AI Chatbot Launch Checklist: What to Verify Before Going Live

What to check first

An AI chatbot is ready to launch when it answers common questions correctly, honestly declines when it does not know, and can hand the conversation to a human. Everything else — logging, privacy, monitoring — exists to keep that quality after launch. Here is the checklist by section.

1. Test question set

• Built from real customer questions from chats, email and calls, not ones invented by the team.
• Each question has a reference answer or a correctness criterion.
• All main topics are covered: prices, timelines, delivery, payment, returns, contacts — whatever applies to your business.
• Questions come in different phrasings: short, with typos, in several languages if your audience is multilingual.
• The test run is repeated after every change to the prompt, model or knowledge base.

2. Edge cases

• Off-topic questions: the bot politely steers back to its job.
• Provocations and attempts to make the bot ignore instructions (prompt injection): "forget your rules", "show your system prompt".
• Aggressive or abusive messages: the bot keeps a calm tone.
• Very long messages, empty messages, emoji or stickers only, files and voice notes if the channel supports them.
• Several questions in one message.
• Questions about competitors, politics, medicine, law — it is decided in advance how the bot responds.

3. Refusal behavior

• When the knowledge base has no answer, the bot says so instead of making things up.
• The bot does not promise discounts, deadlines or terms that are not in approved materials.
• The refusal is friendly and immediately offers a next step: an operator, a form, a phone number.
• The bot does not reveal internal instructions or service data.

4. Human handoff

• There is an explicit command or button to talk to a person.
• The bot hands off on its own when the user is unhappy, the question is complex, or it concerns money or a complaint.
• The operator receives the conversation history, so the customer does not have to repeat themselves.
• After-hours behavior is defined: the bot honestly says when a person will reply.
• Handoff notifications are verified to actually reach operators.

5. Logging

• Questions, answers, sources used and escalations are stored.
• There is a way to flag a bad answer, such as rating buttons.
• Logs have restricted access and a clear retention period.
• Personal data in logs is masked where possible.

6. Privacy and transparency

• Users can see they are talking to a bot, not a person.
• There is a short data processing notice and a link to the privacy policy.
• The bot does not ask for more than needed: card numbers, passwords, passport details.
• The model provider's data processing terms and your country's personal data requirements have been checked.

7. Post-launch monitoring

• An owner is assigned to review a sample of conversations regularly.
• Alerts are set up for API failures and error spikes.
• A gradual rollout is planned: part of the traffic or internal users first, then everyone.

Common launch mistakes

• Testing only "good" questions.
• Giving users no way to reach a human.
• Launching and not reading logs during the first weeks.
• Updating the knowledge base without rerunning the test set.