How Vox Media went from no-code workflows to running AI agents

Why it matters

Most ad operations teams hit the same wall. The work is repeatable and the data already exists somewhere, but the volume grows faster than the person doing it. At Vox Media that person was Clay Vickers, going through a Google Sheet by hand and making optimizations one ad at a time.

What changed was not headcount. He built a set of workflows that make and record those optimizations at scale, pulling from the sheet through a native connector and pushing into the ad server through custom connectors and custom scripts against its API. Common systems get the connector that already exists. The system that matters most to his team, and to almost nobody else, gets one he builds himself. Both run in the same platform.

The agent work follows the same logic. Rather than starting from what an agent could do, his team looked at the whole intake process and asked where the support was thinnest and where it could touch revenue. That pointed at the very start of the sales cycle, where a seller enters a handful of fields and now gets a research brief back.

→ See how teams deploy AI agents on Tray

In their words

How would you describe Tray to someone in your industry?

I would describe Tray as an integration and automation platform. I think it works best with an example. So most organizations have a CRM platform and some kind of project management tool, or a place where you’re actually getting the work done. Those two systems probably don’t talk to each other, but let’s say that when something changes in your CRM, you need to see that reflected in your project management tool. For us, that’s Salesforce and Asana. Tray is the tool that we use to make sure that those two essential systems are in conversation.

"Tray is the tool that we use to make sure that those two essential systems are in conversation."

Walk us through a workflow you’re proud of, end to end.

So we have a set of workflows that help us optimize ads that we have running in our ad server. So we track these and all of the data about them in a Google Sheet, which is fed through a bunch of different systems. And a few years ago, we might not have that many running, and it was possible for just me, just one person, to go through and manually make the changes and make the optimizations that you need to.

But the business has grown and scaled, and every new ad that we’re running could be additional time to make sure that it’s running correctly. So I was able to design a set of workflows that can make those changes and record them at scale all at once by using a few different things. One is a native connector that Tray has built, which is for Google Sheets. And then on the other side, we’re using custom connectors and custom scripts to be able to connect directly to our ad server through the API there. So that’s one of the workflows that I’m most proud of.

What have you been able to build with Tray that you couldn’t have built otherwise?

I think every workflow that we have running in Tray is something that we’ve only been able to do because we’re doing it in Tray. But that’s because of different reasons for the platform.

When I first started building with Tray six or seven years ago, I had almost no knowledge of how to code, but I was able to still build workflows by using the native connectors and reading through the help docs and doing trial and error and doing different things. As the platform has matured, and as I’ve gotten access to new tools like chatbots like Claude, I’m able to do even more advanced things still with my base knowledge of not knowing that much about Python or HTML or anything like that.

The example I gave about the workflows that are making optimizations in our ad server is, I think, a unique story for Tray. The only way that that works is that we’re using both a native connector for something really popular and common like Google Sheets, but we’re also able to do something custom for a tool that is really important to us, but not necessarily used by millions of people like a Google product is. And Tray lets us combine both of those things in the same platform.

"When I first started building with Tray six or seven years ago, I had almost no knowledge of how to code."

How did you decide where to put your first agent?

We’re pretty new to agents, but we’ve got a couple up and running. When we were deciding what made the most sense in terms of where to place an agent, we just looked at our entire intake workflow and thought to ourselves, what are the areas that need the most support, and what area could have an actual impact on revenue?

So we looked at our sales cycle and we realized that sellers at the very beginning of the process might do better if they had a little bit more information. They’re entering a few things into the CRM, into Salesforce for us, but we built an agent that takes in that information, does kind of a basic Google search, and also pulls from our knowledge base that’s in Google Drive, to put together an intelligence brief for the seller for each new campaign that they’re entering into the system.

My advice for other people is look at part of your process that’s repeatable, but also could use some extra support that otherwise might be really manual. It might take a seller or an account strategist a lot of time to do all that research themselves and compile it into one specific brief, but by setting up the agent, we can have that run and happen automatically right at the beginning of the process.

"Look at part of your process that's repeatable, but also could use some extra support that otherwise might be really manual."