AI
AI auto-reply in the DM.
The person messages you. The rule matches. The model writes a reply from the goal and the note. The fallback is there when you want a fixed line instead.
3 minute read
Limits
The reply still has to sit inside the messaging window. AI does not message someone who never wrote to you.
Noise is skipped: empty text, emoji-only text, and long repeated characters.
Answer variations without rewriting the offer
A fixed keyword is ideal for a repeated link request. An AI-enabled reply can help when different customers ask related factual questions in different words. Give the rule a narrow goal and maintained knowledge instead of a broad instruction to sell anything. The model can use incoming text and recent conversation history in the live path, so avoid promising an isolated response with no prior context. It remains a text-generation layer inside the normal account and messaging rules.
A workflow to try
A salon can provide its service list, preparation instructions, and booking URL in the note. The goal might be to answer preparation questions and direct booking to that URL. Test a basic service question, a request for an unavailable appointment, and a medical-advice question. The output should not claim to reserve a slot or diagnose a condition. Keep a safe fallback and saved static message, then read the first actual conversations to see whether the note needs clarification.
Checks before you change the setup
If a reply is not generated, inspect AI entitlement, enabled configuration, noise detection, daily cap, and remaining credits. Monthly exhaustion returns the static template in the implemented service; daily-cap and provider-error paths can use fallback text. If generation succeeds but delivery does not, inspect the messaging window and send result separately. Neither an instruction nor a credit balance opens a cold thread. Pause overlapping rules deliberately when a person needs to handle a sensitive exchange.
Distinguish answering a question from completing a task
The reply can explain a service and point to an authorized destination, but its wording should not imply a calendar, payment, or order action happened without evidence. Preview requests phrased as commands, such as book this now or refund my order, as well as ordinary information questions. Keep the expected boundary in the instructions and fallback. Read the live response beside earlier thread messages so an answer does not accidentally contradict something a person already explained. When a customer needs a decision, coordinate human handling deliberately. A concise correct limit is more useful than an unsupported confirmation designed to sound helpful.
Questions
Yes. Use a fixed keyword for the offer and AI for questions.
No. Supply and maintain the facts in the pasted knowledge note. This content does not claim a website scanner or automatic price synchronization. Update the note when an offer changes and test representative questions before continuing the campaign.
No. Credit availability allows model work; account connection, recipient eligibility, and the messaging window govern delivery. Investigate generation and dispatch separately so an ineligible send is not mistaken for weak AI wording.