In this Braze Innovation Series session, the Stitch team explores the rapidly evolving landscape of AI tools in Braze — focusing on what’s available today, how it actually works, and where it can meaningfully improve customer experiences. From a broad overview of Braze’s AI capabilities to a deep dive into the newly released Agent Console, we unpack how teams can start using AI in practical, production-ready ways.
You’ll learn about:
- The full suite of features that make up BrazeAI™, with a high-level overview of where each tool fits and where to focus if you’re just getting started with AI
- Agent Console: what it is, how to get started, and high-impact use cases that focus on enhancing your existing marketing strategies and workflows
- Designing and implementing AI-powered experiences with Agent Console, including how to embed agents in a Canvas or Catalog and leverage them for more dynamic, real-time interactions.
Throughout the session, we share practical guidance, implementation considerations, and lessons learned from applying these tools in real client environments. If you’re looking to better understand BrazeAI™ and start using Agent Console in a way that’s grounded, strategic, and easily attainable, this session is a great place to start.
Transcript:
Thank you everybody for joining. Welcome to our innovation series. If you haven’t joined us before, where we deep dive into Braze functionality and solutions.
My name is Bobby Tichy, and I lead our solutions team here at Stitch, and I’m joined by Steven Rosenfeld, who is our CTO. For those who aren’t aware of who Stitch is, we are a consultancy that’s focused exclusively on Braze. So it’s the only platform that we consult on, and we help marketers use it to get the most out of their martech stack.
There are three things that we’re going to focus on today. First is a general overview of Braze AI and all the features within it. I think a lot of people overlook just how much AI Braze has launched over the last couple of years. Then we’ll spend the majority of the time today focusing on Agent Console. Agent Console is a feature that went GA a couple of weeks ago from Braze that allows you to implement agents inside of your canvas workflows as well as your catalogs as well. So there’s a lot of good functionality that we’ll walk through, and as part of that, we’ll go through our use case deep dive as well.
So the first thing we want to go through is just a quick overview of all of the different AI features that Braze has. So on this slide, you’ll see the nine key areas of Braze AI. I’m starting at the upper left, then working our way down. First is the Braze MCP server and the content optimizer. Both of these are currently in beta, so they are not generally available quite yet. But the Braze MCP server will allow brands to be able to interact with Braze in a more flexible way than just from using the current REST API that’s in place today.
Next is the content optimizer, and the content optimizer is an agent that allows you to build content much faster and optimize that content much faster than if you’re doing it manually today.
The third piece is Intelligence Suite. This is probably what most of you are familiar with within Braze, which is intelligent timing, and then also intelligent channel, and then also personalized paths, where you can make sure that you’re sending the messages to the right people, at the right time and on the right channel.
Fourth is Operator, which was recently released as well, which allows you to have essentially a chatbot assistant wherever you are in the dashboard of Braze.
Next is Generative AI. So the copywriter for helping write subject lines or push notification content, the liquid assistant, as well as the image generation.
Agent Console, we’ll dive deeper into this the rest of the time we have today. Item recommendations and Predictive Suite, allow you to provide specific recommendations for what products somebody should see, within any kind of a message that’s deployed from Braze. And then next is the Predictive Suite, which allows you to have churn prediction as well as likelihood to convert. And then lastly is Decisioning Studio. Decisioning Studio is the acquisition that Braze made of OfferFit last year, that allows you to create an action bank of anywhere from forty to eighty different offers. And the machine learning and the forward deployed engineers at Braze will decide which version of that content and those offers should be sent to the right person. So truly one-to-one personalization across those different channels.
So let’s go ahead and dive into agent or the kind of three things that we think are most beneficial for people to use when it comes to AI within Braze. There are three components here. First is Intelligence Suite. This is the lowest barrier to entry to use AI within Braze. And using intelligent timing and figuring out when people are most likely to engage, or convert with our messages is very easy to implement at a catalog level or at a campaign level.
Next is the creative variant and then also the channel itself. So if someone is subscribed to every single channel, how do we make sure we’re sending the message to the most likelihood for them to convert or to engage on that specific channel?
Next is Agent Console. Like we already mentioned a little bit, the Agent Console allows you to create an agent inside of Braze, and we’re going to walk through a specific example during the use case deep dive here in just a moment. But the really cool thing about Agent Console is you can use this inside of a canvas, within a specific step, or you can use it in a catalog as well. So for example, you could create a net new custom attribute based off of all the data we have on the Braze profile, or we could create segments. So let’s say that we wanted to create five different segments to bucket our subscribers and our customers in. We could use the Agent Console to build that out.
And then lastly is Decisioning Studio. Like we mentioned before, this is the acquisition of OfferFit. The Decisioning Studio element is a little bit more of a heavy lift when it comes to the initial implementation, just because the machine learning needs some time to learn from your data. But the actual implementation of it is very straightforward. It just takes some time for the model to understand your perspective and understand your data before it jumps into everything else that it needs to. So this is where if we were you, obviously there’s a lot of elements to this as well of where you are as far as marketing maturity is concerned, where you are in the customer life cycle, all those things. But these are the three that we found our clients get the most value from within Braze AI.
So next, let’s jump into the Agent Console fundamentals. So first, we’re just going to talk through what it is, how it works, and some key considerations before we jump into the use case. So Steven, I’ll turn it over to you.
Hey, thanks, Bobby. So when we think about what Agent Console is, it essentially allows you to interject an LLM conversation into a marketing workflow. And so I think while a lot of people may have not used Braze’s AI functions in the past, most everyone has interacted with an LLM at this point, right? And the LLM always does its best job to answer the question you’ve asked, which means I can ask it, “Hey, what’s the best time to send Steven an email? What time should Bobby receive this text message?” That doesn’t mean it’s the best job for the tool or for the use case or the best tool for the job, right? Because Braze has released other AI tools that are specifically designed for that.
And so when we talk about exactly what Agent Console is, we bring that LLM into the workflow, and that’s going to allow us to configure those agents to access various data points across your account. And when that agent runs, it’s going to use your prompt and the Braze data to create a unique response that is tailored to your specific use case and your specific campaign for that specific individual.
And so, of course, there are considerations we have to make when we’re trying to do something that’s sophisticated, and that’s going to be to understand exactly what data needs to be prepared in the account ahead of time. You can start by thinking about the type of campaign you want to run, and sometimes individuals or our clients are really looking for an AI use case. What’s the best use case for Agent Console? And it’s always fun to try to take a step back and think creatively and find something net new to provide value to our consumers. But sometimes it’s also best to just look at the use cases you have running today and say, “How can they be enhanced if we put an LLM in the middle of this?”
And that LLM— I was going to say, Steven, I think that’s a great point because I think a lot of times when people think of an AI use case, they think, “Well, what kind of data do I need to collect, or what do I need to add to Braze to be able to do this?” And I think that’s what’s great about Agent Console, is there’s nothing net new that you need to be able to do. You can leverage just the existing data you have on the profile or an existing canvas or existing catalog.
Yeah, exactly. In those situations, you’ve already thought through how you would use that data, but it may be slightly enhanced by adding an LLM in steps just before a key message or before a key decision split. And the LLM that you’re interacting with, Braze, of course, integrates their own in-house LLMs, but you can bring your own as well. So if your organization already has a contractual agreement with Anthropic or OpenAI, Gemini, et cetera, you can configure that to be brought into Braze, and they’ll send those messages to the LLM of your choice. At the end of the day, however, that LLM’s only going to be as good as the data you can provide it. Otherwise, it’s falling back to the data that it was trained on, and the more context that you can give an LLM, the more specific of a response it’s able to provide for your use case, for your individual, for your campaign.
So that takes us to how do we actually work with an agent in Braze? And there’s three key steps, and we’ve kind of hinted around these already. There’s the account-wide configuration. So you have to make sure that you do have the catalog feed set up, that you have the appropriate data points on your profile. That’s not necessarily specific to Agent Console, right? You’re probably using those catalogs and those profile attributes today. But if you do need something new, obviously, we need to set up a new feed at the account level and make sure that Braze has access to them.
Then we’re going to focus on campaign-specific data collection and organization. So that might be data that you need to drive decision splits within your campaign or specific attributes that are for this campaign that you need in order for the agent to be successful and be able to make an informed decision. And after you have those two data sets, we can move into the agent configuration itself.
Perfect. I forgot to mention earlier as well, if anyone has questions or thoughts, feel free to just throw them in the chat at any time as we’re chatting. Don’t feel like you have to wait until the end, and we’ll try to answer them in real time. And anything we don’t get to, we’ll answer at the end as well. So we’ve gone through the Braze AI overview, all the different features there, what we would recommend you start with, and then the components of Agent Console. So let’s bring this to life through a specific use case that we have here.
So for a retailer that we worked with, there was a specific challenge that they were trying to overcome, and that was when they had a new customer that was going through their onboarding series, it was a lot of information. They had a number of influencers. They had a different number of different products that they wanted to promote. You can imagine any new customer joining any of our brands for the first time, it can be a lot, especially when they’re probably not thinking about all of the things that we’re thinking about as marketers that they could be looking at or that they could be interested in moving on past their first purchase.
So the goal here was to really simplify that experience but also make it even more personalized than it was before. So first, how do we kind of pare down the messaging, make each message that we’re sending as part of this onboarding series just have one specific call to action, not seven or eight or different recommended products, but really each message have intent to it. And so what we did was we leveraged a Braze agent within Agent Console to pair each customer as they signed up with a specific influencer of this brand. So that way, as they continued on with this brand, they had someone that was kind of their almost like ambassador for this brand as they went through the process.
So what this looks like in practice across the solution, the first thing that we did was we just identified the relevant data on the profile. So even if someone hadn’t necessarily bought something before, as soon as they make that first purchase, there’s typically quite a bit of information that we have on the Braze profile already. That could have been from anonymous views that they had had, as well as things that maybe they added to cart but they didn’t quite purchase yet. All of these different things that we have access to. From there, we configured the agent as the first step within the canvas. So that way, as soon as someone made a purchase, the first thing that would happen is that agent would assign the relevant influencer back to that person. And then from there, we deployed and built out the rest of the messaging in that canvas.
So, the couple of key things to call out here is you don’t need new data, right? And like Steven mentioned, the output is only going to be as good as the data that you give it. But in this scenario, what we had was a catalog full of all of the influencers, more than 50, that this brand could potentially put forward to their customers, and then let Agent Console handle the heavy lifting of determining which influencer was right for that person, and then also manage it ongoing. Maybe that influencer changes over time after someone makes the second purchase or the third purchase or their tenth purchase.
One of the things that, again, that stands out to me is that you don’t, to Bobby’s point, you don’t need that net new campaign. We’re going to focus on enhancing one already. So, the influencer catalog was already in place. And all we then had to do was do a review of all of the data that’s going to be fed into the agent. And so with that, we did need to create brand guidelines that are going to be used by the LLM so that the system itself understands what words it shouldn’t use, what guidelines it should follow, and has the account-wide context. And then we’re going to narrow in within the actual agent’s prompt itself to provide the instructions around how to go and fetch an influencer from the catalog and pair it with the individual.
And so after we organize all of the data, then we start to say, “Okay, how are we going to actually capture the data itself?” And for this client, we had an in-app message doing this. And sometimes that may have been actually done through a preference center or landing page in the past where you’re surfacing one question at a time and trying to understand a little bit more about your customer. And that might influence the type of drip campaign you send them into, or we always talk about how the data that the consumer gives you should influence how you’re communicating with them. And in this way, we’re able to do it in near real-time, rather than a drip campaign that’s going to take 30 or 60 days to complete.
Yeah. So in this scenario, the company already had an IAM quiz ready to go. It was something that they’d already launched, and they were already writing it to the Braze profile. So even if you had never purchased before, they had really good conversion on doing this IAM quiz the first time you visited their website or the first time you visited their app. And so based off of this, we were able to take this information plus all their browsing history, and then all of the elements of what their first purchase that they made, take all of that together to ultimately provide the recommendation inside Agent Console inside the canvas itself.
Yeah. And so when we talk about creating the agent, of course, you need to define a name for the agent and description and select the model. But you’re also going to select at that point if this is going to be an agent that goes into a canvas and essentially impacts the consumer’s workflow, or if it’s going to be an agent that sits on top of a catalog. And we haven’t mentioned this too much at this point yet, but agents can sit on a catalog record itself as well. You can dedicate a field in the catalog to an agent so that anytime a record is inserted or updated in that catalog, the system automatically makes a request to the LLM using the prompt you’ve configured and then sets a revised value in that AI field. And so this allows you to provide detailed tagging or a description based off of other components in the catalog. You might want to do cleansing of a field that came to you as a JSON or as a structured data, and you just need to revise it a little bit and clean it up and get it ready for production use.
In our case, we’re selecting a canvas agent where we’re going to insert a step into the canvas flow that actually reviews the response of the user’s in-app message and cross-references the catalog that we mentioned. So we’re looking for what type of fabric do they like, and how frequently, or what’s their just various status points about the consumer itself. And then we’re cross-referencing that with data we have about the influencers. And so our prompt here is designed to do just that.
And because when you are configuring the Agent Console itself, you can select catalogs and fields to insert. We don’t have to worry about sending the entire catalog record into the agent and consuming a lot of tokens that can be expensive. We’re able to really narrow in on exactly what data we want to send to the agent and what records are most relevant. We’re also not going to send every record in the catalog. We’re going to narrow that data set down to maybe it’s the top 15 that are relevant for this consumer, and then letting the LLM decide from those 15 which one is most relevant.
And what we get out, or here when we’re configuring it, we also define what we want to get out of the agent, and that is driven by how we’re going to use that output in subsequent steps within the canvas that we’re defining.
Stephen, there’s a question from Valeriano. I hope I said that correctly. But to interact with the LLM or AI agent, do we have to teach it translating the meaning of each specific event or attribute before starting asking for a task? So for example, if I want to communicate a message after the third day after the customer activation, do I have to explain which of the events represent the customer activation?
Which of the events represents… So it would be best to give it that context, yes. You could infer it from the attribute name that you’re sending to it, but it’s not going to have all of the insight around what led to that value being populated. So that would be part of the instructions of the agent. So you have different layers of the cake, per se, right? So you have your brand guidelines. That’s just talking about in general, how you interact with your consumers. But I think what you’re talking about would be more of, in the instructions, telling it to review this data point and give it that understanding around that data point so that it can use the value that you’re referencing appropriately. Hopefully that answers the question.
Yeah, I think it’s a good point because I think the one thing to think through as well is that we’ve all seen a number of different Braze instances. Some of them are very clean, others have four different custom events or custom attributes just for a purchase activity, right? So, the more context, like Steven said earlier, that you can give the agent, the better it’s going to perform. Just like anything else that we do in marketing, right? The more deliberate we are with what we want the LLM to accomplish, the better the output.
Yeah, agreed. And you really don’t want the LLM to be making any assumptions, right? So the more you can spell out for it very clearly, the better your response rate will be.
Okay, so after we’ve got everything set up, you can actually run test examples as you’re setting up the agent to understand how all of this works. But in real time, the agent’s going to pull the content from the catalog. That’s going to be the influencer’s name, their image URL, some content, references, a brief description of them. It’s then going to take that brief description of the influencer, and it’s going to combine it and represent it through your brand guidelines. So that influencer’s description may not align with your brand, but because we’re pulling their description out of the catalog and asking the LLM to evaluate it against your brand guidelines, we’re able to properly represent the influencer within your brand context. And that’s then going to generate a set of content that is geared towards this individual, appropriately representing the influencer, but all within your brand context.
And so after we get that final output from the LLM, we are going to use that in a downstream message. And so that message, in our case, was a digest that came afterwards. So, we’re sending that to the individual a few minutes after they’ve completed the survey, and it’s incorporating it directly into their messaging based on their response.
And Valeriano had a follow-up question of— Yeah. “I have to teach, basically every time the AI asks me to insert what I mean with customer activation.” And yeah, I think Valeriano, the key thing to keep in mind here is that the LLM is not a mind reader, right? It does not know what you mean by customer activation. I don’t know what you mean by customer activation. Right. Is that a purchase? Is that engaging? Is that downloading an app? Is that subscribing to messaging? So you’re exactly right. If there is an event or an attribute that is called customer activation, then obviously it’s going to be able to deduce what you’re trying to do. But the more deliberate you can be with that information, the better.
And I would say also, it depends on what you’re wanting it to do with that attribute, right? If what you really need it to know is the date of their activation, and that’s what’s in the field, then you don’t have to tell it everything as to how that field came to be populated, right? If you’re just saying, “Hey, look at customers who were activated within the last six weeks, and send them, give them this response. Customers within a different window get a different response.” It doesn’t need all of the historical context as to how that attribute got populated in order to make something meaningful. It’s just going to use the value. But to say, to Bobby’s point, if you’re going to, say, reference the business scenario of look at users who have purchased recently, and that’s the data signal, it’s not necessarily going to be able to assume. You don’t really want it to assume that customer activation is associated with recent purchase.
So next is how do we actually utilize the output itself? So obviously, we’ve got the agent console inside of the canvas. We’re leveraging all of the data on the Braze profile like we’ve talked through. Next is how do we actually use it? So the really nice thing is, one, we can write it to a catalog, we can write it to a Braze profile and then reference it using Liquid. But all of the output from agent console is stored as canvas context. So you don’t have to write it somewhere else if you don’t want to. If you want to save it for later so that way you have it, so you know that Steven is matched to Mark, the influencer, you can absolutely do that. But you can then also use it in every single message or every single decision split or anything else inside of that canvas flow because it’s written and stored as canvas context.
So a couple of key takeaways as we finish up here. And like I mentioned, if anyone else has any other questions, feel free to just throw them in the chat. Thank you so much for the engagement so far. Mm-hmm. But number one, like Steven mentioned, start with the use case. What are we trying to accomplish? In the use case we walked through today, we were trying to increase the likelihood of a first to second purchase by personalizing the onboarding series for a retailer. So we want to start with what problem are we trying to solve? How can Agent Console help us do that in a way that provides more personalization, more engagement, or provides a way of marketing to folks in a different way than before?
Next is context is everything. That’s a really good point of just kind of what we walked through with some of Valeria on those questions, is that the better inputs we can provide it, the better outputs will be provided back to us. So the Agent Console is not going to necessarily know what customer activation is, especially if we’ve got multiple profile fields that might mean the same thing. So the better the context we can provide Agent Console, the better the output. The other thing to keep in mind here too is, think beyond just copy. I think as marketers, we just default to leveraging an LLM or AI for copy and for developing a push notification or email body or different variants of that copy if we’re trying to do multivariate testing or more personalized offers. But there’s so much more we can do for it, like in this scenario of matching an influencer or leveraging it to be able to create segments within our own subscriber base.
Someone asked, “Is Agent Console pricing based on token consumption or flat subscription rate?” Steven?
Yeah. So there is a limited usage that I believe comes with the license. Obviously, Braze is going to know the most current pricing strategy there. But I believe a portion of it is, it’s not like you have strict tokens like other LLMs, but you do get so many uses per month. And then there’s a difference if you’re bringing your own LLM versus using Braze’s LLM.
And then Carla asked, “As you think about the next wave of use cases, what’s the category of problem you’re most excited about that Agent Console can finally solve that wasn’t possible before?”
Man, that’s a great question. I think I get excited when I think about incorporating Agent Console into additional portions of the application and seeing just how an LLM can interact with it. So, a lot of the use cases that we see, at least in the beginning when we’re discussing them, could be done by a decision split. We don’t need an Agent Console step to necessarily do some Boolean condition that decides the direction in the canvas. But when we start doing things like this, where we’re actually asking AI to combine various components and create something unique for an individual, it’s really interesting to me. I think also, I’m just excited to see where else Braze is going to add Agent Console because it’s not going to be just canvases and catalogs forever, I don’t believe. I think that they’re going to continue to incorporate it throughout the app.
The piece of functionality I’m most excited about is creating data through Agent Console. So for example, taking multiple pieces of the Braze profile and then creating a new custom event or custom attribute based off of that. So for example, if in this scenario with the retailer, we have an influencer that’s matched to every person that makes a purchase, and that can be added to their Braze profile. That was something that was fairly difficult to do in the past, of kind of creating new data from existing data. Another really good example would be for streaming customers. So let’s say that you want to infer what kind of genre someone likes to watch based off of their viewing history. But because of VIPA laws, we don’t want to bring all that into Braze itself. But with Agent Console, we could read what kind of titles they’re looking at, not host or store any of that data inside of Braze, but then have a genre that might be their favorite or a second genre as well, so that way we know how to personalize campaigns moving forward. So I think the thing that gets me most excited is the ability to create brand new data from existing data with Agent Console.
Great. Well, thank you all so much for attending. We really appreciate it. If you have any questions at all, obviously, you can reach out to us anytime on our website or on LinkedIn, and we’ll share out the recording of the webinar shortly. Thank you all very much. Have a great day.
