TopolearnAI steps let a Zap handle messy text (emails, notes, reviews) that fixed rules can't understand.
AI by Zapier is a built-in app that adds an AI step to any Zap. Inside it, a large language model (the kind of AI behind chat assistants) reads text from earlier steps and does what your prompt asks.
The step has two panels. In Configure you write the prompt (your instructions to the model). Type / to insert fields from earlier steps, such as the email body. Templates exist for four common jobs: Summarize, Write, Classify and Extract.
Then you define output fields: the named results you want back, each with a type and an optional description. Example for support emails:
category(text): one of Billing, Bug, Questionurgent(true/false)summary(text): one sentence
Because the answer comes back as separate fields, later steps can map them, or a filter can check urgent. This is far more reliable than asking for one paragraph and hoping. You can also have Zapier suggest fields from your prompt.
In the Preview panel you test the step on sample data before saving.
You choose a model tier (Standard, Advanced, Premium) or connect your own AI provider account. Stronger tiers use more tasks per run, so pick the simplest tier that does the job well. Specific model names change often; the tier list in the editor is the current reference.
AI output can be wrong or invented (a hallucination). Keep the step's job narrow, list allowed values, and check results in Zap history before trusting it fully.
Remember
- An AI by Zapier step = a prompt with mapped data + output fields
- Output fields turn free text into values later steps can use
- Pick the simplest model tier that works, and check results
Watch out for
- Asking for one long free-text answer and then trying to split it in later steps instead of defining output fields.