AI

​

When to use this node

• Classify, summarize, extract, or draft from mapped workflow data.
• Turn unstructured text into a concise result before a human review step.

​

Example: Summarize a support email

{
  "connectionId": "ai-work",
  "prompt": "Summarize the customer request and list required follow-up:\n{{trigger.email.body}}",
  "systemPrompt": "You are a support triage assistant. Do not invent facts.",
  "temperature": 0.2,
  "maxTokens": 500,
  "timeoutMs": 90000
}

1. Choose a connection whose provider and data policy are appropriate.
2. Map only the context needed for the task.
3. Review response before using it in communication or another side effect.

​

Configuration

​

Workflow JSON

{
  "type": "ai_prompt",
  "version": 1,
  "configuration": {
    "connectionId": "",
    "prompt": "",
    "systemPrompt": "You are a helpful workflow assistant. Return a concise, useful response.",
    "temperature": 0.2,
    "maxTokens": 1200,
    "timeoutMs": 90000
  }
}

​

Inputs

​

Outputs

​

Execution behavior

• The local runner waits for the provider and returns response text, provider-reported usage, and model ID.
• Prompt, system instruction, and mapped context leave sndbox for the selected provider; the credential remains in the host vault.
• The editor catalogue declares response (string), usage (object), model (string) for mapping. The execution inspector can contain additional evidence fields; inspect a real result before selecting nested paths from object or any outputs.

​

Common issues

• Model output is untrusted data: validate structure and preserve human review for consequential actions.
• Reduce prompt size or raise the timeout within limits when requests consistently time out.

​

Where it can run

• Desktop local runner

​

Test and inspect

​

Related guides

• AI connections
• AI Builder