AI in Automation: the AI builder, the AI step and what it costs

AI helps in two places: it can build a flow from your description, and an AI step can do part of the work inside a flow — read a message, decide, write a reply.

Build a flow by describing it

Create with AI on Flows opens a new flow with the AI Flow Builder beside it; in any flow, the ✦ button in the toolbar opens it.

1. Describe what should happen, in your words — When a deal is won in CRM, e-mail the customer a thank-you — or pick one of the suggestions.
2. The builder adds and fills the steps on the canvas (Flow updated on canvas). When a step needs an app you have not connected, it shows a Connect card.
3. Ask for changes the same way — wait a day before the e-mail. New conversation starts over.
4. Check its work before you publish: open each step, test it, then Save and Publish: The flow editor.

The builder only uses the apps that can be connected here. The Knowledge Base teaches it other services' documentation — see below.

The AI step

AI Step (in the AI group) runs an AI model inside the flow, with Run AI Agent:

• System Prompt — its instructions: who it is, what to do, what never to do.
• User Message — what to work on this time, with values from earlier steps, such as {{trigger.body.message}}.
• Model Spec — the model; leave the default, the one Monfri uses.
• Tool Whitelist — the actions it may call, written as connector.action, for example crm.search-contacts or crm.create-note. Empty means it can only answer. It really does what the tools do — a write tool writes — so allow only what it needs.
• Budget — maxRounds (8 by default) caps how many times it may think and call a tool.
• Output Schema — optional, for an answer in a fixed shape.

The next steps get finalMessage (its answer), structuredOutput, toolCalls, status and rounds.

What you put in the prompts, and what its tools return, goes to the AI model of Monfri's provider; pass only what the step needs. The AI step's use counts in your plan as AI credits: Usage and limits.

OpenAI and Anthropic with your own key

The OpenAI and Anthropic Claude connectors call those providers with your own API key, saved in Connections: their use is billed to your account with them and does not count in your plan: Connectors.

AI Usage, AI Runs and the Knowledge Base

• AI Usage — the AI used by this workspace's Automation: API Calls, Total Tokens, Avg Duration, Total Cost, Daily Usage, Usage by Model, and the Conversations, each with its prompts and answer.
• AI Runs — the requests Monfri's AI chat handled for the workspace, where your workspace has the chat, each with its plan and steps.
• Knowledge Base — documentation the AI builder may read. Add Documentation → Documentation URL and Name → Start Crawling: it follows the links within that site, up to 200 pages, then Review Crawled Pages lets you choose what to keep before it is indexed.

If something doesn't work

The AI step stops without an answer

It used its Budget — every round, without finishing. Make the instructions shorter and clearer, or raise maxRounds. The run's page shows each round and tool call: Runs.

AI step lists tools that do not exist in the connector registry: …

A name in Tool Whitelist is not a real connector.action. The first part is the app — crm, cdp, push-platform, docs — and the second the action, in small letters with hyphens: Create Note is crm.create-note, Search Contacts is crm.search-contacts. A few differ from their label — Qualify Lead is crm.convert-lead, Enroll Contact in Sequence is crm.enroll-in-sequence. Ask the AI builder when unsure.

The AI builder made a step I did not ask for

Tell it so, or delete the step yourself. Nothing it builds runs before you publish.