Skip to main content
An AI Action runs one model call inside an otherwise deterministic Flow. Use it when you need language understanding at a single step, then return to scripted branches.

​
AI Action vs AI Agent

​
Common uses

  • Classify free text (“billing” vs “shipping”) → Condition
  • Extract structured values into custom fields (order ID, city)
  • Summarize a long user message before human handoff
  • Score lead quality and tag the contact

​
Build the step

1

Capture input

Use User Input or the latest inbound message as the model input.
2

Add AI Action

Configure the prompt, model/provider settings available in your workspace, and where to store the result (custom field).
3

Branch on the result

Add a Condition (or Go to Flow) that reads the saved field.
4

Add a fallback

If the model returns an unexpected value, send a clarification message or hand off to a human / full Agent.
Keep prompts narrow and output formats strict (for example Reply with exactly one of: billing, shipping, other). Tight outputs make Conditions reliable.

​
Cost and latency

AI Actions consume model tokens and add latency versus pure scripted blocks. Use them where language flexibility matters; prefer Quick Replies for closed choices.

​
Next

AI Agents

Full conversational agents when you need multi-turn reasoning.

Recipes

Example Flow patterns that mix script and AI.