complaints — each message falls into one topic.
2. 2FrequencyThe share of each topic. Usually five to seven topics make up most messages.
3. 3ComplexityFor each topic: is the answer fixed, does it need live data, does it need a decision?
4. 4Pilot choiceFrequent topics with a fixed answer and no decision needed — those are the bot's first job.

Step two: answering authority

For each topic, what the bot may do is written as one of three levels.

1. 1Full answerThe bot answers and closes the conversation: opening hours, delivery rules, product details. The source is the knowledge base.
2. 2Collect and hand overThe bot gathers what is needed — order number, the problem, a photo — and passes a ready conversation to an agent. A return request, a damaged item.
3. 3A person, nowThe bot does not try to answer: complaints, legal matters, payment errors, an angry customer.

Without an authority document the bot either tries to answer everything and makes mistakes, or passes everything on and adds no value. This document is the head of support's most important decision.

Step three: the material

The bot should answer only from the material it is given. For the pilot topics, that material should be short, approved and owned:

• One approved answer text per topic — checked, not copied from the FAQ page
• Rules quoted word for word: the return period, warranty terms
• Common follow-ups: «what about sale items?», «what about the regions?»
• What the bot says when it does not know

How the material is checked before launch is covered in detail in the chatbot knowledge base audit. In Vexvon, material is kept in the knowledge base by topic category, and each answer shows which fragment it used — the fastest way to find the source of a wrong answer.

Step four: escalation

Handover to a person is not an add-on for later — it has to work from the pilot's first day. At minimum, four rules:

• When the customer types «agent», «person» or «manager» — at once
• When the bot cannot answer the same question twice
• When the topic is on the «a person, now» list
• When the customer is angry or uses abusive language

In Vexvon, an «I want an agent» message sends the team a notification, the bot stops when an agent joins the conversation, and it goes silent on its own if the customer swears. When the bot comes back after the agent finishes is configurable too. The full mechanics are in handing over from chatbot to agent.

Step five: the pilot plan

1. 1Weeks 1–2: one channel, five topicsThe bot answers only the pilot topics; everything else goes to an agent. Twenty conversations are read every day.
2. 2Weeks 3–4: fixesWrong answers go back into the material as fixes. Handover reasons are analysed — which question is handed over often, and could the bot learn it?
3. 3Weeks 5–8: expansionA second channel and the next five topics — only if the first stage's figures have improved.
4. 4Weeks 9–12: routineA weekly quality review, a monthly material audit, a quarterly scope decision.

Example: three levels on one topic

On returns, the same conversation can end three different ways, and the bot chooses which from the authority document.

1. 1«How many days do I have to return something?»Full answer: «14 days if the item is unused. You'll need the receipt and packaging.» The conversation closes.
2. 2«The kettle I bought doesn't work, I want to return it»Collect and hand over: the bot asks for the order number, purchase date and a short description, requests a photo, then passes the conversation to an agent ready to go: «I've sent the details to the team; they'll contact you today.»
3. 3«This is the third time I'm returning something, I want my money now!»A person, now: the bot offers no explanation, acknowledges the frustration and passes the conversation straight to an agent.

Roles: who is responsible for what

• Head of support — the answering authority document and escalation rules
• Knowledge owner — keeping the material current and entering fixes
• Agents — handed-over conversations and logging wrong answers
• Technical lead — connecting channels, live-data integration
• Management — the expansion decision, based on the figures

When one of these roles is empty, the bot weakens in that area. The role most often left empty is the knowledge owner — and without it the bot goes stale within weeks.

What to watch in the first 90 days

• Share of conversations closed entirely by the bot — on pilot topics
• Reasons for handover — the top five
• Information the agent has to ask again after handover — should fall
• Wrong answers — how many and what kind
• Customers writing the same question again soon after — the answer did not work

The share of conversations handled entirely by the bot is a section of its own in the analytics report.

Common mistakes

• Launching every topic on every channel on the same day
• Loading the existing FAQ page into the bot without checking it
• Planning handover to a person as «phase two»
• The bot not introducing itself as a bot
• Measuring success by the number of bot replies

Limits

A support chatbot does not do work that needs a decision: compensation, exceptions, out-of-warranty repairs, contract disputes. Nor does it know information missing from its material, and it should not — «I'm not sure, I'll pass you to an agent» is an honest and correct answer. Questions that need live data, such as order status, are answered automatically only once the relevant integration is in place.

Conclusion: a support bot is built in stages

A customer support chatbot starts from the top requests and continues with written answering authority for each topic, approved material and escalation that works from day one. With a narrow pilot the result is readable, fixes are fast, and expansion rests on evidence. A bot built this way does not replace agents — it frees them from repetitive work.

The rest of the support section is in this category. To map your top requests together, ask for a demo.

Frequently asked questions

1. 1Which businesses is a customer support chatbot suited to?Any business whose support messages repeat the same topics: e-commerce, services, SaaS, clinics, education.
2. 2Will the bot replace agents?No. It will close repetitive questions with fixed answers and pass the rest to an agent ready to go.
3. 3How long does a pilot take?With one channel and five topics, usually two to four weeks. Expansion is decided by the figures.