What happened in Vegas didn’t stay in Vegas this year.
Our session at Braze Forge 2025 was overflowing (literally) with interest — and we’re inviting those who couldn’t make it into the session to check out our encore on-demand webinar. Learn how we’re helping brands like Taco Bell and Whataburger transform their campaign processes with AI and automation.
Our key takeaways:
1. Automate the busywork. Keep humans on the big stuff.
AI and automation are no longer experiments. They’re how leading brands scale smarter. By building “agentic workflows,” teams can hand off the repetitive, rule-based tasks (QA checks, asset formatting, campaign setup) and redirect focus toward strategy, creative, and customer insight. The best work still requires human context and judgment — agents just clear the way so that work can happen faster.
2. Attack the bottlenecks first.
Every campaign process has one or two frustrating pinch points. For Whataburger, it was campaign composition. The process to get messages from Figma to Braze was a heavy manual lift. Their new agent now handles that automatically, saving ~4 hours per campaign and more than 20 hours a week. For Taco Bell, it was QA. Checking hundreds of configurations before launch taxed their already lean team and still left room for error. Their QA agent runs 200+ tests in under a minute, reducing errors and freeing up another 25+ hours per week. The takeaway: you don’t need to automate everything. Start by automating your biggest bottlenecks.
3. Do more with the stack you already have.
These agents weren’t built with new software; they were built by making smarter connections. Using Braze’s existing APIs, n8n for workflow automation, and tools like Figma, ClickUp, and Puppeteer, both teams created powerful AI agents without adding another platform. The message: the modern marketing stack you already have can do far more than you think — it just needs connective tissue.
4. Start small. Then stack agents.
You don’t need a dozen agents or a huge task force to get started. Identify one clear, repeatable use case. Build one agent. Measure the time and quality impact. Then build the next. Each new agent compounds your efficiency and extends your foundation. Like any good campaign framework, it’s iterative: test, learn, scale.
Transcript:
Let’s get started. I’ll do a quick housekeeping note. We are open to taking questions throughout the webinar. So if you have anything that comes up, you are welcome to drop it in the chat or in the Q&A panel and we’ll most likely answer them at the end. But Bobby may end up answering your question throughout. Go ahead and get started.
Awesome. Thanks, Bri. Hey everybody. My name is Bobby Tichy. I lead our solutions team here at Stitch. Stitch is a Braze and martech consultancy, and we are doing a follow-up session from our presentation at Forge, which we did with Troy, who is on with us as well. Troy is the senior manager of CRM Data and Solutions at Taco Bell.
So today what we wanted to focus on was walking through how Taco Bell and Whataburger have changed their campaign process through AI and automation. There will be three key areas of focus today. Number one will be AI automation in the campaign process — how Whataburger and Taco Bell are using AI and automation, and what an agentic workflow and AI agents look like.
And I think one thing that’s really important that we’ll go through throughout this time, and what we’d love to have your questions on as well, is what are the key differences between AI automation as well as agentic workflows.
So before we jump in too far, and before we jump into the Whataburger piece before we get to Taco Bell — Troy, I’d love for you to just do a quick introduction of yourself, and then we’ll jump in from there.
Yeah. Hi everyone. Troy Vo, as Bobby mentioned, senior manager of customer data and solutions at Taco Bell. I lead the marketing automation and technology solutions. Been with Taco Bell almost five years, technology background, and been working with Stitch for at least a year and a half. So we love the Stitch team and appreciate them, and I think this is a very awesome AI workflow that Bobby will share with you guys.
Awesome. Thanks, Troy.
So to start — at Stitch, we have this hypothesis that in the next three to five years, AI and automation will really heavily take over the campaign process, specifically the repetitive and copy-paste things that we’re used to today. And we think that a platform like Braze — and for those of you who are already Braze customers or evaluating Braze — you may have seen over the last couple of weeks just how quickly they are innovating and adding onto their platform to be as open and flexible as possible. Most recently building out same-day integrations with the new ChatGPT browser Atlas, as well as building an integration within ChatGPT so you can leverage things like content cards or in-app messages within a prompt within ChatGPT.
And so we think the combination of Braze, AI, and automation is really where all of these things are going to come together and help marketers do things a lot faster and a lot more efficiently. And the reason why we think that is because obviously the advancements in AI have really opened the door for automation as well. We really believe that humans should focus on anything that provides creativity or innovation. The repetitive tasks, the swivel-sharing between platforms, the copy and pasting — we really think that will go away in the next couple of years.
So before we jump in and kind of walk through these agentic workflows and these agents, what we first want to do is ground ourselves in understanding AI and automation, but then also establish a quick vocabulary foundation for things that we’ll be talking about today.
So first, AI over the last couple of years has really opened eyes to automation in a way that automation has been around for a very long time. Platforms like Workato, MuleSoft, or Tray have all had these kinds of automation or workflow automation elements to them. And what’s been really interesting is that AI has completely blown up automation in a great way, where automation is really now at the forefront of a lot of marketers’ minds. But they are very, very different things.
What you’re seeing on this slide are the main elements between automation and AI. The way you can think about automation is we’re telling it exactly what to do, and it’s executing those things based on specific actions we give it. AI on the other hand is much different because AI can actually perform tasks on its own. And that’s what we’ll walk through today with Whataburger and Taco Bell — these agentic workflows could do a combination of automation and AI, where it’s doing what we’re telling it to do, but it’s also making some assumptions and taking actions on its own as well.
So Troy, before we jump into the quick vocabulary lesson, what are a couple of ways that Yum and Taco Bell use AI today?
Sure, happy to share. At Yum, I would say three categories. We have Byte at Yum, which is a proprietary AI platform that powers our restaurant operations and guest experience across all the brands. We use it to streamline our ordering, labor forecasting, and menu management.
We also have voice AI at the drive-throughs, powering ordering. It helps with speed of service and improves accuracy.
And then we’ve also started working with OfferFit, which is now Braze’s AI Decisioning Studio. OfferFit is AI decisioning that helps us select the right messages at the right time to the customer in near real time.
Yeah, I like the acronym now of BAIDS — Braze AI Decisioning Studio for short. So that’s great.
So a quick vocabulary foundation to make sure that we’re all talking about and thinking about the same things as we go through the rest of this presentation and the agents that we’ll walk through. First — automation. We’ve already talked about this. It’s something that is executing something predefined, doing exactly what we give it. Next is AI, where it’s able to reason and make assumptions on its own. We’ll walk through a couple of examples of that with each of the agents.
Agentic workflows are a combination of those two things. And AI agents are specific elements that are taking actions on their own. You’ve all probably interacted with at least one AI agent recently, whether it’s a virtual assistant when you’re booking a hotel room or making a flight change with Delta. All of those different things could be considered AI agents because they’re taking an action on their own — it’s not predetermined, and it’s not relying on a human to make a decision before ultimately making that decision.
So as we think about AI and automation within the campaign process, almost every single aspect of the campaign process could be enhanced or automated through AI and automation. It’s everywhere, as you can see on this slide. The yellow elements and the pink elements each outline an element of automation or AI that could be attributed to or make a piece of the campaign process better. Some of those are Braze elements, some of those are elements of automation within Braze, but also how LLMs could help — whether that’s Gemini, ChatGPT, or Claude, whichever LLM you use — as well as a couple of different workflow automation platforms like n8n, and now Braze coming out with Agent Console as well.
So before we jump into specifically the Whataburger and Taco Bell use cases — Troy, I’d love to hear your perspective on, since you all are using OfferFit and some of the Braze AI capabilities, how has working with those things helped you all at Taco Bell?
Yeah, I would say OfferFit is definitely an exciting — I wouldn’t say new anymore. We’ve been using it about a year now. It’s definitely a functionality that we use for reinforcement learning. We’re able to scale our testing much quicker. It uses the multi-arm bandit approach, so it balances exploration with exploitation — it tries new offers, captures positive reinforcement, and then serves the best performing offers based on what it sees, in near real time. So it creates a pretty positive impact for our customers. And it just does a lot of smarter personalization at scale. We’re able to test and find results much quicker at scale compared to normal A/B testing.
I think that’s where what we’re trying to do — whether it’s through these agents that we’ll talk about today, or through Braze AI Decisioning, or even just workflow automation within Braze — what we’re really passionate about is figuring out how we can leverage Braze and technology and AI automation to make marketers’ lives easier but also more effective. So that example of not having to do manual A/B tests and being able to do dozens and hundreds of permutations on those things without having to manually do it is really exciting and really impactful.
So we kind of saw through the campaign process where we can leverage AI and automation throughout those different things. The other key element we wanted to think through as we talked through this process with Whataburger and Taco Bell was where in their process could AI and automation help, and where could an agent make a process either more efficient or more effective?
And the other part of this too is it doesn’t always make sense, right? Just because we could build an agent or could build automation for a particular part of the campaign process doesn’t mean that we should. And the reason for that is because, as my dad used to say when I was growing up, nothing happens magically or for free. And that’s the same thing here within these processes, right? Building an agent or building an automation does not necessarily have an effective ROI attached to it. These things take time, they take money, a lot of times they take cross-functional teams, so it doesn’t always make sense to build them out for every single piece of the process.
So what we did was we went through each of Whataburger’s and Taco Bell’s full end-to-end campaign process. We identified where an agent or automation could help in that process. And then we went through the proper approvals internally. As I’m sure many of you on the call have experienced, there’s always going to be an approval process, or maybe there’s even an AI task force or team internally that’s reviewing these things. We went through that with both teams. And then we ultimately went through this process of designing and then building these agents. So identify, discover, solution, build, test, deploy — and then today, these things are live and we’re constantly optimizing them week over week and month over month to make sure that we’re constantly adding new functionality to make them even better.
So let’s dive in. First, let’s jump into the Whataburger composition agent. And like I mentioned, the first thing that we did was go through their end-to-end campaign process. So you can see here all the tasks on the left-hand side, and then the time it took in the pink bars to the right. And you can see with all the different cross-functional teams that were required for these campaign processes — on the bottom right there — as well as where the Stitch team, serving as an extension of the Whataburger team, is helping as well.
So what we did is we went through every single one of these activities and identified — and you’ll see that in purple and orange — where AI or automation could help make this process either more efficient, so it got done faster, or more effective, meaning it could actually be done better.
And so the really exciting part about this is, number one, identifying where it could be used — which is just about everywhere — but also where it makes sense to do that. Like I said earlier, it doesn’t always make sense to create an agent or automation, because building an agent for a particular activity may take a ton of time, effort, and money, which we wouldn’t get that ROI back from. It doesn’t save us enough time or doesn’t make the process that much better.
So to give a primary example, we went through the campaign and offer ideation — that first row at the top — and asked, would this be a good example of building an agent or automation? And we determined this was not, and the reason being was we wanted to figure out where automation would create the biggest lift without compromising any kind of creativity or judgment. That campaign and offer ideation is where the human side of marketing really thrives within Whataburger. So looking at that particular activity didn’t make a whole lot of sense. There’s so much brand voice and creative direction, financial approvals, and they have an offer engine that they use — all these different things are really important to that process.
Another one that we looked at was prioritization. Again, this was not a huge pain point for Whataburger. So when we mapped out this entire process, what really stood out was just how much time and energy was being lost in the composition stage of this campaign process. So taking assets from Figma, copying text from briefs, building templates in Braze, making sure everything was formatted correctly — it was incredibly repetitive manual work, but it’s also very critical to building out these campaigns. It’s required, right? But by leveraging this and deciding ultimately on a composition agent, it tackled the single biggest bottleneck that Whataburger had, but also that the Stitch team had as part of this.
So the reasons why we chose this composition agent — because at the end of the day, we realized this was the largest amount of time it took for the Whataburger team or the Stitch team working with the Whataburger team to complete a campaign. Typically around four hours per campaign. And there were multiple pieces of technology involved — manual emails, ClickUp as the project and campaign management tool, Figma for designs, and Braze for actual configuration of those elements as well.
So how does the composition agent work? At a very high level, the outline for this composition agent is: it creates multiple tasks in ClickUp from the creative brief. It takes the build process from ClickUp, identifies everything that needs to be built, and then analyzes and creates those assets from Figma. So the creative team is building all their assets — whether it’s an email, an in-app message, gamification, whatever it might be — all that lives in Figma. And then it creates all of that content from Figma into HTML, CSS, and JavaScript, and then configures the actual campaign and all the elements inside of Braze.
So let’s quickly look at the before of this process before we jump into the after.
So first, looking at this process — everything is manual. This is a sped-up version of a video that will loop through. But today everything is manual for us and the Whataburger team in building out these processes. So you’re having to manually create tasks in ClickUp based on a creative brief that’s provided from the Whataburger team. We use ClickUp to access the Figma file — so you’re going in and clicking a link within ClickUp to access the Figma file. You’re manually analyzing Figma, and for those of you who use Figma, you know that could be a pretty big effort because there could be multiple campaigns, multiple versions — sometimes creative teams do all of their creative for a quarter and sometimes even a year in just one Figma file. So it can be difficult to decipher what actually is being leveraged for that particular campaign.
Then you’ve got to manually extract that asset from Figma to the desktop. From there, you’re actually going through the process of manually creating that in Braze. So just like you all would normally do, you’re configuring the campaign, you’re selecting the email template, you’re creating the content in the template. And then once you’re actually in that template, you’re going back and doing all of the elements that we all do every day — uploading the image, adding it to the template, adding alt text, adding a URL. So again, all very manual processes that we have today. And then from there is previewing and testing this process. As you can probably imagine, you’re seeing a manual trend here, right? Every aspect of this process is completely manual.
So from there, we go to the after. And now with an agentic workflow and agents, this is where it gets really fun.
So the first step of this process is actually creating all of the elements inside of ClickUp automatically through workflow automation. So we’re taking the creative brief — which is either a Word doc or a Google Doc — and we’re analyzing that doc. We’re creating all of the tasks that need to be done inside of ClickUp through the API.
So from there, once all of those tasks are there to actually build the email or in-app message or whatever assets we need, it all just starts with a button. You’ll see that button in the middle of the screen there that just says “Build from Figma.” So what’s happening there is that our agent is analyzing the tasks in ClickUp along with the campaign brief to identify the Figma file that we need to go to to pull the assets from and create.
From there, the agent identifies all the assets within Figma. And again, this is a really big deal. As we were going through it the first time — actually the first couple of times that we pulled the payload back from Figma — it actually crashed our server that we were using to build the agent, because for those of you familiar with Figma, the API is very robust. It is a phenomenal API — it has every aspect of a Figma file that you could pull down. But what was really cool about it was we were able to decipher exactly what we needed from Figma and which actual piece of content we needed to pull down.
So once we pull that down, our agent translates those assets into HTML and CSS, and then builds the components in Braze through the API.
From there, the process goes through an update to ClickUp. So once that is completed — the whole Braze build — a comment comes back into that same task in ClickUp where we initiated the agent, and it tells us that the agent is done, that the configuration is complete, and that it’s ready for review. And so that’s when you would go in and manually take a look, make sure the agent configured everything correctly. The other cool thing too is you can review all of the data within a catalog, because that’s what’s powering every single one of these campaigns.
So let’s walk through the agent itself. What does an agentic workflow look like, right? So we just walked through what it looks like from the front end, but what does the agent actually look like behind the scenes?
This is the high-level overview, and we’ll dive into each one of these different elements. But this workflow is configured and maintained in a platform called n8n — just the letter N, the number 8, and then the letter N. And this is what we’ve found so far to be the best platform to create these agentic workflows. Number one, it can live on any cloud platform. So for example, if we first build one of these agents in our Azure instance, we can then migrate it to a Databricks or Redshift instance or Google Cloud Platform later on. So as companies migrate from different cloud platforms or different providers, they can move them from place to place. Also, the integration mechanisms to build out things like MCP servers are really strong within n8n.
So there are three key workflows to this. Agent number one is analyzing the brief and identifying the Figma assets. It’s then extracting those assets from Figma and converting them into HTML, CSS, and JavaScript. And it’s configuring those things in Braze. So let’s dive into each one of these workflows a little bit deeper.
Step one — that top row of the workflow we just saw — is identifying and extracting the creative from Figma. And there are multiple individual agents here along with MCP servers, and those are really the heroes of this workflow. Because what these agents allow us to do is make API calls or make decisions on certain things that make the manual process of these workflows completely go away. And the MCP servers are the way that we’re able to interact with these platforms back and forth.
So many of you may be familiar with an API. What’s nice about an API is it gives you access to things, but it’s only one way, right? When I make an API call, I’m just expecting a response with a particular piece of data or an asset or something like that. MCP servers are bidirectional. So I could make a call to an MCP server and essentially have a conversation with that platform that I’m accessing — like Figma or even Braze. And so you can do a lot more with it. It’s a lot more advanced version of an API.
So the first half of this workflow is identifying all the assets within Figma, and more importantly, identifying the assets that are specific to that particular campaign, because like I mentioned earlier, you could have multiple campaigns within one Figma file.
The second half is validating the images. Like I mentioned, there are multiple layers in the Figma file. We only want the images that are related to that campaign. And so that’s how we’re determining which images and which assets we need to pull down.
Step two is converting the Figma files to HTML. We found that Anthropic or Claude is the best LLM for doing any kind of HTML and CSS development. We also built our own Figma MCP server. Figma has their own MCP server, but we built our own primarily because we only wanted to make sure we were looking at the assets that were specific to the campaign process. We didn’t need to pull down things for paid media or direct mail. So we wanted to pare down that MCP server.
And then step three is building the content in Braze on the right-hand side. Some of you who are familiar with Braze and really in the weeds of the Braze API might be thinking, well, there’s no image API inside of Braze, so I can’t upload images through the API. And I couldn’t really do that in a workflow automation. Well, this is what’s really cool about a function and feature called Puppeteer. Puppeteer essentially mimics human interaction in any way that we want to. So what we’re doing within this Puppeteer agent is we’re actually logging in to Whataburger’s instance of Braze, and then we are uploading the image that we’ve gotten from Figma through that Puppeteer process, so that way we don’t have any manual steps as part of this workflow.
So that was a lot of deep technical elements and kind of how agents flow. But today, the biggest impacts for Whataburger are the saved hours per campaign. We talked about about four hours per campaign. On average, Whataburger sends at least six campaigns per week. So that’s 25 hours a week — just saved time and money, either internally or money that they’re having to spend with Stitch.
And what’s really cool about this, as you’re probably all thinking through, is that normally you’d need 10-plus people to go through this campaign process — creative, CRM, legal, brand, QA. Each step has its handoffs, meetings, and reviews. And I’m sure you all have gone through a process where a campaign was delayed because one approval was missed, or you had to stay up late because QA found an error at the last minute. With this composition agent, we’re basically imagining a world — and being able to execute on this — where we brief our AI agent at 9:00 AM and a couple of minutes later we have a fully built campaign inside of Braze that we’re ready to launch in hours instead of weeks.
And so it really allows the Whataburger team to focus more on creative strategy, listening to their customers, and innovating with their partners as well.
Let’s jump into the QA agent and Taco Bell. So Troy, I know you shared a little bit about your background already, but I think it’d be great if you shared a little bit more about how you got to Taco Bell and kind of how your career ended up where it is today.
Yeah, as I mentioned, my background has always been technology. I like to define it as being at the intersection of technology and strategy. So I’m a solution architect at Taco Bell and I lead a team of solution architects. When I joined Taco Bell was back in 2020, in the middle of the COVID days, and we also launched our loyalty program there. So a lot has changed ever since. We also launched our SMS marketing program, and we’ve been working with Stitch for a lot of in-app gamification projects. So a lot of exciting things happening there.
As Bobby mentioned earlier, I think there should be a distinction between AI and automation. I think that’s how we also view it as well. I think automation for us is about efficiency — there are certain rules that we know repeat and we can automate and schedule in advance — whereas AI is more about intelligent adaptation, using algorithms, predictive modeling. We also like to view both of them as working together in a complementary way. So automation creates consistency, while AI helps us uncover insights, make better decisions, and personalize. And then yeah, the QA agent that Bobby and I will share today will hopefully show how we are scaling our process and making it more efficient for the team.
Awesome. So let’s jump into it. Like we walked through with Whataburger, we have the campaign process today. Troy, if you wouldn’t mind walking us through this end-to-end campaign process at a high level for each row. Don’t need to get too detailed into it, but walk us through this and what your current — or previous — process was before the QA agent.
So our process has a lot of steps. As you can see, very time-consuming, a lot of manual work prior to all the automation processes we put in place. But I’m sure similar to other brands — we have a brand strategy and brand calendar experience that we ladder up to. So there’s a lot of ideation and planning. Then after that there’s briefing, and then our team submits tickets to us. We use Jira today for requirements and all the details. And then there’s offer and content development — I’m skipping ahead, obviously running a little quick. Then we build our segmentation, configure the canvas and templates in Braze. The key part next is the QA and proofing, which takes a lot of time. So we identified that as a big area for improvement. And then final scheduling and launch. After that we monitor performance and provide feedback and reporting. So pretty quick run-through of that process.
Yeah, so as we went through that together, we again identified where AI or automation could help the most in these processes. And you can see there are kind of two main areas where the bulk of the time is spent. One is around QA, and then the other is around content block and build. And kind of like what we walked through with the Whataburger scenario, we know that agents and automation are not always a great fit because it takes time and effort to do these things. Troy, what was it about building content blocks that was not a good use case or candidate for building an agent or automation?
It was something that we feel like we can rule-base and support pretty quickly. It’s not something that we need to put an agent against at this time. I think there are more areas where we can focus our resources to put the agent against.
Yeah, well I think too, the way that you all have architected your content blocks is highly modular. So like relying a lot on Liquid logic and message extras to drive a lot of those things — so that way you’re not beholden to more of a manual process in that particular element, even though it takes some time. There’s a lot of versioning and personalization there.
But that’s really kind of the point, right? We all know AI and automation are not going to take over everything. It’s really important to map out this process and prioritize where it makes sense to launch these different things.
And that’s how we landed on the winner, which is the QA agent. The QA agent took a lot of time and effort — at least two hours for every single campaign. Troy, if you wouldn’t mind sharing a little bit more about why we ultimately chose the QA agent and why we thought it would be the most beneficial.
Yeah, I mean, I mentioned earlier the QA process is kind of tedious for us. We spend a lot of time reviewing and scheduling campaigns. The team would manually check everything, which takes hours. We also pivot a lot here at Taco Bell, so there are a lot of last-minute changes that we have to support. It just helps us move much quicker. It wasn’t sustainable, the current process we were doing, so we definitely reached out to the Stitch team to say, hey, how can we leverage technology with all the AI agents these days to help improve our process? So it definitely makes perfect sense.
So the outline of this QA agent at a high level — number one, it executes over 200 test scripts in less than a minute. And how it does that is through this high-level outline, which we’ll dive into here in just a moment. But you can think about each of these tasks as essentially different types of agents or workflows.
And I’ve highlighted those in the way that we kind of have roles here at Stitch. So one is a technical producer. A technical producer is a message development savant, right? HTML, CSS, Liquid, connected content — all those things are what they do all day every day. And so we have a technical producer agent. Then we also have a solution architect agent, which is actually going through all of the elements inside of Braze to make sure everything was configured correctly based on the campaign brief itself.
So what we’re also highlighting here is that there are multiple failure points, right? Just going through a list of test scripts is helpful, but what’s even more helpful than that is identifying where something has failed and then the specific reasons why it failed. So that way someone can quickly go back in and make the adjustments needed, and then run the agent again to make sure they’ve corrected it.
So let’s jump into the before process — a quick review and highlight of just how manual this process is. So again, this is a very highly sped-up video, but the Taco Bell team was spending over two hours for each of those campaigns. And it’s a constant back and forth between this test document and Braze — to look at the test script, go back into Braze, identify what those things are.
Troy, what was the biggest pain point of this manual QA process?
I’d say there are three major pain points. As you mentioned, a lot of manual work. The team has to do a lot of manual work — there’s checking, we use different tools, check against Google Sheets, against Jira, and then just reviewing all the scripts. And like you mentioned earlier with the content and Liquid, that was one of the big pain points.
Second one is being slow and repetitive, right? For us, we want to make sure we move quicker, be able to support the team and the business as we scale.
And then late-night pressure. I’m the biggest culprit of it with my team. A lot of times we have to stay late and put in extra hours. I definitely want to give my team the space for work-life balance. So the more that we can remove repetitive manual work is what we’re looking for.
Yeah, for sure. And I think too, especially for you all at Taco Bell, those last-minute changes, or trying to get a campaign out that is relevant to something that might be happening in the world or in pop culture — all those things are why we want to get campaigns out faster.
The other element that we want to highlight was what does this look like now with the new QA agent? So let’s jump into kind of this sped-up video first, and then we’ll walk through the actual automation and workflow itself in detail.
So first, the agent is kicked off through a test email. So what that ultimately looks like is a test email sent from Braze, and then it goes to this agent and that kicks off the entire process. And that agent cycles through every single test script that is on that test document. Right now there are over 200 different tasks on that document, but what’s really cool is that we’re able to take whatever that test document is and update it in real time. So maybe Troy and team have identified five different things that today were different from yesterday that we need to change. All they have to do is update that test document and the agent just reads from that document.
In addition to that, there are kind of two main elements of the QA — rendering and configuration. So it’s looking at different things like on the rendering side — it’s not just looking at, was this campaign built correctly? If it’s an email, for example, it’s looking at, is this email accessible and ADA compliant? Is it going to look good in dark mode and light mode? What is it going to look like for mobile and desktop as well?
In addition to that, the configuration side is, was this campaign actually built to the specifications of the original brief? So for example, if this is a canvas with eight different touch points and two decision splits, it’s looking at those elements as well.
So Troy, what’s most exciting about this for your team?
I would say just the benefits — it will make us move faster and smarter, and it allows us to scale quicker. Our business is always growing. We have a lot of very aggressive goals. So for us, the more that we can scale and execute more efficiently and more accurately, that’s the biggest benefit I see.
Awesome. That’s great.
So the next portion of this — let’s actually walk through the agent itself and the agentic workflow. So again, this is just the view of what we built within n8n, that automation workflow platform. And there are four key areas here that we’ll dive into.
Number one is how that test email works and is sent to that agent inbox. That’s what kicks off the entire process. So once a campaign has been built, that test email will be sent to the inbox that kicks off the entire process.
From there, what we’re doing is we’re integrating to Jira, pulling the business requirements document down to analyze it within that agent, and we translate it to plain text for LLM consumption. So a small element there, but what works really well for LLMs is plain text versus a formatted document. So that’s what we’re doing to make it very easy for the LLM to understand it and for the agent ultimately to do what we need it to do.
The next row is all of the configuration details. So what’s happening here is we’re actually pulling the configuration details initially from the brief in Jira, but then we’re pulling everything from Braze to make sure that everything was created correctly. So for example, it’s identifying the canvas, it’s pulling all the components of the canvas, and then it’s pulling all the content blocks within messages inside of that canvas as well to make sure everything was configured correctly.
From there, there are two main QA workflow checks that are happening. Number one is the rendering QA — that’s the top line there, where we’re going through everything like we talked about before: rendering, accessibility, delivery, best practices. Next we’re going into configuration QA. So these are the steps and the actual setup of the campaign. And again, the 200-plus QA tasks are actually looked at and completed in under a minute. And the task list is dynamic as well — it changes any time based on the campaign or based on the workflow that’s specific to Taco Bell at that particular time.
The other really cool thing about this is the detail it provides. So out of those 200-plus, maybe there are three errors or three things that don’t pass the QA process. The output back to that test script file is what failed and why specifically it failed, down to the specific detail. So for example, if we didn’t include alt text as part of a URL inside of an email, it will say “missing alt text.” So that way it’s very easy for the person who’s actually doing the QA to go back, make the adjustment, run the agent again, and be able to deploy the campaign really quickly.
So we talked about the hours saved per campaign — about two hours per campaign. Taco Bell is typically sending at least 12 campaigns per week. So again, at least 25 hours saved through this QA agent. Troy, what are the biggest benefits that you and the team have received from this QA agent?
Yeah, as you mentioned, major time savings is one of the biggest benefits. As we start using this process more, our error rate has definitely decreased. So fewer chances of errors. And for the team it’s better upfront requirements — as the QA results run, they give the team more insights into where the majority of issues usually happen, so the team can make corrections. And we are starting to see fewer of those in our BRD process, which is the document that the strategy team provides, and that helps with that.
And as I mentioned, it just allows us to get to faster deployments, get to the door much quicker, and more accurately. And again, these are not just wins for the team — I think it’s wins for our entire business as we scale and are able to execute and communicate better to our customers. So it’ll give us more time to free up and do more advanced strategic work as well. I think it’s a win on both fronts for sure.
Yeah, definitely. I think what’s cool to think about too is once we combine these campaign processes together — or these agents I should say — so for example, once you take the composition agent for Whataburger and the QA agent and start to combine these together, our next goal is implementing the QA agent from Taco Bell for Whataburger and vice versa — the composition agent for Taco Bell. We can really start to bring down the amount of time it takes to launch these campaigns from start to finish. We really get a good view of what that could look like — from a 14-day process down to six to eight days — and saving the team a ton of time and effort as well.
The other key thing I would say — oh, go ahead Troy.
I think I would add to that — yeah, I agree. I think for us working with you and Whataburger as well, that is definitely our roadmap to include the composition agent. I think the value here is being able to associate a bunch of agents together, that way you can have the full AI flow. That’s the advantage of not just using one agent, but a combination of agents together.
Yeah, for sure. And it’s a good point because the other key thing to think about as well is once you build an agent, then it’s one, optimizing that agent moving forward, but then also how can we layer that foundation to build additional agents as well? Just like we all know marketing is never finished, this kind of agentic workflow and foundation isn’t finished either. So what we want to do is set the foundation with that first agent, and that’s what we would really highly recommend that you all do — think about what’s the one agent or one thing that you could build to make a process better, then leverage that to build other agents on top of it. The first agent will always take the most amount of time, but then the other agents after that will become much more efficient for you to build on top of that foundation.
So I’m sure you’re all wondering a couple of different things here. One is how long did it take to build these things out? So sharing with you all — it took three team members on our side about 300 hours to build each of these agents. Each agent took about the same amount of time between the QA and the composition agent.
How much do these agents cost? Agents can vary across a very wide range. And the reason for that is they completely vary in complexity, how they’re being leveraged, and what we’re actually asking them to do. So they can range anywhere from $25K to $250K per agent, just based on the different platforms that we’re integrating with and what we’re actually asking the agent to do. The other thing to keep in mind as well is the infrastructure and maintenance of that first agent. But once you’ve got that first agent, you’ve got a really good foundation to build other ones.
What is the ROI? This is always a hard question, right? But it’s completely dependent on the use case. As you can see, for the composition agent with Whataburger, it saves them about 25 hours per week, and the same amount of time on the Taco Bell side. And I think the other element too on the Taco Bell side is just not requiring people to be around late at night, being able to strike more of a work-life balance — not only for the Taco Bell team, but for the team at Stitch that they work with as well.
So what you can take from this — there are a couple of key things here. One is, go through your campaign process, identify opportunities to incorporate AI and automation into your workflows. Again, there is a higher level of upfront work to start, like kind of building that foundation. But once you have that foundation, you can quickly add on additional agents.
Also, if you’re not leveraging some of the AI components within Braze, start there. Start small, see how it works for you. Whether it’s intelligent timing, intelligent channel, or intelligent selection, there’s a lot that comes with every Braze license that you can get a lot from — just to start testing and see how it works out for you before going full blown into an agent.
So if we were in your shoes, what we would be doing next are some of the elements that you’ll see here on this slide. First, conduct an end-to-end review and analysis of your campaign process. Identify the biggest pain points or repetitive tasks in that flow. Just like what we saw with Whataburger and Taco Bell, not every single piece of that campaign process is going to be a good candidate for an agent or automation. But the ones that take the longest or are the most repetitive typically are. And then start to build a roadmap. Obviously you want to start with one — don’t start with trying to build out eight, right? Just start with a single agent or a single automated workflow.
And if you’re having trouble thinking through that, at Stitch we have a complimentary Braze and MarTech assessment that we run. So what we do is we come on site for three to four hours or do it virtually to run through your entire campaign process and your MarTech stack to figure out — if we were in your shoes over the course of the next 12 months, what would we be doing to make this a more automated workflow, and to get the most out of your MarTech and out of your Braze platform?
Before we jump into Q&A — and if you do have any other questions, definitely type them into the chat — we also wanted to announce our Braze Innovation Series. This is our first webinar as part of that series. You’ll see a couple of upcoming dates for December, January, and March. We’re really, really excited to do that and deep dive into more of these different situations, examples, and solutions with our team as well as our clients. If you’d like to register for that, go to bit.ly slash Braze Innovation Series.
Great. Let’s jump into Q&A. First I’m going to ask Troy a couple of questions and then we’ll open it up based on the questions that have been put in the chat.
So first, Troy — what advice would you give other marketers who are curious to adopt AI-driven automation in their organizations?
Yeah, great question. I think you touched on it a little bit there. Foundation is key, so make sure you have the right foundation. Clean data is essential. I think we’ve all dealt with data that isn’t clean, and that raises a lot of issues at the beginning. And then begin with a use case — make sure you have a very specific use case in mind that you can tackle, start small. And then as you learn, you’ll start to be able to grow and scale the process.
What’s the next thing that you’re excited about being optimized by AI in the next 12 months outside of QA?
Well, that composition agent, right, we want to put that together. But I think we see opportunities in AI everywhere. It’s going to help our team be able to focus more and do more advanced work. So I think we have a lot of stuff in place. I can’t really go too much into forward-looking things at this time, but AI is definitely a major component of our roadmap for the next year.
What’s your go-to order at Taco Bell?
I’m pretty boring. I just like two soft tacos. That’s it. I like the classic stuff. We do have our Decades menu coming up — so a little shameless plug — I like to run with the classic items. Definitely, definitely go there.
And if you haven’t tried the crispy chicken burrito, I highly recommend it. It’s fantastic.
Where are humans still the most valuable in your campaign process today?
Yeah, I think our CTO and CLO say the same thing. I think obviously AI is something we want to be at the forefront, leveraging AI, but we see humans as still the most valuable component of our organization. So we definitely want to free up space for our team members, as I mentioned, to focus on more strategic opportunities. AI can handle all the repetitive stuff, predict patterns, but I think humans provide the best context, judgment, and the ability to really connect with customers. That’s how we envision humans and AI working together.
And for those on the call who haven’t started implementing AI at all and are starting from scratch — where would you recommend they start?
Well, again, I would say identify a use case that’s simple, that you can tackle and measure. Make sure you can measure accurately, because the key is making sure there’s a positive return on investment. And definitely get the cross-functional team involved, because you’re going to need help from all sorts of different teams, not just one. Don’t silo it. And then start learning small, and as you scale, you’ll be able to perfect the process.
Awesome. Well, thank you very much, Troy. It looks like we’ve got some questions in the Q&A, which is great. Keep them coming through if you have more.
We’ll start from the top. Is OfferFit for offers or messaging only? Troy, actually, since you have so much experience with OfferFit, would you mind answering that question?
No, not really. It depends on what you want to use it for — it’s an AI decisioning engine. So it’s part of Braze now. You can apply other types of campaigns there as well. It’s not just for offers. Hopefully that answers it.
Yeah, I think OfferFit is really great for any kind of revenue-driving use case. So for example, it could be for a welcome journey or onboarding campaign. It doesn’t necessarily have to be specific to an offer. It could just be different types of content, different types of creative, different subject lines, but it just allows you to automate those different types of testing at a much broader scale. Typically you’re able to do anywhere from 50 to 100 different versions for a particular use case. So things like welcome, re-engagement, abandoned cart, taking a next best action — those different things are all really good candidates and use cases for OfferFit.
Yeah, I would just add to my earlier point. I think same thing — you want to identify a very specific use case, because the way it works is you want to identify a KPI. It could be sign-up, could be any type of conversion you’re looking for. It doesn’t have to be an offer, doesn’t have to be revenue — it could be something else as well. And then the AI will make positive learnings toward that KPI that you’re trying to achieve.
Yeah, for sure. The next question from Nicholas — does the agent enable or auto-populate Canvas setup, or does that still have to be manually created?
No, the agent does create the Canvas configuration inside of Braze.
Next is — what are some of the basic marketing ops or workflow building blocks that you need in place to be able to work with these specific agents? For example, asset naming conventions or metadata?
I think what’s really cool about both of these agents is that there was no pre-work that we did with Taco Bell or Whataburger to get their data in a specific place or leverage specific naming conventions. Those things are certainly always helpful, right? Anytime you’re trying to automate or train someone new on processes, the more organized it is, the easier it is. But what was really great was, especially with an LLM, you’re able to decipher a lot of those things with AI. So we didn’t have to do a lot of that upfront. There wasn’t a large amount of upfront work that we had to do before we were able to start solutioning and building these agents.
The next question for n8n — who typically owns these workflows, or who should? Can our CRM team own it if we’re able to scale up?
That’s a great question. Every organization is much different. Some organizations have the CRM or the marketing team own them. Others, it’s the IT or tech or engineering team that owns them. What I would recommend is identifying someone within your team — it doesn’t have to be the marketing or CRM team, it could be someone from a cross-functional team — who is kind of the leader of these agentic workflows or AI in general, specific to marketing. And that’s really important — it has to be someone that’s specific to marketing.
We’ve met with a lot of companies who have an AI team or task force, but there’s not someone that’s dedicated to marketing. And the hard part there is that someone who’s working on AI use cases for customer service is not going to understand the main elements or use cases we might have from marketing, and vice versa. And so having someone who owns it as far as the use case and the ideation is concerned is important. Who owns n8n at a platform level isn’t as important, because ultimately anybody who needs access to it can build it out.
Next question — in order to build the agents like shown here, are the Braze agents an add-on required, or is this a separate platform?
Yeah, so you do not need to buy anything additional from Braze to build out these types of agents. We use n8n, that workflow automation tool, but you could use Workato, you could use MuleSoft — any workflow automation tool that you have licensed could likely build these similar types of agentic workflows.
Kevin asked, how many hours were saved in your speed to market for Whataburger?
About two hours per campaign, about 25 hours per week on average. And obviously when there are more campaigns that need to be built out, there’s more time saved. So recently, as campaign volume has come up, we’ve actually gotten over 40 hours per week saved through that composition agent.
Mike? Yep. This is being recorded, so we’ll share that out with you afterward.
Lexi asked, was there anyone on your team that was nervous about AI taking their jobs? And how did you communicate the benefits to your team members? Troy, do you want to take that one?
Yeah, I can take that. Happy to take it. No one on my team — I can say that with a hundred percent confidence — is afraid of AI taking our jobs. I think we actually love having AI. Again, it speeds up our process. Our mundane, redundant work will go away. It frees up the team to do more advanced strategic work. I lead a team of solution architects, so we’re always looking to solve different problems. So no one at Taco Bell is afraid of AI replacing us.
Next question — you touched on a tool, Puppeteer, that uploaded content within Braze. Were there security concerns allowing it within Braze, or did you perhaps only limit it to image upload?
You’re exactly right — you answered your own question. We just basically pared down the permissions for that particular user that Puppeteer was leveraging to only do the tasks that we wanted it to.
The next question is, what type of resources are needed to build one of these AI agents? I know we can hire Stitch, but I’m wondering if our engineering team may have the right resources.
Yeah, you’re exactly right. There’s nothing proprietary to how we’re building these agents. Anyone who is familiar with how LLMs work, building integrations — especially any kind of API or MCP server integrations — across multiple different platforms can do this. So as you can see, for Whataburger we integrated ClickUp, Figma, and Braze. For Taco Bell, we integrated Jira, Google Docs, Braze, and a couple of other platforms. So as long as you’ve got some folks on your team who understand how to build these types of integrations and workflows, you can absolutely do it on your own. Typically those resources are going to be inside of an engineering or a product team.
Martin asked, how are you creating the canvases programmatically? Is this a feature of the Braze MCP?
It’s not a feature of the Braze MCP. I think where your question is coming from, Martin, is that you can create a canvas through the API based on an existing canvas, but you cannot create a brand new canvas from the API. So what we did in certain situations where there was already a canvas configured that we could leverage, we just made a copy of that canvas — because that’s available through the API. For the ones where we have to create a new canvas altogether, we’re leveraging Puppeteer to create that.
Zach’s question is, what’s your strategy to reduce AI hallucination when using agents? And do you think hallucination risk increases as agents become more autonomous?
A hundred percent — on the last part of your question around hallucination. The big thing that we’re doing within these agents is that we are giving them very specific instructions and we’re giving them a scope of what they’re allowed to do and what they’re not allowed to do. So we certainly had to go through a number of testing and QA iterations to get them exactly where we wanted them to be. But because we’re in a workflow and we’re not relying on these agents for things like copywriting or content generation, we’re able to really hone in on exactly what we want them to do. Even though they’re able to go out and make an API call or create an asset through Puppeteer, we’re still governing exactly how they do those things. So very specific instructions is how we’ve made sure that there aren’t any random hallucinations that would cause it to create something that it shouldn’t.
I could also add to that — I think at the beginning, our team also reviewed a lot of the results. One of the cool things about the workflow Bobby shared is that after the workflow runs, it will generate a file of the results. So definitely at the beginning, spend the time to review and make sure there are no hallucinations. Continue to do that, but as time goes by, you’ll become more confident with the agent. Also, the way we have guardrails in place, the agent is not able to schedule anything or deploy anything, right? There are still a lot of guardrails in place to make sure our team still has full control.
Yeah, that’s a great point. Especially like what we talked about earlier — just like marketing, AI is never done. As the agents are using an LLM, as you all know, there are different versions of LLMs being created and updated all the time. So that’s something we’ve got to make sure we’re continuously on top of. And specifically on the QA agent, like Troy mentioned — that’s why it’s so nice to see the actual output of what’s happening and how those things are working, so we can prevent those hallucinations.
The last question we’ve got here is, what other Braze enablements would you consider a must-have as we move into utilizing AI as a company? Decisioning Studio is one that was mentioned.
What’s nice about Braze as a platform — just on the base platform — is that a lot of the AI components, like the copywriting assistant, intelligent timing, intelligent channel, and intelligent selection, all come standard with the Braze platform. I would say as you’re looking to go into AI as a company, I would first look at what can be done with the publicly available features through the API as an example — more of the process side that we’ve walked through today.
AI Decisioning Studio is a feature that is kind of a tool in and of itself, right? It’s a very, very powerful tool, but it requires a lot of thought, planning, and execution to make sure you’re using it the right way. So that’s one that I would say is more for advanced use cases or advanced companies. But before that, you could definitely jump into other elements of Braze and still leverage AI. Troy, anything you would add there?
Yeah, I agree. AI Decisioning Studio is definitely very powerful, but you need to make sure you have the right plan going in. And it is a little more advanced. You need the right resources to support it, but that would be it.
Cool. Well, thank you very much, Troy, for doing this with us. We really appreciate it. And thank you to everybody who joined — we really appreciate it as well. We love this stuff. This is our dream job, so we really appreciate you coming and we’ll share the recording afterward. Thanks everybody.
Thanks everyone. Bye.
