A reliable answer is more than fluent text. YundaDesk selects information that fits the question. When evidence is missing or risk is high, the AI should clarify, state a limitation, or hand off instead of guessing.

Four sources for an answer#

• Knowledge base: Stable facts such as product information, policies, service procedures, and FAQs.
• AI skills: Reusable response methods, business steps, and governed actions.
• Customer context: Confirmed preferences, history, and information that applies only to the current customer.
• Live capabilities: Orders, inventory, shipping, or other changing data from connected apps.

Knowledge cannot replace live data, and customer memory cannot override an official policy. When the workspace lacks a required capability, the AI should explain what is missing.

Safe general questions#

The AI can directly answer simple utility questions and safe public-fact questions when they do not depend on merchant knowledge, customer-private data, or external live data. A tool is not required first, and a question is not refused merely because it is unrelated to the merchant's business.

For relative time questions such as “today” or “what time is it,” the AI uses only a reliable visitor time zone supplied by the channel. If the visitor time zone is unknown, it asks first. Live weather, prices, and news still require a reliable live source; the AI does not guess the current result from old knowledge.

Verify an answer#

1. Ask a real customer question in the Playground.
2. Open the answer details and confirm that the correct knowledge, skill, or customer context was used.
3. Check that the answer does not add promises or numbers absent from the source.
4. Test a rephrased version and one unsupported question.
5. Before enabling a real channel, confirm that unsupported questions lead to clarification, a clear limitation, or handoff.

When the AI should not answer directly#

Refund commitments, price changes, identity or permission actions, and requests missing live order, inventory, or shipping data should not be completed by guessing. When a customer explicitly asks for a person, automation should not delay the handoff.

A temporary AI or integration failure does not mean that handoff occurred. Check the current reception state and let the team take over when needed.

Improve an unreliable answer#

Correct the knowledge when the source is wrong. Create or improve an AI skill when a reusable handling rule is missing. Correct customer memory when the information applies only to one customer. For a specific wrong answer, use Correct AI to create a learning suggestion, then review and test it before activation.

Do not hide the issue by adding conflicting content. Retest the original question, a rephrased question, and an unrelated question after the change.