A business chatbot should answer from an approved knowledge source, identify questions it cannot resolve and hand the conversation to a person. Begin with one bounded task, such as explaining product options or collecting an inquiry, before allowing it to change orders or make commitments.
Define the job before choosing a model
| Use case | Required input | Failure response |
|---|---|---|
| Product and service questions | Reviewed descriptions and current policies | Show the relevant source or ask the team |
| Lead qualification | Agreed questions, contact consent and routing | Offer the normal inquiry form |
| Order lookup | Authenticated user and authorized order API | Hide private records and route access issues to support |
| Document assistance | Approved document set and access rules | Say evidence is missing instead of inventing an answer |
Build a reviewed evaluation set
Collect real questions without personal data. Include questions with a known answer, ambiguous requests, missing knowledge, outdated policies and attempts to obtain another customer's data. Define acceptance criteria for source support, escalation, response time and cost per completed conversation. Repeat this evaluation when the model, prompt or knowledge changes.
Control cost and authority
Keep model credentials on the server. Limit input size, usage and external calls. Retrieve only content the user may access. Treat retrieved text as untrusted information; business rules and permission checks belong in the application. Require confirmation for actions that affect orders, payments or customer records.
Provide a usable human handoff
The visitor should be able to reach a person without repeating the whole conversation. Send the consented contact details, relevant messages and unresolved question to the agreed queue. When the AI provider is unavailable, display the regular contact route.
Review AI integration services or send a chatbot brief with the target channel, approved knowledge, example questions, privacy constraints and handoff destination. Model selection follows the measured task requirements.