What an AI step does inside a workflow

An AI step is an ordinary step with a language model inside. Text goes in, the model writes text back, and the workflow carries on with that text as data.

What you need

• A job that works on text: sort it, shorten it, pull out names and amounts, or draft a reply.
• An account or connection for the model. Some tools supply one and some ask for your own.
• A written prompt, which is the instruction the model receives along with your text.
• A rule for who checks the reply before it does anything real.

What goes in

Two things go in. The first is your prompt, such as "Sort this message as billing, delivery or other." The second is data from earlier steps, such as the message body. You drop the earlier field into the prompt the same way you'd fill any other step's field. The page on mapping walks through that.

Whatever sits in that prompt leaves your workflow. Make says every AI feature in Make sends your request to an AI provider for processing.¹ Read your own tool's documentation to see where a prompt goes.

Treat the prompt box like a public postcard. Passwords, API keys, customer records and card numbers don't belong in it, unless a written rule at your business says they can go. If you aren't sure, ask the person who owns that rule before you build.

What comes out

Mostly text. It might be a label, a short summary, a list of fields or a draft email. The workflow treats it as one more field that later steps can use.

Because the output is just text, a later step can't tell a good answer from a bad one. It simply receives whatever arrived.

What the model is doing

Wikipedia says generative pre-trained transformers (GPTs) are a type of LLM that is pre-trained to predict the next word.² In plain terms, the model is built to produce text that fits what came before it. A fluent reply isn't the same thing as a checked fact.

That's why wrong answers can sound fine. Wikipedia describes hallucinations as cases where models generate syntactically fluent text that appears factually sound, but is internally inconsistent with training data or factually incorrect.²

Fixed step or agent

The tools offer two broad shapes, and n8n's glossary names both.³

• A chain calls a language model and other parts in a fixed sequence. Unlike an agent, a chain doesn't decide which steps to take.³
• An agent uses a model to interpret a request and decide which actions to take, such as calling tools.³ n8n's agent node works by connecting a chat model and one or more tools, and the agent decides which tools to call.⁴

For a first workflow, the fixed shape is easier to follow, because you know in advance which steps will run.

What it costs to run

Zapier's task page, checked on 4 November 2026, lists AI by Zapier with its own rate, which depends on the model tier you select.⁵ Look up the current rate on the maker's own page before you plan volume.

Where the data goes

If the tool lets you choose a region or bring your own connection, read what the maker says about it. Make notes that bringing your own AI provider connection removes Make's EU guarantee entirely.¹ Whether a region matters depends on your country's rules and your own policy, and for legal questions you should go to the official source.

Put a check after the AI step, not just a hope. The page on filters and branches covers filters and a human approval step.