Adding AI Steps to n8n and Make Workflows: A Practical Guide

To add an AI step to an n8n or Make workflow, use a node or module built for the job rather than a general chat prompt where you can: in n8n, nodes such as Text Classifier, Information Extractor or an LLM chain; in Make, the Make AI Toolkit or a provider app such as OpenAI or Anthropic Claude. Give the step one narrow job, ask for structured output, check the result before anything acts on it, and set retries and an error route so a slow or failed AI call never loses the record.

This article used to compare n8n with Make. That comparison now lives, with Zapier added, in n8n vs Make vs Zapier. The overall shape of a reliable workflow, including storing the input before anything else happens, is in how to build an AI-powered workflow. This page covers the practical part in between: putting a language model call inside a workflow on either platform.

What an AI step in a workflow is good for

An AI step earns its place when the input is unstructured text that a fixed rule cannot read reliably. AI vs traditional automation covers where that line falls. The common jobs:

Names are as listed in the n8n documentation and the Make AI Toolkit documentation at the time of writing (October 2026). Both platforms add and rename AI features often, so check the current names in your own account.

Adding an AI step in n8n

1. Add credentials for your model provider, such as OpenAI or Anthropic, in n8n's credentials area. Keep keys there, never pasted into a node's text.
2. Choose the root node for the job. Text Classifier and Information Extractor are built for the two most common tasks. Each needs a chat model sub-node attached, which is where you pick the provider and model.
3. Map the input with an expression pointing at the field from the previous step, such as the message body from your form trigger.
4. Define the output. For the classifier, list each category with a short description, and add one for unclear cases. For the extractor, describe each field or supply a JSON schema; the Structured Output Parser does the same job for LLM chains.
5. Set the node's error behaviour. In the node settings, turn on Retry On Fail with a sensible wait between tries, and set On Error to Continue (using error output) so failures go down a branch you control.
6. Test with pinned data. Pin a set of real, anonymised examples to the trigger so you can rerun the AI step without sending new submissions.

Adding an AI step in Make

1. Choose between the AI Toolkit and a provider app. The AI Toolkit's ready-made modules cover classifying, extracting, summarising and translating, and can run on Make's own AI provider or on your own provider connection. The OpenAI and Anthropic Claude apps give you full control of the prompt and model.
2. Map the input from the earlier module into the text field.
3. Shape the output. With the AI Toolkit, define the categories or the items to extract. With a provider app, ask for JSON with named fields, then pass the response through the JSON app's Parse JSON module so later modules can map each field.
4. Add an error handler route to the AI module. Make offers five directives — Resume, Ignore, Break, Commit and Rollback — set out in its error handlers guide. Resume lets you supply a fallback value, such as the category unclear. Break stores the run as an incomplete execution so it can be retried.
5. Run once with sample data and inspect each module's output before switching the scenario on.

Prompts a workflow can rely on

A prompt in a workflow is not a conversation. It runs unattended thousands of times, so it has to behave the same way every time.

• One job per call. Classify, or extract, or draft. Not all three at once.
• A closed list of answers. Name the allowed categories, and include one for unclear or none of these.
• An exact output format. Field names, types, and what to return when a value is missing.
• A couple of examples, including an awkward one.
• Low randomness, where the platform exposes a temperature setting.

A classification prompt might read:

Check the output before anything acts on it

Models occasionally return a category that is not on the list, a malformed date or an empty field. Treat the AI step's output like any other untrusted input.

• Validate it. An If node in n8n or a filter in Make checks that the category is one of the allowed values and the required fields are present.
• Send unclear results to a person, not to a default bucket.
• Let rules act, not the model. Once you have a category, deterministic steps decide who to notify and what to create.
• Keep a person in front of anything a customer sees, at least until you have reviewed a few weeks of drafts.

What an AI step costs

There are two bills. The platform charges in its own unit: n8n counts a whole workflow run as one execution however many nodes it has, while Make charges credits per module, and the AI Toolkit's credit use varies with the tokens consumed and the model chosen when it runs on Make's provider. If you use your own key, the model provider bills you separately by tokens in and out.

For short classification or extraction calls at small-business volumes, the model cost is usually small next to the platform fee and the time spent building and maintaining the workflow. Long prompts, large attachments and drafting with the most capable models are where it climbs. Test with your real message lengths, check the usage in your provider dashboard after a week, and use AI automation cost and ROI to put the numbers together.

Personal data in prompts

Every AI call sends data to a third party. Send only what the step needs — the message text, not the customer's whole record. Check the provider's data-processing terms, including whether inputs are retained or used for training, and record the processing in your privacy notice. GDPR and AI automation for UK businesses has the checklist.

A worked example: the same step in both tools

Say a heating engineer's website form posts to a webhook. The workflow stores the submission first, then needs to know which service it is about and whether it is urgent.

In n8n: a Webhook node, then a Google Sheets node to store the row, then a Text Classifier with five categories and a chat model attached, retries on and the error output routed to a Slack alert. A Switch node sends each category to the right person, and unclear goes to the owner.

In Make: a custom webhook, then Google Sheets, then the AI Toolkit's Categorize Text with the same categories and a Resume error handler supplying unclear as the fallback. A router sends each category on, with a filter on each route.

Same design, same safeguards, different screens. The platform matters less than the order of the steps and what happens when the AI call fails, which is the argument of the AI business automation guide as a whole.

Next step

Pick one workflow you already run and find the step where a person reads text to decide what happens next. That is your first AI step. If you want it built with the safeguards above, the AI automation service does this on n8n or Make, whichever suits the rest of your systems. If you run GoHighLevel, look at GoHighLevel workflow examples first, because some of this can happen inside the CRM.