How to plan a useful customer-support chatbot

A useful chatbot is a support workflow with clear limits, not just a text box connected to a model. It should help customers find reliable answers, know when it is uncertain and transfer the conversation to a person when needed. The planning work begins with your support questions and source material.

Find the questions worth automating

Review real support conversations, search terms and help-centre gaps. Group questions by frequency, consequence and whether the answer can be grounded in approved content. Delivery status, account access and policy explanations have different data and privacy needs. Begin with questions where a wrong answer is easy to detect and correct.

Create a dependable knowledge source

Assign an owner to each policy or help article. Record when it was reviewed, which markets it applies to and what should happen if information conflicts. Retrieval can bring relevant content to a model, but it cannot make an outdated source correct. If the source changes often, plan an update process before launch.

Design the boundaries of an answer

The assistant should distinguish a general explanation from an account-specific action. Require authentication before exposing personal order or billing information. Avoid promising refunds, delivery dates or policy exceptions unless the system has the authority and current data to do so. When the evidence is missing, it should say so and offer the next useful step.

Make human handoff a first-class path

Give customers an obvious way to reach support. Pass along the question, relevant context and what the bot already tried—without leaking unnecessary personal data. Define who owns urgent or sensitive requests and when the bot must stop answering. A smooth handoff matters as much as a polished automated answer.

Evaluate before expanding

• Test a set of real, anonymized questions in both supported languages.
• Check whether answers cite the right source and avoid unsupported claims.
• Measure unresolved conversations and handoff quality, not only answer volume.
• Retest after changes to prompts, models or source documents.
• Let support staff flag answers that need correction.

Should a chatbot answer every question?

No. Its scope should match the quality of its sources and the risk of being wrong. A bot that reliably handles a smaller set of tasks and escalates the rest can be more valuable than one that responds confidently to everything.