AI agent

The AI agent step calls a large language model (LLM) once per record. Use it to draft messages, classify leads, extract fields, or summarize context before a channel send.

Prerequisites

1. Add an LLM under Integrations → AI providers (OpenAI, Anthropic, or Mock) and save the API key.
2. On the AI agent step, select that provider.
3. Choose a model from the dropdown (current models only). Use Custom only if you must paste a special API id.

Retired model ids (for example old Claude Haiku dates) return provider errors. Prefer the dropdown.

Screenshot needed

Integrations → AI providers → LLM list with one configured provider.

Where: Integrations → AI providers

Save as: src/assets/screenshots/19-ai-providers.png

Screenshot needed

AI agent step: provider selected, model dropdown, role and task fields.

Where: Spinner builder → AI agent

Save as: src/assets/screenshots/20-ai-agent-step.png

Agent role & rules vs task for each record

These two fields are easy to mix up.

Field	Meaning	Changes per contact?
Agent role & rules	Who the AI is and how it must behave (tone, policies, “never invent balances”)	No — same for every row
Task for each record	What to do with this contact’s data	Yes — uses templates for metadata fields

Example role:

You are a polite collections assistant. Be brief. Never invent account balances or legal threats.

Example task:

Customer name and phone come from metadata fields. Draft a short payment reminder in plain language.

Output modes

Mode	Result
Text body	One string, usually stored as ai_response
Key–value fields	JSON object with keys you list
Classification	Single label
List	Array of strings

For text mode, later steps read the result with a template for output.ai_response (or your custom output field name).

Timing and failures

• Calls can take up to about two minutes per record.
• Runtime logs show AI agent calling LLM then AI agent completed or a clear error.
• Failures appear on the pipeline journey for that record.

Next

• Log step
• AI providers