I've built hundreds of Zaps over the years. Most of them handle simple tasks. Copy this field to that field. Send an email when X happens. Create a contact in the CRM.

But there's one Zap I built six months ago that quietly saves me more time than any other. It takes form submissions from our WordPress contact forms, runs them through OpenAI's API, gets back a summary and routing decision, then sends the lead to the right person or system.

Before this, someone on the team had to read every form submission, figure out what the person actually wanted, and forward it manually. We were getting about 40 submissions a day. That's easily an hour of work, often more if the submissions were long or unclear.

How it actually works

The trigger is a Gravity Forms webhook. I tried using Zapier's built in WordPress integration first but it was unreliable with our caching setup. The webhook fires immediately when someone submits the form.

Next step is a Formatter action to clean up the submission text. I strip out extra line breaks, remove the field labels, and concatenate everything into a single block of text. This part matters because OpenAI charges by token and clean input means fewer tokens.

Then comes the OpenAI step. I'm using GPT-4 with a temperature of 0.3. I tried GPT-3.5 Turbo first because it's cheaper but the routing accuracy was only about 70%. With GPT-4 I'm getting 92% accuracy based on a sample I checked manually last month.

The prompt is straightforward. I ask it to:

  • Summarize the inquiry in two sentences or less
  • Identify the primary intent (support, sales, partnership, or other)
  • Assign a priority level (high, medium, low) based on specific criteria I provided
  • Return the response in JSON format

The JSON format is important. Zapier's built in JSON parser can then split the response into separate fields I can use in later steps.

After that, I have a path split based on the intent. Sales goes to our GoHighLevel pipeline. Support goes to our help desk system. Partnerships come directly to my email with the AI summary at the top. Other goes to a general inbox.

What didn't work

I initially tried to have the AI write a draft response as well. The idea was to save even more time by giving the team a starting point. But the responses were too generic and people ended up rewriting them completely anyway. It actually added friction instead of removing it.

I also learned that OpenAI's API can occasionally time out or return errors. The Zap failed about once a week in the first month. I added error handling with a fallback path that just routes everything to the general inbox if the AI step fails. Not ideal but better than losing the submission entirely.

The other limitation is cost. We're spending about 40 dollars a month on OpenAI API calls for this Zap alone. That's still way cheaper than an hour of team time every day, but it's not free.

Practical takeaway

If you're handling a decent volume of inbound requests and someone's reading through them manually, this pattern works. The key is getting the AI output into a structured format you can actually route on. Don't try to get too fancy with the AI response. Let it do the categorization and summarization, then let your team handle the actual human interaction.