Let AI write the DM reply.

Steps

The fallback is required. It sends when you need a safe default.

A connected assistant can read results and create a fixed comment rule. AI replies stay on this page. The setup is on /how-to/connect-your-agent.

• Open the rule and enable AI replies.
• Write the goal in one sentence.
• Paste prices and links into the knowledge note.
• Set the fallback text.
• Run a test draft. It reserves 1 AI credit and may consume it even if the provider fails.

Prepare facts, boundaries, and static text together

Enable AI only where a generated answer improves on a fixed response. Growth or Business is required, and a goal plus fallback must be configured. Put maintained prices, policies, service areas, and destinations in the knowledge note. Describe how to handle a missing fact instead of asking the model to close every sale. Keep the saved static message useful too: noise or exhausted monthly credits can return static text rather than a generated answer.

A workflow to try

For a shipping rule, state the service area and current shipping explanation in the note. Set the goal to answer the shipping question and provide the relevant support URL. Write a fallback that acknowledges uncertainty without inventing an arrival date. Preview a routine question, a request about an unsupported destination, and a message telling the model to ignore the note. Each preview reserves one credit and may retain it on provider failure. Review the output for facts, tone, and actual destination, not just fluency.

Checks before you change the setup

After a controlled launch, inspect conversations and the displayed credit balance. Topic matching can reserve credits separately from generation. Keep sensitive disputes and exceptions with a person; do not advertise an automatic takeover timer that this review did not verify. An AI answer remains an API send subject to the messaging window and connection. If dispatch fails, separate permission and transport problems from the model wording. Turning AI on cannot repair a closed thread.

Keep a factual review log for the first rule

For each representative preview, note the question type, supplied fact, expected boundary, generated answer, and anything corrected. Include at least one answer that should admit uncertainty. After launch, compare real replies with the same boundaries because recent history can change the context. If an incorrect answer appears, pause or narrow the rule, repair the relevant note, and test that case before resuming. Do not turn a corrected sample into a claim that all future answers are guaranteed. The review record should explain what the model is allowed to say and what remains a person’s decision, rather than only recording that the output sounded natural.