Filters, branches and paths

A workflow needs rules for when to carry on, when to stop and when to take another road. Three tools do this: filters, branches and a person who checks the result.

What you need

• The yes-or-no question, written plainly. "Is the order over the limit?" works. "Is this a good lead?" doesn't, unless a field answers it.
• The field the answer depends on.
• A decision for the "no" case: drop it, send it down another road, or hand it to a person.

Filters: carry on or stop

A filter checks a condition, or a set of conditions, and either lets the run continue or ends it. Zapier says that with filters, you set a specific condition and the Zap will only continue if the data from your app meets that condition.¹

Make works the same way. Bundles that meet the conditions are passed on to the next module.²

You can combine tests. In Zapier, use AND when all conditions must be true for the Zap to continue, and OR when any one condition can be true.¹ Negative tests such as "does not contain" are easy to get wrong, so test with real data.

Branches: choose a road

When "no" needs its own action, you need more than a filter. Zapier draws the line this way. Filter stops the Zap entirely for runs that do not match your conditions, while Paths routes data to different actions based on conditions.³

The three tools name this differently:

Order can matter. Make says routes are processed sequentially, not in parallel.⁴ Zapier says path branches run one at a time, left to right.³ Zapier adds that its sequential running can change how a Zap behaves when a branch uses a Delay or other slow action, so check how your tool orders branches.³

Always give the branch a fallback. A run that matches nothing should end up somewhere you can see.

Where a person checks an AI answer

An AI step returns text, and a filter can only test that text. Sorting a message into a label and then filtering on the label is a fair use. Letting a filter decide whether a draft reply is good is not.

For anything that goes outside the business, add a person. n8n documents this for AI agents. You can require human approval before an AI Agent executes a specific tool.⁶

Here's what n8n says happens. The workflow pauses and sends an approval request through your configured channel. If the reviewer denies it, the action is canceled and doesn't run.⁶

That page names sending messages, modifying records and deleting data as the kinds of tool that deserve extra review.⁶

This page doesn't cover approval options in Zapier or Make, so check their own docs for what they offer.

A short checklist

• Filter on a field you trust, not on the AI's opinion of its own work.
• Give every branch a fallback.
• Send risky actions to a person first.
• Run a test with data that should fail the filter, to make sure it stops.
• Leave secrets and customer records out of the prompt and the approval message, unless your business's written rules say otherwise.

For why an AI answer needs this treatment, read about what the model is doing.