Intelligence and Reply Behavior

Guide intro

Reply behavior is where the product starts to feel real to your customers. A fast connection is useful, but the quality of the replies depends on the instructions, language direction, and reply mode decisions you set here.

The safest early approach is to keep the logic simple. Decide how the AI should sound, what information it can rely on, and when it should prefer text versus voice. Then review real conversations closely before expanding into more complex automation.

Instructions, language, and business context

Write the business rules in plain language. The AI works better when instructions are practical and specific rather than long and abstract. Include what the business offers, how it should describe pricing or availability, and which topics should trigger handoff.

Language direction matters too. If your customers switch languages often, define the preferred style clearly so replies remain understandable and consistent.

Best practices

• Use short practical rules instead of long marketing copy.
• Update instructions when product availability or support rules change.

Reply modes explained

Always Text keeps replies in text even when customers send voice notes. Always Voice biases the workflow toward audio replies. Customer Preference follows the customer's format more closely, while Text Preferred keeps text as the safer default even when voice messages arrive.

Text Preferred is usually the best starting point for teams that want cost control and easier review. Customer Preference can be useful when audio matters to the audience, but it should still be monitored carefully.

Steps

• Start with Text Preferred unless your business has a strong reason to default otherwise.
• Review where customers genuinely benefit from voice replies.
• Keep escalation rules short so support exceptions still reach a human quickly.

How to review reply quality

A reply mode should be judged by clarity, not novelty. Review whether customers understand the answer, whether important details are captured, and whether the team needs to correct the AI too often.

If the AI is too broad, tighten the instructions. If it is too hesitant, improve the business context it can rely on.

Related docs