The Replify AI agent is a controlled customer responder, not a general chatbot. It receives the current customer message, recent conversation context, approved business inputs, and relevant knowledge, then returns one structured action: send, skip, handoff, or close.
Choose the response mode
Open AI Agent and select whether AI replies alone, automations reply alone, or response-claiming automations run before AI. Human-owned work and active guided replies remain ahead of this setting.
Describe the business clearly
Add a concise business description covering what you sell, who you serve, locations, languages, and important boundaries. Do not use this field as a dump of changing prices or inventory; put detailed and maintainable facts in the knowledge base.
Select assistant goals and tone
Choose the goals that reflect the actual job: support, sales, FAQ, booking, orders, or pricing. Select a tone your team would approve in Arabic and English. Replify is designed to match the customer's language and can use natural Saudi Arabic when appropriate.
Build the knowledge base
Add approved websites, PDF or Word files, Markdown files, and manual knowledge in AI Knowledge. Keep sources focused. Remove conflicting, expired, or duplicated policies. Add links the agent is allowed to share.
Define safe handoff boundaries
AI should hand off when a customer asks for a person, account access is required, a refund or exception needs judgment, information conflicts, the request is sensitive, confidence is low, or the required answer is not grounded. A handoff includes a customer acknowledgement, a concise internal summary, a reason code, and a human task.
Test before activation
Use the built-in knowledge test with realistic phrasing, misspellings, short follow-ups, Arabic, English, and mixed-language messages. Include cases that should answer, clarify, hand off, close, and stay silent. Review the matched sources and confidence rather than judging only the wording.
Monitor after launch
Review AI response runs, depleted reply balance, plan limits, provider failures, low-confidence decisions, human handoffs, and cases that reopened. Update knowledge when the business changes; do not try to solve missing facts by making the prompt more permissive.
Still need help?
Tell us what you were trying to do, the channel involved, and what you expected to happen.