Set Up AI Automation
Start with a website and your own OpenRouter key. Follow this guide from a first test reply to advanced knowledge retrieval.
Configure your assistant in the dashboard without building an AI integration yourself. Start with replies disabled, review your knowledge and test before enabling automatic replies. OpenRouter is the supported provider; model access and charges belong to your OpenRouter account.
1. Pair your number and select a device
Create a WAStack account and connect your number using Devices and WhatsApp Linked devices. Open AI Automation in the dashboard and choose the device. Automatic replies require an active plan with chatbot access.
2. Connect your own OpenRouter key
In Setup & persona, save your OpenRouter API key. WAStack encrypts the saved credential and does not return it to the browser. Choose a supported text reply model from the live catalog or enter its exact OpenRouter ID. Available free and paid models can both be used. A blank key field keeps the saved key; an explicit removal revokes it from this device.
3. Add knowledge from your website
In Knowledge & files, enter your public website and optionally specific page URLs, one per line. Keep sitemap discovery enabled to find more pages and choose a limit of up to 10 pages per import. Review extracted titles, text and source links, deselect unwanted pages, and add the reviewed pages. Save settings. Imports are snapshots and require review; public static text is supported, while login-only, PDF or JavaScript-only pages may need manually entered facts.
4. Give the assistant your business voice
Set the business name, personality, tone and reply instructions. Add factual products, services, opening hours and policies as editable knowledge entries. Explain when the assistant should ask questions or hand over to your team. Upload approved images or documents only if you want the assistant to share them; uploaded documents are not automatically read as knowledge.
5. Optional: use embeddings for relevant facts
Enable Retrieve relevant knowledge, choose an embedding model from the authenticated catalog and save. Click Index saved knowledge. Indexing batches your enabled facts and caches vectors. After editing facts or switching embedding models, index again; unchanged chunks are reused. Adjust relevant excerpts, knowledge character budget and minimum semantic similarity if needed. Incomplete indices use bounded keyword matching for missing chunks; previews show full, hybrid, partial hybrid or lexical retrieval.
6. Generate a test reply
Enter a customer question and click Generate test reply. This uses saved settings and your OpenRouter credits without sending a WhatsApp message. A preview starts a fresh conversation; live replies use configured recent conversation history. Inspect the response, selected sources and separate question-embedding and reply-generation costs. Ask missing-knowledge questions as well as easy ones. Reasoning models may need a larger output token allowance.
7. Enable replies and keep control
Enable AI assistant and Automatic replies, then save. AI replies to new direct customer messages on this device. When AI is enabled, turning automatic replies off also suppresses keyword chatbot fallback; disabling AI returns to existing keyword rules. Pause a conversation in Conversations for human support and resume when ready. Replies use WhatsApp text formatting and approved media; official Meta interactive buttons, lists and URL buttons are not generated.
8. Review usage, privacy and limits
Usage & logs separates knowledge embeddings, question embeddings and generation operations. OpenRouter costs are separate from WAStack plans and message allowances. A free reply model does not make paid embeddings free. Provider-reported missing cost is shown as unavailable. Daily AI request limits count previews, indexing batches, question embeddings and generation claims and reset at 00:00 UTC. Full prompts and results are off by default; enabling them stores sensitive business/context content in diagnostics. Turning a capture flag off clears that diagnostic content. Knowledge and conversation memory are separate; set previous messages to zero to clear and stop stored memory.
Ready to implement?
Start building with WAStack APIs. Connect your WhatsApp in under 5 minutes.
Start FreeOn this page
Frequently asked questions
Common questions about set up ai automation.
Related documentation
Browse all documentation
From getting started to advanced API patterns, find everything you need to build WhatsApp automation.