n8n Now Builds for You: How to Use the AI Assistant Workflow Builder

n8n has added a chat-based workflow builder directly inside the product. In the launch video, the team describes it as n8n Assistant: a builder agent that lives inside the n8n interface, creates workflows from natural language, lets you set parameters in chat, handles credentials through n8n's normal credential screens, and can test the workflow while you watch it run.

That last part matters. A generic AI coding tool can generate workflow JSON or call an MCP server, but it usually separates the conversation from the visual workflow. n8n's Assistant keeps the chat, canvas, credentials, test runs, and workflow editor in the same place. Much less waving a wand at a black box. Sensible, at last.

This guide explains what the new AI Assistant is, what it is not, and how to use it safely.

Sources used for this guide:

• n8n Now Builds For You (n8n Assistant)
• n8n docs: Use AI Assistant (Preview)
• n8n docs: Use AI Workflow Builder
• n8n docs: Use n8n MCP server
• n8n pricing and AI Assistant credit notes

What n8n Assistant Actually Does

n8n Assistant is a preview AI agent inside n8n that helps build and manage workflows from natural language.

According to n8n's docs, it can help you:

• Create workflows from a plain-English automation request
• Edit workflows by adding nodes, changing logic, or adjusting configuration
• Test and troubleshoot workflows by checking errors and suggesting fixes
• Work with credentials without exposing your secret values in the chat
• Create or update supporting resources, such as Data Tables, when a workflow needs them
• Research approved websites when web access is enabled, after asking permission for the domain
• Help with instance tasks, such as renaming or publishing workflows

The result is not a proprietary AI-only artefact. It is a normal n8n workflow that you can inspect, edit, test, and publish like any other workflow.

Availability: Cloud, Preview, and Self-Hosted Reality

There is a small but important distinction here.

The current n8n documentation says AI Assistant is available on n8n Cloud only and is still in Preview. The pricing page says AI Assistant preview is included in n8n Cloud plans, with monthly AI credit allowances, and that self-hosted AI Assistant support is coming soon with bring-your-own API key support.

The launch video also notes that self-hosters can run it by bringing their own API keys during preview, but that the process is more involved and should become easier later.

So the practical guidance is:

Because this is a preview feature, assume availability, UI, credit usage, and supported actions can change. Build with it, but do not blindly trust it in production.

n8n Assistant vs AI Workflow Builder vs n8n MCP

n8n now has several AI-related ways to build workflows. They overlap, but they are not the same thing.

Use AI Assistant when you want the safest, most n8n-native experience. You stay close to the workflow canvas, credential cards, and execution results.

Use n8n MCP when your work starts in an external AI tool. For example, if you are already planning a feature in Claude Code or Codex and want that agent to create or modify the matching n8n workflow, MCP is the bridge.

Do not confuse the two. One is an assistant inside n8n. The other lets outside tools talk to n8n. Similar destination, different control surface.

Before You Start

You need three things before using the AI Assistant properly:

1. Access to an n8n instance with the feature enabled
2. Permission to the project, workflows, credentials, and resources involved
3. A clear description of what the workflow should do

The Assistant uses your n8n user permissions. It can only access workflows, credentials, and resources that your user can access in the selected project.

Do not skip that permissions check. If the Assistant cannot see the project or credential it needs, you will waste credits debugging a permissions issue disguised as an AI issue. Tedious, avoidable, and very common.

How to Open n8n Assistant

The launch video shows the entry point like this:

1. Open n8n.
2. Go to the plus button.
3. Choose New AI Chat.
4. Describe the workflow you want to build.

The exact label may shift during preview, but the important idea is that you are starting a chat inside n8n, not pasting instructions into a separate AI tool.

The Prompt Format That Works

n8n's docs are clear: specific prompts work better. Include the trigger, services, data, success behaviour, failure behaviour, and whether the Assistant should ask before publishing.

Use this structure:

Create a workflow that:

Trigger:
- <when the workflow should start>

Apps/services:
- <apps, APIs, databases, or tools involved>

Data:
- Read <source data>
- Transform <what needs changing>
- Send or save <destination data>

Success behaviour:
- <what should happen when it works>

Failure behaviour:
- <what should happen when it fails>

Safety:
- Ask me before publishing the workflow.
- Do not delete, overwrite, or send external messages without confirmation.
Click to copy

That is much better than:

Make me an invoice workflow.
Click to copy

Vague prompts produce vague workflows. Shocking, apparently.

Example 1: Build an Invoice Collection Workflow

A good first test is a workflow that checks for invoices and stores structured data.

Prompt:

Create a workflow that checks Gmail every weekday at 8:00 AM for unread emails with PDF invoice attachments.

For each invoice email:
- Save the PDF attachment to Google Drive in a folder named Invoices/YYYY-MM.
- Add a row to a Data Table with sender, subject, received date, Drive file link, and processing status.
- Mark the email as read only after the file is saved and the Data Table row is created.

If saving to Drive fails:
- Leave the email unread.
- Add an error row to the Data Table if possible.
- Send me a Slack message with the email subject and the error.

Ask me before publishing the workflow.
Click to copy

Why this works:

• The trigger is specific.
• The workflow has clear data movement.
• Success and failure behaviour are different.
• The Assistant knows not to publish without confirmation.

When the Assistant starts building, watch for these things:

• Does it choose the correct Gmail trigger or polling pattern?
• Does it handle multiple attachments?
• Does it avoid marking the email as read too early?
• Does it ask for the correct Gmail, Google Drive, and Slack credentials?
• Does the Data Table schema match the fields you asked for?

Example 2: Debug a Failed Workflow

The Assistant is not just for creating new workflows. It can also troubleshoot existing ones.

Prompt:

Debug the latest failed execution of this workflow.

First explain:
- which node failed,
- the exact error,
- what data reached that node,
- and the most likely cause.

Then suggest a fix before changing anything.
Do not publish or overwrite the workflow until I approve the change.
Click to copy

This is the correct pattern for debugging: explain first, change second.

If you ask it to "just fix it," do not be surprised when it changes things you did not mean to change. That is not clever automation; that is handing it the keys and hoping for manners.

Example 3: Update an Existing Workflow

Prompt:

Update this workflow so failed Shopify orders are sent to Slack and then retried once after 10 minutes.

Requirements:
- Keep the existing successful-order path unchanged.
- Add clear error messages to Slack with order ID, customer email, failed node, and error message.
- Retry the API request once after 10 minutes.
- If the retry fails, stop and leave the execution marked as failed.
- Ask before publishing.
Click to copy

Key detail: tell it what not to change. AI builders are often too eager to reorganise a working workflow. You want a surgical update, not an interior design project.

Handling Credentials Safely

n8n's docs state that credential secrets are not sent to the AI and the AI does not see them.

When the workflow needs a credential, the Assistant should prompt you with a credential card. From there you can:

• select an existing credential you have access to
• create a new credential
• enter the secret in n8n's normal credential screen

Do not paste API keys, passwords, OAuth tokens, private URLs, or customer data into the AI chat unless it is genuinely required. Use n8n's credential system.

A good rule:

Describe the credential by purpose. Enter the secret only in the credential screen.

Example:

Use my existing production Slack credential if available. If not, ask me to create one. Do not ask me to paste the Slack token into chat.
Click to copy

Testing the Workflow

The launch video emphasises that the Assistant can test the workflow and show it running in front of you. Use that. Do not publish a generated workflow without test execution.

Before publishing, check:

• Node choice: did it use the right trigger and actions?
• Credential mapping: are the right accounts connected?
• Data shape: does the output from one node match what the next node expects?
• Error path: what happens when an API call fails?
• Side effects: will it send messages, delete records, charge cards, update customers, or publish content?
• Permissions: is it operating in the correct project?

For production workflows, run a test with safe sample data first. If the workflow sends emails or Slack messages, use a test recipient. If it updates a CRM, use a test record. If it deletes anything, stop and make it ask for confirmation.

Publishing Safely

n8n Assistant may ask for confirmation before high-impact actions such as publishing, deleting, or making important changes. Still, write the instruction yourself:

Ask me before publishing this workflow or making any destructive change.
Click to copy

A safe publishing checklist:

1. Review the workflow visually.
2. Open each node and check the configuration.
3. Confirm credentials are correct.
4. Run a manual test.
5. Confirm expected output.
6. Test at least one failure path.
7. Publish only after you understand what the workflow will do.

Preview AI features can make mistakes. n8n says this plainly in the docs. Believe them.

Credit Usage: How Not to Waste It

The AI Assistant uses credits based on the work it performs. n8n's docs say longer conversations, larger workflows, debugging sessions, and repeated iterations use more credits.

To reduce waste:

• Start with a precise prompt.
• Ask the Assistant to propose a plan before building.
• Keep unrelated tasks in separate conversations.
• Do not ask it to regenerate from scratch unless your requirements changed.
• Make small refinement requests instead of vague complaints.

Instead of:

This is wrong. Try again.
Click to copy

Use:

Keep the Gmail trigger and Drive upload steps. Change only the Data Table schema so it includes invoice_total, currency, vendor_name, and due_date. Do not rebuild the rest of the workflow.
Click to copy

Specificity saves money. Imagine that.

When to Use n8n Assistant Instead of Building Manually

Use the Assistant when:

• you know the business process but not the exact n8n nodes
• the workflow touches several apps and needs a first draft quickly
• you need help debugging a failed execution
• you want to iterate on a workflow while watching the canvas update
• you are prototyping and can review before production use

Build manually when:

• the workflow controls destructive operations
• regulatory or audit requirements demand exact change control
• the workflow is already stable and only needs a tiny edit
• the data contains secrets or sensitive customer information you do not need to expose
• you do not yet know what the workflow is supposed to do

The Assistant is a builder, not a product manager. If you have not decided the process, it cannot do that thinking for you.

n8n Assistant vs Claude Code with n8n MCP

The launch video makes an interesting comparison: Claude Code plus the n8n MCP can work well, but n8n can optimise the Assistant for its own interface in ways an external coding tool cannot.

That is the real advantage.

With n8n MCP, an external AI tool can connect to n8n and build workflows. That is powerful if the rest of your task is already happening in the external tool.

With n8n Assistant, the build happens inside the workflow editor. You can see the workflow, set parameters, select credentials, and test the execution in the same environment.

A practical split:

• Use Claude Code/Codex + n8n MCP for developer-heavy projects where workflow automation is one part of a larger codebase.
• Use n8n Assistant for workflow-heavy work where visual inspection, credentials, and test executions matter most.

Common Mistakes

1. Giving it a one-line prompt

Bad:

Make a lead automation.
Click to copy

Better:

Create a workflow that triggers when a new Typeform response arrives, enriches the company domain with Clearbit, scores the lead, adds qualified leads to HubSpot, and sends a Slack alert only if the score is above 70.
Click to copy

2. Publishing without reviewing

Generated does not mean correct. Review every node before production.

3. Pasting secrets into chat

Use credential cards and n8n credential screens. Do not paste keys into the Assistant chat.

4. Letting it change too much at once

When editing an existing workflow, specify what should remain unchanged.

5. Ignoring failure paths

Every useful automation needs an error path. If you do not specify it, the Assistant may give you the happy path only.

A Reusable Prompt Template

Use this when starting a new workflow:

Create an n8n workflow for this process:

Goal:
- <business outcome>

Trigger:
- <event or schedule>

Apps and credentials:
- <app 1>: use or ask for <credential>
- <app 2>: use or ask for <credential>

Steps:
1. <first action>
2. <second action>
3. <third action>

Data mapping:
- <source field> -> <destination field>

Success behaviour:
- <what should happen when it works>

Failure behaviour:
- <what should happen when it fails>

Safety constraints:
- Ask before publishing.
- Do not delete records.
- Do not send external messages until I approve a test.
- Explain the plan before building.
Click to copy

For existing workflows, use this:

Update this existing workflow.

Change only:
- <specific nodes or behaviour to change>

Do not change:
- <working paths or nodes to preserve>

Before editing, explain your plan.
After editing, run a test and show what changed.
Ask before publishing.
Click to copy

The Bottom Line

n8n Assistant is useful because it brings AI workflow building into the place where workflows actually live. You can describe the automation, watch the workflow appear, set credentials safely, test the run, and refine the result without leaving n8n.

Treat it like a fast junior workflow builder: useful, tireless, occasionally wrong, and absolutely requiring review before production.

If you use it with precise prompts, safe credentials, test data, and explicit publish controls, it can turn a rough automation idea into a working n8n workflow much faster than starting from a blank canvas.

If you use it vaguely and publish without checking, that is not an AI problem. That is you creating a small operational disaster with a chat box.