AI steps

AI steps are steps like any other: fields in, data out, an error output when something goes wrong. The OpenAI and Gemini families are the same steps on purpose, so changing model is changing one step, not rewriting the workflow.

The four verbs

Ask

A prompt in, an answer out, with standing instructions and a temperature you can hold at zero. OpenAI, Gemini and Ollama.

Classify

Your own categories. The result is a value you can branch a line on. OpenAI and Gemini.

Extract

Name the fields you want and get them back as fields, not as prose. OpenAI and Gemini.

See, speak and hear

Read an image, generate one, text to speech, and transcribe audio. OpenAI, Gemini and Whisper.

Step	OpenAI	Gemini	Ollama	Whisper
Ask	✓	✓	✓
Classify	✓	✓
Extract	✓	✓
Images (read and generate)	✓	✓
Speech	✓
Transcribe	✓			✓

Your keys, your bill

An AI step takes the provider's API key on the step itself. Calls are made with your key, so they bill to your account with the provider, at the provider's prices, and you can use any model your key has access to.

Ask

Field	What it does
Prompt	What to ask. Usually mixes text and variables: `Summarise this in three bullet points: {{http_request.data.body}}`
System instruction	Standing instructions that apply to every run: tone, format, what to refuse
Model	Which model answers
Temperature	How much the answer may vary. Hold it at `0` for answers that should be the same every time
Max tokens	An upper limit on the length of the answer

It returns text (the answer), model, finishReason (stop for a complete answer, length when it ran out of room) and tokens, what it cost in tokens.

Classify, then branch

Classify returns one of the categories you gave it. Put a condition on each line out of the step and every category gets its own route:

Mail received ──► Classify ──┬─ category equals "invoice"  ──► Extract ──► Sheets: add row
                             ├─ category equals "support"  ──► Zendesk: create ticket
                             └─ category equals "other"    ──► Outlook: move to folder

When a model fails

A model step has an error output like every other step. A rate limit is then a branch you drew, not a run that stops halfway: route the error to the same step on another provider, or to a message that tells someone. See When a step fails.

To let an AI assistant run your workflows, rather than putting a model inside one, see Assistants and MCP.