Video: Agentforce Workshop: From Design to Implementation | Duration: 3604s | Summary: Agentforce Workshop: From Design to Implementation | Chapters: Welcome and Introduction (2.88s), Mentoring's Transformative Impact (170.21s), Data-Driven Matching Process (290.34s), Matching Process Challenges (619.785s), AI Integration Challenges (990.9s), AI-Driven Matching Features (1185.725s), Building AI Confidence (1387.185s), AI-Powered Match Optimization (1671.75s), Embedded AI Design (2358.18s), Smart Match Impact (2596.44s), Streamlining Match Meetings (2872.52s), Match Closure Reasons (3053.69s), AI-Enhanced Match Support (3108.11s), Human-AI Collaboration (3344.955s), Data 360 Integration (3428.435s), Concluding Remarks and Offers (3512.56s)
Transcript for "Agentforce Workshop: From Design to Implementation": Hello, everyone. Welcome or welcome back to another Salesforce Ben webinar, this time in collaboration with Coastal and Big Brothers Big Sisters of America. Today's session will be a hands on workshop all around agent force. As I'm sure you're all aware, agent force can drive real productivity gains in Salesforce, but typically, many teams can introduce unnecessary friction by starting with conversational AI instead of optimizing their existing workflows. But, hopefully, by the end of today's session, that won't be the case for you. You'll be leaving with a framework to identify your first use case for embedding AI into your workflows, the technical patterns to build it along with a business case to get it approved. But before we get started, a couple of housekeeping bits. This session is being recorded, and it will be shared in our follow-up email. You are in listen only mode, so you won't be able to come off mute throughout the session. This is just to streamline the webinars. We've got a lot to cover, but please do interact in the chat as we move along through the session. And there is also a q and a box. There will be time for q and a q and a at the end, and please submit your questions throughout. So if we move on to the next slide, I just want to cover a little bit about Salesforce, Ben, in case you're new here. So we are building towards a future where there is a clear path to success for anyone wanting to embark on or accelerate a career within the Salesforce ecosystem. As and as a part of that, we offer practical, relevant, and accurate information through many different mediums. We've got articles, our newsletter, events both in person and this virtual session as well, YouTube, and our podcast. So there really is something for everyone. And if you haven't already had a look through any of what you see on screen, then don't forget to check them out later at salesforceben.com. So without further ado, I would love to invite to the stage our speakers for today's session. We've got Travis Gibson. He is the chief technology officer at Big Brothers Big Sisters of America and. the senior technology intern with deep experience advising organizations, LVMH, De Beers, Cox Automotive, Michael Kors, and Vera Wang. Earlier in his career, he held IT engineering roles at Amazon and Choice Hotels, and most recently served as head of cybersecurity infrastructure and IT operations at Make A Wish America, where he led enterprise technology and security initiatives supporting more than 10,000 employees and volunteers. We also are joined by Sam Taylor, director of innovation and emerging technologies at Coastal. As a technologist with over a decade of experience in Salesforce consulting, he helps organizations identify the real tangible value in new to market technologies in the ecosystem. Welcome, Travis and Sam. Great to have you here, and I will hand over to you. Awesome. Thank. you for for having us here today. Let's go ahead and jump in. I'm I'm so glad to be joined here by Travis Gibson of Big Brothers Big Sisters of America. It's a very, very valued client of Coastal's, and we're so happy to be a part of their mission and really helping address and improve their reach and impact nationwide. Travis, thinking about, you know, the size and scale of the organisation that you have, can you talk about, you know, in your role, what does that mean to really have to manage the technical complexity of an organisation that serves over 5,000 communities and over 200 agencies? Yeah. Well, thanks thanks for having me. You know, I I would just say in regards to your question, technology at BBDSA isn't about efficiency alone. It's about protecting trust, accelerating connection, and improving outcomes for the kids that we serve. So just a little bit about Big Brothers Big Sisters of America. We are the largest and most effective youth mentoring organization in The United States. Our mission is simple, but it's hard. We create and support one to one mentoring relationships that ignite the power and promise of youth, and our outcomes are measurable and long term. You know, youth who were littles, they earn 15% more over their lifetime than peers who weren't mentored. They're 20% more likely to enroll in college. Being a little closes two thirds of the economic gap, and more than that, they're just less likely to use illegal drugs, drink alcohol, skip class, get into physical fights. And I think all this is why what we're talking about today, which is gonna be match match quality between mentors and mentees, bigs and littles. This is why it matters so much. Better matches don't just feel good. Mentoring doesn't just feel good, although it does, it also changes life trajectories at scale for the youth that we serve here. That's that's amazing. And and and, Travis, you know, when I see when I see numbers like 92% belonging and 82% of Littles that can strengthen their level of engagement, there's a lot of data as a technologist I'm thinking about, like, how do you quantify that? How do you measure that in terms of your organisation? And as we think about what we're gonna be talking about today, which is integrating AI into your processes within an organisation to kinda help scale that up further, what's been your experience so far in terms of jumping into this organisation and really helping all these different agencies quantify the work that they're doing together. Yeah. I mean, we we've been around for a hundred and twenty years, so we have a lot of data as you can imagine, and turning that data into insights is one of the pleasures that I have working as the CTO for this organization. Really being able to quantify beyond, like I said, the feelings, right, which it does feel really good. But there's real real life impacts that are happening for our young people, but also beyond even the young people, the bigs, the parents and guardians, and that just trickles down to all of our communities. So let's let's talk a little about your your your transition, you know, onto the Salesforce platform historically and and working towards leveraging AI today. You know, I know that early on, in in Big Brothers Big Sisters engagement with Salesforce, had a lot of agencies that were kind of working independently that were following maybe a centralized rubric for performing matching, but there was a lot of work kind of needed to just, you know, lay the groundwork of centralizing and unifying that process and data. So can can you talk a little bit about the challenge that you encountered kind of early on and and your initial steps of of helping Big Brothers Big Sisters kinda standardize and scale up? Yeah. I mean, what you're seeing here are national averages across our federated network, which you mentioned over 200 agencies serving 5,000 communities. These are our national averages, not edge cases or outliers. We have incredibly skilled match specialists across the country. So the challenge wasn't effort or intent. It was kind of variability and process, data, and time to match. You know, some agencies match quickly, others took much longer, and youth experience depended heavily on where they lived. So, you know, matching is is one of the hardest jobs in our organization. Our our matching specialists were reading dozens of profiles for bigs and littles. They're parsing through long form interview notes, applications, assessments, running keyword searches, and and relying heavily on historical knowledge and memory. So, you know, when you see the the beginning of this journey, right, up to two years, it's it's really because of disparate processes and manual matching, incomplete or inconsistent data, and just a a really heavy cognitive load on our specialists. And then over time, we centralized data and workflows and match force. We automated outreach to complete those profiles for bigs and littles. It just created fewer unknowns at match time. And now trying to leverage AI, not not just for faster decisions, but for better starting conditions for decisions so that specialists spend time evaluating and not searching. And to call this out, know, this I know we've got a lot of attendees on this call. Not everybody is in in an industry and organization like what yours is. Looking at this timeline a little bit, want everybody else to think about, you know, this is a classic Salesforce org maturity timeline that we often see. You know, I've been doing this for over ten years at Coastal, and we've brought all sorts of companies and organizations across all industries into the Salesforce ecosystem and enhance their presence. And early on, was for so many organizations. It was, hey. We've got teams that are working in disparate systems. We've got people that are using spreadsheets. We've got data in, you know, 20 different places that we need to kind of bring in and unify, define that process early on, and then we wanna figure out how to optimise from there. And so for those of you that work in an organisation that uses Salesforce or maybe you're a partner and you're you're an implementer and you think about your clients, this is a traditional arc that many, many organisations go through kinda regardless of industry or their use case. And so what we've been able to do over the you know, over years and and what you guys have done was establish, centralize those processes, establish them in a platform, find ways of improving your data quality, for example, here in using Marketing Cloud to kind of nurture those bigs to get more profiles populated more quickly to reduce that timeline. And again, just to just to kind of make that timeline tangible, the numbers we were kind of working from here was to say, hey, I as a as a BIG mentor coming into Big Brothers Big Sisters of America, if I if I started, you know, at the beginning of this timeline, it might take up to two years to place me, with a little. And just by kind of going through that same process that so many organisations have had, by the time we got to the point where we're now talking about AI, a lot of those enhancements have kind of come into the door, but you still had that two month process where you were seeing your match specialist kind of struggle there. And, you know, one of the things I know we talked about earlier on, Travis, was kind of going through. Even if we can filter results, there's still a lot of profiles that people have to sort through there. So, you know, what was the, you know, kind of common complaints that you got from your match specialists and coordinators around why it was still taking up to two months? Was it analysis paralysis? Was it kind of a lack of data quality? You know, can we dig into that a little bit? Yeah. I mean, there's a few things. Right? Analysis paralysis for sure is one of them because you think about just going through dozens of different applications and interview notes and and parsing through all that data to find the the matching criteria, the common matching criteria that we know leads to the longest, strongest, last lasting matches. You know, things like likes and dislikes and hobbies and career aspirations. And when you have dozens of bigs and dozens of littles trying to find all that information and remember that information and then make matches. And then there's also the the the human constraints of, you know, our our matching specialist care, and they really just wanna make sure that they're making the best decision with those matches. Right? And so there's, you know, a little bit of that anxiety. Is this the right match? And what if it goes wrong? And things like that. And so, you know, we're trying to reduce that mental load, and instead of replacing judgment, kind of create that trust that follows fast. It's the classic thing. Right? Hey. I'm looking at a list of people and the next person on the next page might be a better fit for this little, so I better sure go in and look at that profile and read through everybody. Right? You know, so just as if you or I are searching for buying a car or looking at Google results or things like that, there's only so much information that we can process at one time, but we're also not sure where the tools in front of us are we actually getting the best results at the top of the list as opposed to just kind of a narrow down set of results that we still have to fully parse through. That's exactly right. Yep. And, you know, when think about part of the challenge here, you know, there's a lot of organizations that I've worked with where when they have this type of challenge, they end up creating very long or complex surveys and forms to fill out initially, sometimes with hundreds of fields to capture information in all sort of a structured way so then we could build that matching process based on the tools and capabilities that we had available to us. Can you just talk a little bit about the high level? Like, if I'm coming in as a big, what does that process kinda generally look like on my end? What what am I having to fill out versus what's kind of more of that unstructured conversational information that your team is learning at that time? Yeah. So when you initially raise your hand to become a big, you know, you have to fill out applications. There's usually an orientation interview to let you know what to expect, and then we have a rigorous process to ensure youth safety and protection as well. So, you know, there's background checks, fingerprints, things like that that have to get done. So you can imagine as a specialist, lots of follow ups and and tracking tasks and trying to make sure that the big is in constant communication. And then, you know, over those interview processes is where you get some more of that, you know, deeper dive information where you're you're not you're asking, you know, what are your preferences for a match and things like that. And so just collecting all that information and then consolidating it and then parsing through it. You know? So from the big's perspective, you know, there's there's the paperwork aspect from the matching specialist. There's the follow ups and task tracking, things like that. We really just wanted to reduce that load for everyone involved. Absolutely. And and not only are you doing that with the bigs, but you're having to go through, I'm sure, a similar process or maybe even a more pretty big rigorous process with the littles as well. Right? Yeah. Their parents and guardians, especially, of course, the littles. Yeah. So all the constituents involved kinda have to go through this kind of onboarding process where they're initially their initial inquiry, and then we're collecting all that information. Then they get to that stage. We have all the information. They're ready to be matched, and now we're trying to look through and parse that data and figure out who to match to make sure we get the highest outcome for the young person. And there's obviously when we think about that, there's both the the subjective. Obviously, we wanna know someone who they are as a person on both sides and and and try to match based on personality and preferences, like you said, and traits and all those other things. But then there's also some hard and fast rules that we have to stick to that we can't just necessarily just, you know, let the let AI and how it's kind of probabilistic just kinda hope it's always gonna follow those rules every single time. Right? Like, you have some hard business logic that you in filtering that you need to be able to also align on because there may be some, you know, preferences, let's say, the parents around the type of person that they want, you know, engaging with their kid. Right? Yeah. I mean, that's exactly right. And, obviously, front of mind, number one priority for us is youth safety and well-being. And so those those rules are nonnegotiables for us. Right? And these are the things that that must be in place and the data that we must be able to verify before we can move you to the next stage of the process. When you're had you explored obviously, we we built this on the Salesforce platform using AgentForce and the components that AgentForce brings to the table. As as much as, you know, you wanna share enough, are there areas where you had explored or talked about doing this with AI previously, or was this more of a I kinda you thought just the the next natural evolution because you've already built on on Salesforce? Is this kinda your first foray into AI for the matching process or something that you had tried in other areas and and maybe not had success with? No. We didn't try, but we definitely evaluated other areas. And and, you know, for me, it was almost immediate that we need to kind of keep this within Salesforce. We don't want our people leaving Salesforce, right, and having. another tool that they have to deal with. We don't want our data leaving Salesforce. Right? If you're if you're moving your data into other, you know, warehouses and stores of information, That's just more from a cybersecurity perspective to have to protect. And so reducing cybersecurity, reducing tooling, and the load on people was really front of mind for us. So, you know, that's where it became obvious to me, like, we wanna embed this within their same workflows, within the same screens they're already using. It helps us protect information. It helps us, you know, with economies of scale, reducing cost, and things like that. So And it helps with change management too. Right? If your teams are already working in the Salesforce interface in that ecosystem, they already have a level of familiarity there, then we're not asking them to kind of swivel chair into another tool and have to learn that whole process and then kinda move back into the Salesforce platform where they're otherwise working. Right? I'm sure change management was probably something really front of mind around adoption and actually gaining the ROI and the efficiencies that that this capability should be bringing in. Yeah. Absolute I mean, yeah, our specialists, like I said, didn't need another tool. They just needed better signal inside the screen they already trust. And so, you know, embedding AI inside Salesforce made adoption, like, very easy as it relates to you know, in general, adoption can be quite difficult for any sort of change, but especially technology. So hearing these requirements that we were getting from Travis and team, Epic Brothers Big Sisters of America, that really told us a few things as the implementation partner. One was that, you know, they've already invested and really done a lot of work in standardising a process within their organisation, you know, as the nonprofit is the entity itself, but then also building that out within the Salesforce org. And one of the things kind of hearing all of these requirements, we wanted to make sure that, one, whatever we built wasn't necessarily going to, again, cause that friction of adoption where people had to learn kind of a new click path or an entirely new way of engaging with the interface in order to gain the benefits of this AI matching. It also told us that more likely than not, you know, we wanted to give people the option of potentially pivoting into maybe some of the existing tooling that they already had so that as especially as we're doing a pilot, we're not necessarily doing as much of a a a breaking change to a process. If for some reason the match tool and our initial iterations wasn't performing the way that we wanted it to, we wanted people who were working in the pilot to be able to go back to their existing technology very, very easily. So we're not necessarily just doing kind of a hard cut over. And on top of that, we wanted to make sure that whatever we built was going to be maintainable by Big Brothers Big Sisters of America and their team that was working in Salesforce as well. So we didn't want to introduce a ton of external dependencies also for the cybersecurity concerns, but just making sure, like, look, we wanna build on Agentforce. We wanna leverage the investments that Salesforce has made. So how do we make this kind of as out of the box as possible as low friction? But really all that we're doing is just adding AI into the process so that their team can benefit from that? So we knew that there were gonna be a few capabilities that we wanted to be able to launch within what we were calling SmartMatch. One is, obviously, we wanted to have AI driven matching. And so the idea is taking both structured data and unstructured data that we have. And by the way, yes, we will show a little bit of a demo of this here in a second. But we wanna take all that data, and we want to allow AI to lever you know, use the tools that we have available within agent force and and Salesforce ecosystem of products to be able to find matches both on that subjective preference trait all of that information that we have, but also adhering to the red lines, the hard requirements, the things that we absolutely must not violate in terms of rules both for Big Brothers and Big Sisters, but also for parents of these kids that are going be having these mentors come into their lives. On top of that, we knew that we wanted to be able to display recommendations in such a way where the matching specialists would be able to understand why those recommendations were presented in front of them, so that they could still be able to see kind of the end result of that data, how things line up between the bigs and the littles, where they do, maybe where they don't, so that it's not a black box. We want it to be visible. We want it to be reportable. We want a level of transparency so that while we are using AI to make the process easier, the users internally are not feeling like they are kind of getting cut out of the process or not understanding why certain matches were getting made. So we knew that we wanted to be able to give summaries of that information and we wanted to build it in a user interface that was gonna require as little change management as possible. Any comment, Travis, on any of these features before we jump into the demo here and I kinda show people, you know, what we built? Yeah. I think what you were just talking about is is just so important, Sam. Like like, trust trusting these tools is is so important, especially for such a human centered business like ours. And and trust came from three things. You know, you mentioned transparency, why a match was suggested. They have to know the reasoning for why this was a recommendation. Number two was control. So the specialist always makes the final decision. They they look at the recommendations. It really gives them a prioritized list instead of manually parsing through all those things themselves and searching. And they, you know, using their expertise and their human judgment, make the final decision. And then last is consistency. Number three, consistency. You know, making sure that those results, those recommendation results are aligned with their lived experiences. And so from a change management perspective, what we saw some of the pilot agencies doing where they would have regular weekly match meetings, and their matching specialists would come to these meetings with recommendations that they had from the the current process that they used. They would, in the meeting, run the AI smart match recommendation engine, and then they would start to see alignment between their recommendations and what the AI was doing. And the AI was doing it, obviously, faster, but still aligning with their own experiences and their recommendations. And so we really positioned AI as, like, second a brain, not a decision maker. And once the specialist saw that it reduced the mental load instead of replacing judgment, like, trust followed a lot faster from that. You know, you you you hit on a really good point there. And another way that I tend to phrase it when talking with a lot of people who are Salesforce clients or prospective clients of ours is that Salesforce has done a great job. You know, you'll hear Salesforce talk about the trust layer behind agent force and their AI solutions. And that trust layer, paraphrasing here a little bit, but a lot of what that is oriented around is making sure that the standard good practises around implementing technically, you know, generative AI within an order follow. Right? So it's not going you know we're trying to reduce hallucinations just generally in how it operates. We have checks and there's a checks and balances process to make sure that there isn't any sort of toxic results or, you know, profanity or things like that that you don't want kinda coming back from your AI. And then also just making sure that they kinda take on that responsibility around that security, that data security layer to make sure that, hey, whatever models you're using on the back end, we either can apply masking on that to protect your data or, you know, it's within Salesforce's trust boundary. But to me, that trust of saying we can trust what's going on with your data and that our AI is safe and secure is not the same as building confidence. And I think that's the challenge that we're presented as implementers and you guys as as customers of Salesforce are presented with, which is to say, okay. I've got this great platform. I'm gonna go build an AI agent or some sort of embedded AI, generative AI solution on there. How do I build confidence in what I have built and configured in the instructions that I'm giving to that AI because it's nondeterministic? Because sometimes I can put in something very similar and get a little bit of a different answer. So how do we, you know, tune that enough to be able to really meet your team's needs? And and a couple of things that we learned as part of this process. One, I think there was a two week period where we were able to rearchitect the solution about three times based on new learnings and understandings around exactly what your team members were expecting for the process and the experience and and kinda how we apply the searching and filtering. So one, was we're a bit it'll be very iterative in that approach. And then I think to to a certain degree, Travis, what you're talking about is also we had this, like what we eventually started calling a test and tune cycle. You know, with a lot of traditional Salesforce and IT implementation projects, you kinda have your design. You go you do your build, and when we go, we go build somewhere. Maybe we might be doing agile, might be doing waterfall, but you kinda have that build phase. And then you have that UAT phase. Right? So it's okay. The build's done. We're gonna get our users in there, test it, make sure everything meets the specifications, and aside from any fixes, we're gonna go deploy. I feel like with this process, it was much more of an iterative, hey. We're gonna build a framework for something. We're gonna get some of your pilot, you know, subject matter experts in there kinda testing it, and we're gonna immediately take that feedback and kind of iterate whether we're talking about the prompting or the architecture or things like that. So I'm I'm curious just to get, you know, a little bit of feedback from you, Travis, in terms of, you know, the evolution and that process of working together to build that solution. It really sounds like, you know, that's what we needed to do to make this work. Yeah. I think I mean, it's always been my philosophy too. Like, driving adoption is hard enough. And if. you just drop push it on people, you know, you're gonna lose all that confidence and trust. And so having our agencies involved from the very beginning toward from initial requirements gathering to the design phase and then ultimately testing it, you know, and not just not just having them involved, but when they do provide the feedback, they see that you actually apply that feedback to the solution. All these things help build confidence and trust, and and they feel like they were a part of it because they were. Right? And if if somebody's a part of building a solution, they're much more likely to actually use it and trust it and enjoy it. And and, you know, that's the other reason why we went with, you you know, know, you were talking about hallucinations and things like that. That's that's really why we decided to embed AI directly into their workflow instead of, like, you know, building a chatbot or an agent. Mhmm. People asking different questions, get different responses. You're trusting that black box response. And so, you know, we we learned from them, gathered the requirements, and then we did all the prompt engineering. And we have all that kinda running in the background, and and we're we're selecting the data, you know, based off of what we know is the best matching criteria. And you get more consistent results that way, and and the more consistency that's aligned with their lived experience, the more they're just gonna trust it and use it. And, you know, I'm I'm looking forward to talking about that, I know, in in later slides on some of the some of the impact that we're seeing from from the consistency of results. Absolutely. So let's pivot over into our sandbox real quick. Obviously, we're not gonna show a production order because we wanna make sure we're being very cognizant of the data that we're showing here. But I'll just kinda give a little bit of an overview and and, Travis, while I'm not seeing you on the screen here, feel free to chime in as I'm kind of doing a quick little walkthrough here. Every big or little that comes into the org goes through a whole application enrolment process. And I'm not going to go through every field or every tab that we have here on the screen, But one of things just to really call out is that there's a lot of data that gets collected, right? There's a lot of contact level data that we can see here, whether we're talking, for Big around where they work, who they are, you know, there's a background there's background checks that are collected and we get information from that, even personal references that they have to go through. I mean, this is even more serious than when I'm when I've applied for jobs in the past, let's say, right, like this this comes down to really Big Brothers Big Sisters taking all that very seriously. And, you know, when we then get to a point where we are talking about actual traits and and kind of assessments in an interview process. And, Travis, I'm not sure if you wanna kinda chime in on some of the the critical information that we're collecting, you know, when we start getting into these stages. Yeah. I mean, we're we're collecting, you know, likes. Dislikes are just as important as the likes are. We found over the years through this matching process, You know, age ranges, you know, variability in age ranges have an impact on matching gender, sexual orientation, preferences. You know, all all this information kinda goes into the final matching decision. And as you can imagine, as you can kinda see here, it's it's a lot of information for a human to have to kind of sift through and and, you know, make sure there's there's that alignment. And, again, that creates that human anxiety of is this the right person, what if, you know, and those types of questions start to come up. Yeah. And so and, again, we can see here, right, some of the most critical information around the quality of the match, let's say, is very unstructured. I mean, it's interviews you guys are doing. There's assessments that your team is writing up and kinda giving summaries here, even giving kinda match recommendations and feedback, you know, not even just what the bigger the little is sharing with you directly, but, you know, you're even kind of getting multiple layers of that information and analysis. And so I don't wanna call it the legacy, but let's say the prior way of finding matches, in this case, finding a little from a Biggs record. And for anybody who's been in Salesforce for a while, I'm sure a screen like this is probably very familiar to them, where we were taking a lot of structured information, like anything that could be structured so that we can add filters for, you know, different sites, different programme types, what genders we're looking for, that sort of thing. And in running a search, that would give us a related list at the bottom here of all those potential matches. And, you know, when you think about this type of a solution, what it's really doing is filtering out maybe what we think doesn't work. But it's not prioritising what does work. A solution like this is not necessarily going to rank at the very top who your best match is. It's basically saying, I've got this pool of ten, fifty, a 100 people, maybe even more, who could be matches for, in this case, littles for this big, but now someone someone has to still click through every single one of those to kind of go through all that assessment information. So that's what we were really trying to change. So when we go to the SmartMatch tab, there's a couple of things that we did here. And when we think about, you know, as we kind of kicked off this call talking about frameworks and and and kind of how we approach problems, one of the things that we did was we took kind of two categories of information, one being characteristics about a person and the other being preferences. And, you know, as you saw in those prior tabs, there's a lot of data that's locked up in assessments and all that info other information. We actually use Salesforce prompt templates under the hood and flows to build to rebuild these whenever we need to or maybe when that first person first comes into this particular stage to give us a more distilled summary of kind of each of these categories of information. And so now what we're doing is we're taking what was maybe a combination of unstructured and structured data and putting that all into these just these are just long text fields, folks. Right? There's nothing special behind this aside from the prompt outputting that info prompt and flows outputting that information. The reason why we do this is then we take these and we we use the capability that Salesforce is called search index within Data three sixty, formerly Data Cloud, where we are able to take these usually, will go in as, like, one or two chunks. That's part of the reason why we made them smaller, so there's as little chunking as possible in breaking apart that information. And what we can then do at a follow-up task that you'll see here in a second is we actually perform a similarity search on these chunks of information. So we've kind of flipped the script on its head. Instead of saying, hey, I need to have information structured so I can filter on it, we're actually turning this back into unstructured data so that AI can do what it does really well and do that semantic similarity search between these sets of characteristics and preferences that we have here about a person. Now, when we think about how this executes, again, what you're not seeing here on the page and, again, nothing against us whatsoever. But in terms of what we needed to accomplish here, what we're not seeing on the page is actually a chat interface for for SmartMatch. And this goes back to, you know, the adoption priority that that Travis has been bringing up and that we've been keeping in the back of our minds, which is, hey. These users already have a level of familiarity working in this interface, working in a flow and lightning web component kind of architecture. And we also know, frankly, the more questions we have to type sometimes, that's like one of the slowest ways that humans communicate. We know clicks are some of the fastest. So how do we keep this kind of this click oriented as possible? And I'll just show what this process looks like. So we can see here, we've got an embedded flow on the page. So I'll say, okay, we're gonna go search for some new matches. One of the things that's very critical with this process is distance. Right? I can have really good matches, but if I'm in Florida and the little I'm getting matched with is in California, that's probably not gonna work. Right? Like, we wanna have the ability to have in person real human to human connection here with these matches that we make. So one of the things is that based on guidelines that we got from Big Brothers Big Sisters of America, we were actually able to kinda have both a raw distance filter, but even hooking into, I believe, Google Maps APIs, we actually were able to calculate drive time distances for each of the matches that we initially return so that we can filter those down further. So in this example, I'll use drive time. Now, again, we think about some of those hard rules and requirements on what we want to filter on, we try to narrow those down a little bit and we have about five or six different inputs that we can have here for what we wanna have in terms of those hard and fast requirements. And again, first at the top, we have the slider so I can see what distance we want. I'm just gonna put it pretty high because of the limited sample data that we have in the sandbox here. But then we can see down below, we can say, hey, do we need to enforce, you know, gender rules on the search or sexual orientation or specific locations of agencies as opposed to just looking at, you know, let's say, this drive time heuristic. And so we can set these if we want. I'm just gonna go ahead and leave them blank right now, and then we go ahead and run next. Now this is the part where I have to sometimes damp from about thirty seconds or so. What's happening behind the scenes is this is all within the aside from the Google Maps API here, everything else that's happening is within the Salesforce platform. What we're running is a combination of flows and Apex. Well, hopefully, we'll see the results refresh here in a second, but they may start to be the same. Click finish. Yeah. Pretty much the same results here. And what's happened when that search runs is it's running a combination of prompts and queries within Data three sixty to the search index to find similar results. So it takes both some of the hard filters, you know, structured filters that we have here and then takes this information that we have up at the top is actually a prompt query where then we are using that to find the most similar littles that based on the overall meaning of each of these fields that we have here. So now what we are getting is instead of potentially dozens or hundreds of results, we can you know, this can this is toggleable. We can do the top three or top 10 or whatever kind of hard limit that we wanna have on there, but we're actually even able to rank these in terms of closeness of similarity. We get some data that comes back in from the search index that gives us kind of a numeric score. We don't display that to the users. We just use it behind the scenes. But then we're actually able to have AI even give us a summary of why this person is a good match. And so we can see right here down below, you know, it's not just that black box of generating these matches, but the AI is giving us kind of that reasoning so that the match specialist or coordinator can feel very confident as to why this match was linked up in the first place. And on top of that, we still retain a lot of the structured information and we're actually saving these results so we can report on them and monitor the quality of those matches over time. On top of that, we still are keeping the human in the loop. If I want to go in and read the profile of this little directly, and it's probably something that I should be doing with any of these results that come in, I can still do that. But now, hopefully, what we've done is maybe we have these coordinators that are looking at maybe the the top two or three results rather than having to go through page and page and page of results and trying to find that best person that maybe is, you know, in the next set of 10 results that we show on the page there. Any commentary you wanna add on this, Travis? Yeah. I would add that, you know, you think about the variability across our country. We have all these micro cultures in different regions, and so that's why these these filters here that you're seeing were so important. Right? Religion matters more in some parts of the country than others or even things like distance. So, you know, five miles in Los Angeles is a lot different than five miles in rural Ohio. And so that's where we implemented drive time versus just distance on the old matching tool that we have that that Sam was showing earlier. But we also integrated not just drive time, but public transit time, cycle time. Right? Because if you're in a city, you may be walking. Right? So walking distance is a little bit different. So it just depends on where you are in the country and and what matters to you. And so instead of trying to build a one size fits all solution, we put some of the customization capabilities in in the hands of our agencies directly, which is what Sam is showing here. Fantastic. Let's pivot back to our slides here real quick. And so we hit on some of this already in terms of, you know, how the engine functions today. You know, what Salesforce often calls this kind of a hybrid search capability, I would they both have a feature for this, but there's also just kind of a general design pattern or architecture around that that we are following. So one is, you know, we wanna pre filter results based on some of those things like drive time or like a religion or sexual orientation, things like that before we actually are starting to rank who's, you know, who's a better or worse candidate. Right? So we wanna say, okay. Once we've got kind of our our hard criteria filtered out, then show me, you know, the best matches within that set that are eligible. Right now, like we said, we're we're kind of returning the top 20 matches. And then the the travel preferences are kind of evaluated after that, whether, you know, whether we're talking drive time or or public transit, things like that. And then, obviously, there are certain, you know, location matches and things like that that we also need to adhere to. So we're kinda for everybody who's maybe more involved from a developer or an admin or architecture perspective on this call, you know, a lot of what we're talking about are these programmatic deterministic controls, right, over the process. And then we are layering in AI specifically, you know, at that appropriate time to kinda let AI do what it really does best rather than say, I'm going to take all this information, dump it into AI, give it those rules, and hopefully cross my fingers it's going to follow those rules every single time. From a postal perspective, we really talk about that being the difference between a fully agentic solution and an embedded solution. And, you know, just for a little bit of kind of terminology here, the, you know, the general industry terminology that we tend to see. Right? Salesforce kinda has theirs their own and every every other platform does as well. Simon Wilson is a great thought leader in this space, kind of defined it as, you know, an LLM agent runs tools in a loop to achieve a goal. Right? So agentic AI is where we're taking more of that process, we're handing it off to the agent, not just for, you know, the natural language processing that the AI does best, but also deciding what to do and when. That can be extremely helpful and beneficial in scenarios like, for example, a customer service support agent. I need to be able to let that agent handle some of my entry level tier one tasks, decide what information to go look up where, what action to take, who to route someone to. But in situations like this, where we have a well defined process that already exists that's relatively linear, right, but we know that we just have a couple of specific rough patches or points of friction within that process, That's where more of this embedded AI approach can be very attractive. And that's where we're taking AI, and we're simply leveraging that in very specific steps in a process that's otherwise well defined and is human driven. Now we could actually that that embedded AI could be an actual agent that runs in there sometimes that maybe is deciding subtasks to go do and then returns that to the user. We're not doing that in this case. In this case, we're using a series of prompt templates that execute at different points within the process, whether it's doing that initial distillation of someone's profile or whether we are, you know, doing that final step of the evaluation of those matches. And so this isn't to say one is better than the other, but what we are saying here is it's a new way of designing solutions in AgentForce that I personally believe has been somewhat overlooked, but where we are actually seeing a lot of early adoption and success you know, by most of our clients, but especially by Big Brothers Big Sisters. And so on that point, Travis, I'm wondering if you've got any metrics that you wanna share based on your experience so far as we've been live for, what, fully live about ninety days now, I think. Yeah. A little more than ninety days. You know, I mean, we we we've been out a little bit longer than that, but like you were saying, we went through that iterative process. We learned things, and we continue to improve the solution. And so now we've got this scale ready solution out for the last three months. You know, I I think the biggest unlock so far wasn't just accuracy, but just confidence and speed. And we have certs, you know, several early match outcome metrics that are kinda proving the concept of this solution. So, you know, one of them is early match longevity improvement. So what we're seeing is matches created without it are already strong at 95%, but we've seen, like, a 4% lift in match success on an already high baseline. And that's in terms of, you know, the the the length of the match. Of course, you know, three months in, our average match length is is twenty months or more. So we still have some time to go, but we're still seeing almost ninety nine percent success rate now in terms of matches lasting longer with the smart match recommendation engine. There's also this power user signal that we're seeing. So, like, our most experienced specialists, which are the hardest audience to win over, they used smart match repeatedly, and they saw meaningfully longer lasting matches in the early stage of the match, which helps eliminate that early adopter bias argument. Beyond that, you know, SmartMatch is is showing strongest where relationship fit matters most. So we conduct these strength of relationship surveys at the three month match anniversary. And across our core relationship quality indicators, SmartMatch assisted matches are consistently scoring higher than non SmartMatch matches. So some of the early signal themes are a higher agreement with, you know, we are well matched, a stronger sense of closeness, greater overlap in those shared interests that we're talking about. So at just three months in, SmartMatch pairs are already reporting stronger perceived fit and connection, which is the exact things that typically take much longer to establish, and they're beyond surface level outcomes. These are foundational predictors of match durability. Smart match also appears to reduce early relational friction. So so while the average scores are modestly higher, which is expected this early, what's more interesting is what's not showing up as strongly in Smart match matches. So we're seeing lower signals of awkwardness or lack of things to talk about, emotional distance, and early disengagement patterns. So so, again, it's not about making relationships perfect, but it's just reducing the early friction when relationships are most fragile. And that matters because most match failures happen early, not later. And then also just AI impact showing up in human perception, Mhmm. not just a system metrics. So so what's standing out is that the SmartMatch benefits show up in how people feel, not just the operational stats. So some of these survey results, people are feeling well matched early. They're feeling close. They're feeling aligned in interest. And that's important because humans don't persist in relationships because of algorithms. They they persist because it feels right. And so if our AI recommendations were getting it wrong, we'd see that skepticism in these surveys first. And so, you know, the lift is small, but that's actually the point. We're not seeing dramatic swings, and and that's a good thing because Big Brothers Big Sisters already has a strong baseline. Our match specialists are highly skilled, the bar is already high. Yet, Smart Match still shows directionally consistent improvement. And so when you're already good at something human driven, even small gains can pound quickly. And, you know, what's encouraging is that these early relationship signals we're seeing line up with the long term match success data. What what I love to to hear with that, Travis, is that all of that was focused on end outcomes, like the the true true so what at the end of the day and your mission as an organization. You know, we started talking about, well, the length of time that it takes to find a match and that sort of thing. But, obviously, where you where we've been focusing on first is like, Are we making sure we're not, you know, messing with the process in a in a negative way and actually having kind of positive end feedback, which it sounds like we absolutely are. Do you have even whether it's anecdotal or actual collected evidence around around that time savings or, you know, where your team members are spending their time now? And granted, it's definitely still early in this process. Yeah. So, I mean, we we talked earlier about some of those weekly match meetings that agencies were doing. And so even even in the pilot phase, forty five days in, we did a feedback session with them. And these agencies, you know, especially our large agencies that have a lot of matching specialists and a lot of different matches to look through, they would meet on on a weekly basis for at least an hour. Right? And you can imagine ten, twenty people in a meeting like that. That's time, right, that that. the organ spending. Even forty five days in, those match meetings were reduced to thirty minutes. And now we've even heard from the same agency that they've eliminated those match meetings altogether. Right? And so that's sixty minutes times 10 to 20 people a week that goes back to the organization where they are now able to focus more on high value human interactions versus, like, those meetings. I'm never gonna complain about having a meeting getting canceled off the calendar. Like, that is for sure. That's one way to drive adoption of technology, reducing meetings. So on that note, we we're just about ten minutes left. Let's go and jump into q and a from the audience here. So, Simon, what do we have? Oh, well, we did have one question come in with regards to SmartMatch summaries. Were you building agent force agents, tools, or actions from scratch, or were there items already available that you could work into your overall solution? I'll tackle I'll tackle that from a design perspective. We were generally building, building these from scratch because what we're leveraging is kind of the classic Salesforce as a platform functionality. Right? This is something, you know, doesn't exactly apply to to Big Brothers Big Sisters America here. I'll say this from a coastal perspective. A lot of the out of the box agents, tools, and capabilities that Salesforce brings along with AgentForce are fantastic for getting an understanding of what the platform can do, but not necessarily gonna be broadly applicable to everybody. We generally see those at least require customization, if not essentially building custom versions of those. But the good news is this is Salesforce. If you know how to build flows, if you maybe have to write at that Apex and you have, you know, developers on hand that can do that, it's not requiring this whole additional lift to do so. But going back to the specific stuff that we used in this build today, it was the search index capability within Data three sixty for so we we would once we generated those summaries of records, we synced them down into data three sixty through the the regular Salesforce CRM integration, created a search index from those, and then it was just a combination of mainly flows and prompt templates, a little bit of Apex sprinkled in to build the rest of the solution. So it it didn't really require too many of the traditional kind of agent force tools, but more of those components that we use that kind of fall under that agent force umbrella. Excellent. Thank you. Another question was, what is the most common reason of missing the match? Travis? Yeah. Can I kinda need a little bit more clarity on that question in terms of missing the match? I think, you know, the most common reason for match closure in our organization, so when a match actually happens and then those matches end up closing is people moving. Right? People move and they're too far away from each other. In terms of missing the match, it it's really been sifting through all that manual data, and that's where some of the time delays come in, the human anxiety. You know, like, did I look at everything? Am I missing a criteria that's gonna be critical to match longevity? Things like that could could lead to a mismatch. So I hope that answers your question, Andy. Excellent. Thank you. And, yeah, Andy, if you would like to provide some more explanation around the question, we can always address it again. Another one here. Could you please highlight matching process balance, the efficiency of AI driven tools like agent force with the need for human intervention when a mentor mentee relationship faces complex emotional or behavioral challenges that technology might not predict? Yeah. So we have not integrated AI into the match support process. So once a match is actually made, we still have our specialists supporting those matches. That's all human human driven right now. We do have a vision for ways that we can help optimize that for our matching specialists as well, but we're never gonna have tools like agent force supersede human intervention for those things. So so we have monthly match support calls that happen with parents, with bigs, with littles. We're using the data that we have in the system to help support those matches and support, you know, whether it's the parent, the little, the big through some of these, like, you know, emotional or behavioral or complex human challenges that you're talking about as those surface either in our data or during those monthly match support calls. So really what this solution has done is by opening up the time spent for a matching specialist, looking through data, trying to find the next match, it's really freed them up. That combined with the automations we were talking about earlier with getting people through the process, it's really freed them up to have more of those high value interactions and those calls and and working with people on the support side of the match versus some of the administrative side of getting people through the enrollment process and getting people matched. I'll I'll I'll piggyback on that, Travis, with with a follow-up question there, which is, are there areas or or where do you see next where you will kinda go deeper into the AI agent force side of things? Is it gonna be kind of continued in hand like, monitoring and enhancement of the SmartMatch process? And just, you know, you know, kinda just sharing here with the community, are there other areas where you are considering, you know, potentially where AgentForce or an AI solution may kinda help benefit you guys scaling up? Yeah. I see moving us further down into the funnel now. So we've talked about automating some of that enrollment process. Then once they're ready to be matched, making recommendations to help prioritize match selection process, The very next part of the funnel is supporting the match for longevity. And so I see a world where, you know, we can use this data that we're collecting in these surveys that happen throughout the match life cycle in the match support calls, also using some automations to collect the information so people can just respond to a text with, you know, any challenges or struggles that they might be having. And then we have a robust library of youth mental health and wellness sort of artifacts, knowledge artifacts that we can then use AI to kind of make recommendations for activities, conversations, different things like that based off of this library that we have on youth wellness, but also just elevating those to our match support specialists so that they can better support the matches. And then that elevates the monthly match support call that I was talking about instead. of, you know, the the call being surfaced around, you know, did you meet with your little last month, and what did you do, and was there any issues? Some of that can be collected automatically, you know, through text automations, email automations, things like that. Then you go into that monthly match support call with that information already armed combined with our library of youth wellness education, and we can help the bigs and littles and parents through those challenges. So that's where I see kind of the next evolution of AI for us and technology in general. Excellent. Thank you. I've got an interesting question here. If a workflow requires a judgment call that the AI shouldn't make alone, how do you design the handoff between the AI's recommendation and the human's final decision? I think it's something that we've kinda seen in the solution and how we built it so far, right, which is, again, we're not leveraging AI here as the kind of final decision maker. It's making we position it as recommendations, and then we try to build an interface to make it as transparent, but also as simple as possible for the human on the other end to review those, you know, one of three or up to 10 or 20 recommendations and then kind of narrow down what the person needs to do on that side of things. We talk a lot about this at Coastal, believe in our usage of AI tools. At the end of the day, a person is responsible for the decision that's made on there and it's their name on that record, on that document, on whatever asset is being generated there. But, again, it's that's where we come down to, hey. That agent tech, which is where the AI is making and executing a decision versus more of that embedded functionality where we're leveraging it either for its strength or to make recommendations and we're teaming that up for the human. That's where we kinda make that decision early on when we're doing our design with our clients to say, you know, which of these really fits the solution that you wanna have? Excellent. Thank you. And I think we have time for just one more question before we wrap up today's session. How does Data three sixty play into this architecture? At what point does a simple embedded solution need to scale into a multi cloud data strategy? You know, I think Travis was hitting on a little bit here too, but there's a couple of things. Very common question that we get is, I'm doing something with AI and agent force. Do I need data three sixty? Formerly data cloud again for everybody here. The short answer is yes. It needs to be turned on, but it does not require that you actually go kind of for the full licensing component with your account executive. Again, everybody's situation is different, don't take that as gold, but everybody who's enterprise licensed it up with Salesforce has some Data three sixty credits available. So we are using that search index capability that I talked about, which essentially is Salesforce's implementation of a vector database for those that are familiar with the technology, that lives within data three sixty. So we are doing a little bit of processing and storing some data there, and then we query that every time we wanna make a match. Aside from that, yeah, as Travis talks about future plans potentially with doing more with AI and bringing that data in from other sources, we will likely work with them on understanding where Data three sixty will come into play for ingesting that data to be able to surface it up to AgentFlorence. On that note, I know we're coming up on time here. So I just wanted to call out, obviously, as Coastal, we're doing this. We love the work that we've been doing with Big Brothers Big Sisters of America. We wanna be able to put in similar solutions and solutions for AgentForce and all things CRM for an organisation like yours if you're in the market. By all means, check out these QR codes or you can go to our website at coastal.us. You can sign up for either a complimentary Salesforce org scan, which will give us kind of a whole understanding of the lay of the land of technical debt in your Salesforce org in a very easy to understand dashboard. We're also doing data strategy labs complimentary or even, you know, more in-depth engagements where we will get in, understand data quality that you have, which we all know is essential to anything we wanna be able to do with AI. That takes us to 01:00. Travis, thank you so much for for joining us today and sharing your experience. And and thank you to to Simon and the Salesforce Ben team for hosting us here as well. Thanks for having me. And if you wanna raise your hand to be a big, go to bebignow.org. Excellent. Amazing. Thank you everyone for joining, and the session will be shared the session recording will be shared in our follow-up email if you'd like to revisit any of the points discussed today. Otherwise, I hope you all have a fantastic rest of the day. Thanks, everyone. Thank you.