Intent routing
Before doing any work, Promptly figures out what kind of question the visitor is asking and sends it down the right path. This is intent routing, and it's what keeps the bot fast and cheap.
🖼️ [Image] — A diagram showing a message branching into product, knowledge, small talk, and escalation paths.
The categories
Every incoming message is classified into one of a few intents:
- Product — the visitor wants something from your catalog ("show me red sneakers under $80"). Routed to the planner with product tools only.
- Knowledge — an informational question ("what's your return policy?"). Sent straight to your knowledge base and answered, skipping product tools entirely — the fastest path. See How retrieval works.
- Mixed — needs both catalog and knowledge ("do you have waterproof boots, and what's the return window?"). Routed to the planner with all tools.
- Small talk / unknown — greetings, chit-chat, off-topic. Answered with a brief, friendly reply that gently steers back to what the bot can help with — no catalog search, no knowledge lookup.
- Escalation — the visitor asks for a human or is clearly frustrated. Handled before any planning, by your Escalation settings.
Why routing matters
- Faster — a greeting doesn't trigger a database search; a pure policy question skips product tools. The bot does only the work the question needs.
- Cheaper — skipping unnecessary steps means fewer tokens per message, which lowers your monthly cost. See Languages & cost tracking.
- More accurate — each path is tuned for its job, so product queries get product logic and informational queries get your documentation.
Behind the scenes
Routing runs on a fast, low-cost classifier and caches its decisions, so repeated or similar questions are routed almost instantly. When the bot is unsure, it errs toward the broader (Mixed) path so nothing useful is missed. Small-talk replies and language follow the visitor automatically — you don't configure routing directly; it works out of the box.
🎬 [Video] — The same bot handling a greeting, a policy question, and a product search, each resolving differently.