Speaker 0: Hi, everyone. I hope you’re having a great. Um, my name is Amber Juri Amel. I’m on the team at, and I’ll be here moderating this session, build AI driven journeys in marketing cloud. Um, I have Hunter Dunbar here from Cervelo, and he’s going to be presenting, uh, the session today along with Tim Ziter. Um, I just wanted to give a couple housekeeping reminders. If you have any questions, please put them in the q and a tab. And at the end of the session, if we have time, we’ll, um, get your questions answered. Uh, thank you. Take it away,

Speaker 1: Hunter. Great. Thank you, Amber. Um, you can see my screen. Correct? Just wanna double check. Okay. Great. Yep. I can see that you can. That’s awesome. Alright. So, um, my name is Hunter Dunbar. Uh, I’m gonna be joined shortly by Tim Zider. Tim and I work for Cervelo. Uh, we are a wholly owned subsidiary of um, Kearney, management consulting firm, but we specialize in, uh, bunch of different stuff under the Salesforce umbrella. But definitely, uh, marketing cloud is a big part of what we do. And today, we’re gonna be talking about all the new features and, um, solutions that you can build with Marketing Cloud and also how you can augment Marketing Cloud journeys with, um, AI data from Salesforce. So, uh, thank you to all the sponsors. Really enjoyed presenting at this conference and and really appreciate being able to connect with people virtually like this.

So we’re gonna talk about the new datasets, which are are really emerging as as core pieces of marketing journeys. We’re gonna talk about how to leverage those datasets using generative AI. We’re gonna talk about some of the new features that Salesforce gives us to create Gen AI solutions. So, specifically, we’ll talk about prompt builder, uh, and then we’re gonna talk about how this all comes together in Marketing Cloud and what are some of the other features that can be leveraged in Marketing Cloud.

So this slide is supposed to be have animation on it, but, um, let’s talk about marketing cloud journeys. So traditionally, with marketing cloud, we have, uh, contacts, opportunities, accounts, and Salesforce. Uh, we can pull in external data. We can pull in case data, all the kind of standard CRM data model that everybody here knows and loves. That gets integrated to Marketing Cloud using the Marketing Cloud connector. We take that data, we model it into, uh, data extensions, uh, transform our our datasets as they need to be transformed, uh, and then we feed those into journeys, and journeys create these river deltas of different outcomes that can occur. And so we’ve got decisions that are based on criteria in the in the dataset. Well, what we’re gonna talk about is and that that’s traditional kind of core demographic data about about our customers. Um, well, we have a new emerging source of data, which is called interaction data. I’m gonna talk a lot more about this in a bit. But as we introduce, uh, intelligent insights based on the interactions that we have with our customers, we’re going to be able to introduce additional tributaries and paths that our customers can take as we as we guide them down a a marketing cloud journey. Um, so with our traditional journeys, we’ve got, you know, basic demographic data about our customers. We have all the traditional tools that are available in marketing cloud. But with AI powered journeys, we still have that demographic data. Um, but we’re gonna talk about a bunch of new features that can be built on top of traditional marketing Cloud features and functionalities, so advanced segmentation, um, one to one content at scale, uh, plug in tools like typeface, advanced CMS. Tim’s gonna talk about this later in the in the talk. Um, and then we also have interaction intelligence. So this is getting more data points which can be used to segment our customers more intelligently.

So let’s talk about transcription. We have been, uh, writing down, uh, spoken word, um, like, since the dawn of man. We’ve been writing in caves. Right? Passing on stories. And throughout history, we’ve had, you know, a bunch of different technological advances in this area. Um, but definitely in the past hundred years and specific specifically in the past um, ten to thirty years, I think we’ve had we’ve had major, major advancements. But, you know, we go beginning in the early nineteen hundreds. We we have the the introduction of the typewriter. In the, uh, mid twentieth century, you’ve got, you know, these these old pictures of, like, armies of, uh, typist transcribing calls and transcribing information using typewriters, putting that onto, you know, just a piece of paper, and then it goes into a filing cabinet. Um, well, if we fast forward to what we have today, uh, you can speak and, uh, the, you know, the the words that that you speak or the words that are in a conversation can be almost immediately transcribed into, um, a data layer on your, you know, cell phone, your machine, or in the cloud through Zoom integration or or what have you. Um, and the big unlock for this is is machine learning and, uh, that computers are better at kind of understanding how humans talk and and how that maps to, uh, transcription. And so in the past, uh, since 2014, you can see, um, that the error rates on transcription have have plummeted. I remember, like, you used to do talk to text on your phone. It would be just completely riddled with theirs, and you would send a text message to a friend, and you would send something that was completely wrong. Um, and if you do text to text on your phone now, like, it it’s it’s much, much better than than ever before.

So what are the traditional tools that Salesforce gives us to leverage on structured transcription data? Because now we’ve got these large blocks of transcription data, but what are we gonna do with it? Um, well, Salesforce provided a couple tools in their old Einstein AI. And before that, it was MetaMind. Couple different tools to to analyze unstructured text. So named entity recognition takes a block of text and pulls out, like, the think of it like the proper nouns. So Marc Benioff, the CEO of Salesforce, gave the keynote speech at the conference in Paris last week. Um, that’s gonna pull out a bunch of entities from that block of text. So Marc Benioff, CEO of Salesforce in Paris. Um, well, when Tim and I first started kind of thinking about this problem of, um, how to analyze unstructured text, uh, we were thinking about it in the context of a person working with a a wealth adviser. That’s a very kind of uh, personal and human conversation that you could be having. And so we were trying to we’re using kind of like a traditional NER tool, and we said, good to see you again, Alex. It’s been a few years, so I’m pleased we can get caught up on where things stand today across your financial picture. Well, named entity recognition is just looking for entities, and so it sees good to see you again. And it recognizes a a Whitesnake album that I guess there was a Whitesnake album called Good to See You Again. So this tool’s imperfect, and it’s not exactly gonna work for what we need to do. Then we’ve got sentiment analysis. Uh, I think most people are familiar with this. It takes a block of text and just says positive or negative. Is it good or bad? Um, on, like, a kind of microscale, it’s not that helpful. Like, a macroscale, if you needed to analyze, you know, 200,000 tweets, I guess it can give you a good thermometer of how your organization is doing. But for our situation where we let’s say we have a Salesforce organ. We have a ton of call transcripts and we’re trying to get insights from those call transcripts. Well, sentiment analysis isn’t really gonna help us. But generative AI will. And so with GenAI, we can take an unstructured block of text, send it to an LLM, and say, give me the entities that were discussed.

So I’m gonna skip past this one. And, um, let’s talk about the application that, um, um, I’m going to demo, which is, uh, how we kind of build our AI driven insights. And this is going to, um, take a bunch of call transcripts. And, uh, let’s say that these call transcripts are between a wealth adviser and a potential client. And we’re gonna take an unstructured block of text, and we’re gonna say from this block of text, give me what financial goals were discussed, uh, what products were discussed, did the customer raise objections, uh, and were there any follow-up actions defined? And so what we what we get is we start with a bunch of just call transcripts, documents, and we end up with a bunch of structured data. So, um, we have, like, who the customer is, what were the individual products they discussed, what were the individual objections, etcetera. And what this is gonna let us do is we can report on our data, and we can create insights about the types of things that we’re talking about, and then we can provide sales coaching. And, well, the the third thing that should be listed here is we also can create better, more intelligent segmentation in our journeys.

Okay. So when we first started building this application, uh, we did it all custom. So I’m not sure that I’m gonna be able to zoom in very well here. But essentially, uh, the application that we built integrated with a Python API hosted on Microsoft Azure and integrated out to the OpenAI chat completions API. Um, on the Salesforce side, what we’re doing is we’re taking a call transcript. We’re logging it to an object. We’re then, uh, taking that transcript and copying it over to an object called an interaction summary. Uh, and that interaction summary triggers a call out to our Python API, which calls OpenAI, calls the LLM. The LLM is gonna give us a response. We’re gonna save that back to Salesforce, and then we’re gonna parse that response into structured data which we can report on. So we did all this work and got it to work, worked really well. Uh, and, you know, then Salesforce released prompt builder. You know, while we were building this, I think prompt builder was getting released, but I’m sure many people on the call here can commiserate that Salesforce release prompt builder. We were like, wow, that looks really great. That’s cool. But, um, there’s no way we’re gonna be able to get access to that until, you know, 2026 is probably when it’ll become generally available. Um, well, I was able to get access to a a beta version of Prop Builder, and, uh, uh, we essentially took this whole application architecture and rebuilt it using prompt builder in about four hours. And so what did what did that actually entail? Well, in our old application, we had call transcripts saved in Salesforce. We had an Apex trigger, which was taking those call transcripts, passing it to an API, which was calling an OpenAI API. And that OpenAI API was returning response, which we were parsing in Salesforce and then saving as as objects in our database. Uh, well, with prompt builder, we basically, you know, still have the call transcripts, still have a flow, which is calling the prompt builder. Um, but prompt builders is simplifying everything in the middle here. So everything involving creating an integration, creating a call out to, uh, an LLM, getting that response back, and it’s also doing it within Salesforce’s trust layer. I don’t know if Salesforce’s trust layer I don’t know if they’re actually calling an OpenAI API as part of that or if they have their own hosted version of chat GPT on a Salesforce server or not. Um, but Salesforce will tell you that it’s all within a trust boundary, which I think is much more advantageous than what we were doing back here, which is just calling an OpenAI API. So, um, I guess I just kinda talked about these security advantages, but we we definitely really liked this, uh, this solution and and and felt that it saved us quite a bit of time over what we had originally built.

Um, so I’ve got some code here, which I think will be available after the fact. But, um, um, before I pass it off to Tim, let me just share how this application kinda works end to end and talk a little bit about, um, prompt builder. So, uh, oh, you know what? I think I need to share my entire window here because I’m not sharing the right thing. Okay. So I will go to Okay. So if people are not seeing Salesforce, please let me know. But I think you should be seeing a Salesforce window here. And, you know, I talked in the talked earlier about this application flow where we’re creating, you know, we’re logging a conversation to either a custom object or a task. Uh, we’re saving it to an object called an interaction summary and then using an LLM to generate entities that, uh, represent components of that interaction summary. So we go from unstructured interaction, uh, transcript data to structured entities, which describe that interaction. And so back here, when I go to a lead, I’m gonna go here. I’m going to take call transcript. So here here you can see I’ve got a call transcript. I’ve got Walter who works for, uh, an investment firm. He’s talking to Lily. Lily is, uh, remember we were talking about a a wealth adviser and a customer. They’re gonna be talking about financial products. They’re gonna be talking about financial goals. They’re gonna be talking about concerns, and then they’ll have some follow-up actions at the end. So let me make sure that I’ve got all this copied, and I’m gonna log a new task against Lily’s lead here. And so I’ll log this call. Normally, there would be some automatic integration between a telephony system in Salesforce taking that call transcript and saving it somewhere. We’re we’re skipping that phase of this for the for demo purposes, but would not be a hard thing to to do. And when I check this extract transcript insights and click save, I’m gonna get a new interaction summary record. I’m looking at the lead, and an interaction summary record just got generated. So here’s my call transcript. All this data about these two people. You know, Lily’s describing all of her financial assets, goals, etcetera. And, um, we come down here, and we’ve got an API response. And this is still blank because this is an asynchronous process right now. But when I refresh, you can see I’m getting an API response back. That API response includes information about products that might have been discussed, some objections that might have been discussed, goals, and then finally, follow-up actions. So we get this JSON object back, and we’ve got code on the back end which is going to translate that into entities. Now I’ve got a nice structured version that I can use to analyze my call transcripts and nice structured data which I can pull into data extensions in marketing cloud. Um, before I move on, I’m gonna talk briefly about, um, Salesforce prompt builder. And so, you know, I mentioned that I got access to a a beta version of Prompt Builder. Normally, you would be calling Prompt Builder from within your Salesforce org. Um, but just to kind of make sure that I’m explaining every step of what I’m doing, our promptbuilderdemoorg and our marketingclouddemoorg are in separate places. So right now, I’m making a a call out from my marketing cloud org to the org that has prompt builder in it. You would never do this in real life, but it’s just to demonstrate the features and functionality of the of the applications. If I go to prompt builder, you can see we’ve got a bunch of different prompt builders available. I’m going to create a new prompt template, and, um, there’s a couple different options that Salesforce gives you. So you can generate results to a field. You can create a flexible prompt. You can do, uh, answers to, uh, uh, no answers for knowledge articles, or you can do, um, summarization of a of a record. Our prompt is gonna be a a flex prompt, and I can put in, um, a name for this. And I’m not gonna go through every step of this, but essentially, we can define a couple different types of inputs for the prompt. And in our case, what we’re gonna do is we’re gonna create a prompt with an input, which is a call transcript. And so, essentially, what we’re doing is we’re doing a bunch of prompt engineering here. We’re telling the LLM how we wanna get a response back. We’re giving it examples of the types of products it could it could detect, uh, the types of objections it could detect. And we are then at the very end kind of inputting the transcript. Um, and what prompt builder is gonna allow us to do is we can we can test this prompt in real time. We can hook it up to different types of models. So right now, we’re using OpenAI’s GPT four, but you could use an anthropic model. You could bring your own model, etcetera. So in the interest of time, I’m going to, uh, I’m gonna hand this over to Tim, but I’m happy to answer any questions about this this solution, um, later in the, uh, q and a portion. Alright, Tim. So if I stop do you want me to go back to the slide where that’d be best?

Speaker 2: Sure. Yeah. That that’d be okay if if you wouldn’t mind. Sorry. I’m late.

Speaker 1: Yep. No problem. Okay.

Speaker 2: Alright. Thanks, Hunter. So let’s talk about how we can take this amazing, uh, AI driven data and make it actionable in marketing cloud. We’re gonna cover doing this in three separate ways today. One is going and creating that those advanced, uh, using it to create advanced segmentation and those custom journey paths. So we’re gonna take that amazing data in Salesforce and bring it into marketing cloud via the contact builder, uh, those synchronized data extensions. With that, we’ll use those those SQL queries to transform that data to determine which segments should be entering in the journey and then which custom path should they be going down. The second part is delivering that one to one content at scale, where we’re gonna use those cool new features in Marketing Cloud to generate that one to one, uh, text content, use the typeface tool to create those images on a one to one basis, all pulled together with Amscript using a data extension content management system. And then last but not least, we wanna do those next best follow-up actions so that sales knows the best way that’s gonna work for them to close that deal or follow-up with a lead. And then obviously the ones, uh, the the steps that it’s gonna work best for to to maximize that conversion. So we’ll go to the next slide, please.

So before we dive into those three areas, let’s take a look at this data that’s that Hunter has created in the Sales Force side. So on the left, we’ve got all those Salesforce objects. We’ve got those standard objects, the campaign, campaign member, lead, and account. And then you can see that relationship, the linkage between the lead ID and that interaction summary, that new object that Hunter’s created. So with that connection to the lead object that’s being used for the campaign member, uh, that opens up all those entities. And then all that data is flowing into the center part here, the marketing cloud side, which is gonna do a couple of different areas of functionality. Number one, we need those key data points that are gonna give us ability to send that that, um, that email in this example. So we would need a subscriber key. Uh, we would need an email address. Another data point that we’re gonna look to wanna pull in is those those data points is gonna allow us to create those custom journey paths and also personalize the the email. So you can see the the personality data point there is gonna drive the the way the content is displayed. Uh, the call to action is gonna show intent, etcetera. And then last but not least, marketing cloud is gonna be, uh, utilized here to create those custom Salesforce objects for those next best follow-up actions, uh, with sales. So utilizing, um, having the owner ID, for instance, that’s gonna come that’s if you need to create that task or the account ID that’s gonna be fold folded in there if you need to add and create an opportunity. Let’s go to the next slide, please.

Uh, so let’s take a look at this data, the the data that Hunter’s bringing into the marketing cloud environment. We’ve got four different examples of this. The first one is using that data to drive intent. So with using this time frame category, we’re gonna be able to listen to hear for, you know, is this person interested in us reaching out in to zero zero to thirty days, or they may be more interested to talk in six plus months. That’s gonna, you know, signal intent, uh, that could influence on the segmentation side, a ranking of which, uh, type of segment that would fall into. And then it also might impact the creative, uh, with the call to action as well. Another good example of the the data is the, um, the different, um, product categories or and of that next best product. What’s gonna be the most, uh, relevant product that’s gonna resonate with those, uh, with that subscriber. So, um, in this example, we’re looking for that category. Uh, what category is is mentioned the most, uh, related to all those different entities based on those interaction summary? Once we find that that that category that’s mentioned the most, then we’re gonna look for a flag that’s gonna determine, uh, what’s that product within that category that we’re wanna wanna bubble up. Now that’s gonna impact the the content that might also impact how you’re gonna be communicating to sales for them to follow-up with the appropriate product on that.

Next slide. A couple more data example options. Uh, this is how we could determine that different personality or the personas. Little bit more esoteric here where we might be listening for different types of words, different combinations that might influence things like the imagery that might be in the email that’s gonna be generated by that typeface, uh, content block. Or it might be some notes that might be in a task, uh, for sales to follow-up knowing the overall feel, uh, you know, keywords that the, um, the subscriber has been voicing there. And then last but not least, the the different I different indicators for what’s gonna be the best way to follow-up, uh, with this lead. What’s gonna be a a way that they’re gonna resonate, um, you know, whether that’s a phone call or an email address or maybe scheduling another meeting. Pull that out. That’ll be populated in the segmentation and then also in the new notifications that sales will use to follow-up.

Go to the next slide. And let’s take a look at how this, uh, could be pulled together, um, in using SQL in, uh, in marketing cloud. So let’s say this this company has a sales tool, uh, ranking tool like a SalesWings, where they’re pulling in this predictive score where it’s either cold, warm, or hot. Um, so in the center of this query, we’re using that entity detail, that time frame entity detail. And if that entity detail is, you know, zero to thirty days, then we might elevate that rank to a hot rank. Whereas if it’s six plus months, it might, you know, lower that ranking. So that’s what that center of that query is doing, that nested query. On the outer part, we’re using a case statement that’s combining that predictive score from SalesWings and the new data that we’re bringing in from, uh, from prompt prompt builder and the LLM model to create this Uber ranking, where if somebody has a hot ranking from, uh, SalesWings, but they’re showing really low, uh, low intent, then we might elevate them from a hot ranking to more of a warm or a cold ranking. Same thing if them someone was showing, uh, you know, a cold ranking from SalesWings, uh, we might if they’re showing high high intent, we might elevate them to a hot ranking overall. And then if we go to the next slide, we can see how this might manifest in a journey itself where we’ve got three separate paths here, um, depending on on those different segments. So we have a hot path that might go directly to sales to send a hot lead notification, uh, to let them know that, uh, a lead is ready to talk. It might create a task. We’ve got a center path here, which is that warm ranking, which might send an email directly to, uh, to that lead to gauge that intent. If they engage right away, they get moved up to the warm to the hot path. And then if they don’t engage, they probably could just follow down a cold path where they’re gonna be served up a variety of different communications over time. And not only will this determine the different paths, but we’re gonna use the data that’s, um, that’s being enhanced by the l m model to, you know, to customize the email content, also to indicate different, uh, paths like a decision split in the journey itself.

Go to the next slide. Alright. Let’s talk about this one to one content at scale. So I’ve got two great options, uh, right out of the box in marketing cloud under the Einstein Copy Insights family, where you can take create personalities based on the personas of your brand. With those personalities created, Einstein will create, um, a variety of different you know, five different versions of subject lines, body copy, all based on those, uh, those different personality types. And then, uh, not only will it create them, they will improve and and, uh, improve those different variations and different copy points, uh, over time. And then that combined with the typeface content block where you can create one to one images, um, for your brand, uh, no longer having to go photo shoots or look for stock images, Typeface font the the Typeface content block will create them for you. Uh, you can really truly create those one to one, uh, you know, uh, content pieces at scale.

So let’s go to the next slide. Let’s take a look at how we can pull this together in two different ways. One is using this next best product, uh, query. So what we would with something as simple as the, uh, the lead ID from, uh, from your, um, your communication, that can unlock all of those entity details, um, from those interaction summaries, and you can pull in all of this data. We’ll once you sort of unlock the the interaction summary, you’ll wanna do something like, um, have a separate query down below that’s going to determine, uh, for all of those different categories, uh, what’s the top category for that subscriber and then also within the category, the, uh., the top product. So with that determined in a date extension, you can bring that back up into, uh, the the AMP script and use, um, you know, look up ordered rows to find that top, uh, category. And then once you find that top category, the top product, uh, then that would be rec you know, that would give you a recommendation for the the product to feature, uh., prominently in your communications. And then the next, uh, next slide, the next example is using, uh, Amscript to build that AI driven content management system. So a couple of things you need before you actually do this in Amscript. First, you’ll need a query to determine the best personality fit based on those, uh, different entities that you’re listening for. It’s very similar to the query we’ve just looked at in our in a previous slide. Once that’s built to determine which, uh, personality to your each subscriber fits into, you’ll wanna create a, uh, content management data extension that has broken down by each of those personalities, You’ll wanna list the subject line, the body copy, even the actual URL images from the typeface content in that date extension. So with that created, you could use, uh, something as simple as, you know, random function in AMPscript to find that personality for each subscriber and then sort through and find one of the five or so different combinations of that text and imagery, um, and then set those values so that in the email itself, it displays, uh, the values that are in that data extension content management system. So if we go to the next slide, it’ll just sort of give you an overview of how this comes together in the emails themselves. Um, so here’s the same email, uh, two different audiences. So on the left here, we’ve got Abigail 0026 interaction summary. She’s at retirement age. Uh, she’s showing high intent because she’s ready to get serious about her retirement or those products. And then we’ve got Victoria here on the right, 0027 interaction summary. She’s a young professional. She’s probably showing lower intent, you know, certainly important to her, but not as important as Abigail. And you can see from top to bottom, these emails are dramatically different. So you can see the image at the top resonates to the different audience, the different, uh, personalities that they might be sharing in their in that l m driven, uh, content. Section number two, you can see that that, uh, Einstein copy that comes through that body copy that’s determined by, uh, by that personality. Number three is that next best product logic in, uh, in Amscript to determine which products to feature. And even sections four and five, based on that next best follow-up decisions, you can customize the call to action. You can include a follow-up steps. Like, for Abigail, she might wanna reach out to a sales rep. Where for Victoria, based on her personality type, maybe not so much. Maybe she’s just gathering information at this point, so not to be too salesy. So example of the same email, dramatically different for those two audiences on a one to one basis.

And then next slide. The last section we wanted to cover was those next best follow-up actions. So taking and listening for what’s gonna be the best way for sales to follow-up. You can use those data points in something as simple as a decision split in your journey. So you could create a task if that’s the best way for sales to follow-up. You could update an opportunity or create a new opportunity if that’s, uh, if that’s preferred. You could even create a case if it’s not a sales related, um, decision. It’s something that’s more, um, something that service would would, uh, be more appropriate to follow-up with. Um, and not only creating these different objects that would resonate with sales, but personalizing with all of that data. So you’re serving up, you know, what’s the best product they’re interested in, what are the different personality points. There’s a lot of personalization you can do, customization of all of those objects. You can sort of see here using that handlebar code, uh, in the object activity to bring in those different personalization types so that sales isn’t just getting a generic task. They’re getting one that’s got all of this information for, say, they’re following up with, uh, Victoria. Uh, they’ll they’ll know the best about her. Um, so that wraps up those three sections. Um, so hopefully, that gave you a good overview even though I I missed Hunter’s section. Um, I’ve I’ve heard it. It’s been fantastic. Hopefully, it gives you good an overview of the the history of where we’ve come to this date on the transcription history and LLM models. Hope we give you a great overview of how to take this GenAI architecture. Uh, the prompt builder demo is so cool. Uh, and then hopefully, I gave you a couple of ideas on the marketing cloud side for different ways to see this pull together on the Marketing Cloud side, uh, leveraging that data, uh, from Salesforce. Guess we’ll we’ll open up for questions.

Speaker 1: Not seeing anything in the, uh, q and a channel here. So I think we’re we’re probably ready to wrap. Tim, I think our the presentation is gonna be recorded in public for, you know, people wanna review this later, but you and I are also accessible on LinkedIn or, you know, our contact information is not hard to find if people have follow-up questions. So

Speaker 2: Sounds good.

Speaker 1: Awesome. Thanks very much, everyone.

Speaker 2: Thank you.