Choose a bounded use case
Start with low-risk, verifiable questions such as opening hours, available services or access instructions. List what is out of scope: complaints, contractual exceptions, diagnoses and promises about price or timing. A request that needs individual judgement should have an obvious route to a person.
Prepare a reviewed knowledge base
Collect approved pages, policies, answers and owner contacts. Remove outdated information and record when each source was reviewed. Separate facts the assistant may explain from actions a person must confirm. MaviBot's official AI assistant page describes an AI assistant and conversation features; verify which sources, permissions and limits are available on your plan.
Write rules for uncertainty
Tell the flow to ask for clarification when details are missing and to admit when it cannot find an approved answer. Confidence is not evidence of accuracy. For sensitive requests, payments, personal data or contract changes, escalate the conversation and retain only the context the team needs to continue.
Design the human handoff
Customers should be able to request a person without guessing a secret phrase. Send the original question, confirmed details, unresolved point and agreed next action to the team. Show a waiting state while the case is unreviewed; recording a handoff is not the same as resolving the request.
Test normal and failure cases
Create a matrix with correct questions, spelling errors, two intents in one message, outdated information, an out-of-scope request and an explicit request for a person. Ask someone who did not build the flow to test it on a phone. Record the expected answer, received answer and recovery path.
Run a measurable pilot
Start with a small group and a defined period. Count reviewed answers, escalations, unanswered questions and manual corrections as pilot observations, not performance promises. Compare maintenance effort, handoff quality and total cost with a simple FAQ or manual support path.
Questions and answers
Should an AI assistant answer every question?
No. Limit the first use case to verifiable questions and keep a human path for exceptions, complaints, sensitive data and individual decisions.
How should I evaluate MaviBot before serving customers?
Test a small knowledge base, uncertainty rules, handoff, permissions, export and current plan limits. The official page describes features but does not replace testing your process.