Name: Strategies for Scaling Agentforce AI Systems
Uploaded: 2026-04-07T17:32:36.676Z
Duration: 58 min 36 s
Description: Strategies for Scaling Agentforce AI Systems

Transcript for "Strategies for Scaling Agentforce AI Systems": Hi, everyone, and welcome. Thank you so much for joining us today for strategies for scaling agent force AI systems brought to you by Salesforce Ben and the Absolute. We're really excited to have you with us. We know there's a lot of interest right now around AI and agent force, especially as teams move from experimenting to actually scaling these systems in production. So today's conversation is all about what it looks like in practice. Now before we get started, just a couple of quick housekeeping notes. This session is being recorded, and we'll be sharing the recording with you after the event. We'll also have time for q and a at the end, so please feel free to drop any questions you have into the chat, and we'll keep an eye on that throughout this session. Now, John, if I could have the next slide, please. Just a little bit about Salesforce, Ben, where we're building towards a future where there's a clear path to success for anyone looking to start or grow a career in the Salesforce ecosystem. Now as part of that, we focus on delivering practical, relevant, and accurate content across a range of formats. So you'll find our articles, newsletters, YouTube, podcasts, events like this one. There really is something for everyone. So if you haven't explored that yet, perhaps feel free to take a look after this session. But bringing it back to today's topic, scaling AI systems, is something that we've seen come up more and more across the ecosystem, and that's why we're so excited about this conversation. Now if we move on to our next slide, I'm joined today by some fantastic speakers who bring a real mix of industry perspective and hands experience when it comes to AI and Salesforce delivery. I'll let each of them introduce themselves in a moment, but together, they'll help us unpack not just what's changing, what it actually takes to scale AI systems effectively and responsibly. So with that, let's get started. Zane, if you'd like to bring yourself onto the stage, and I'll let you introduce yourself. Sure. Hi, everyone. My name is Zane. I am the director of AI and business outcomes consulting here at dAbsolute. Over the last decade, I have helped companies in various industries leverage automation, predictive insights, and developing operational strategies to unlock new levels of efficiency and user experiences. And now I do the same thing, but with agent force into the mix. So helping companies strategize, implement, and release AI technology into their operations to achieve real business impact. And with that, I'm gonna pass it on to my good friend, Shannon. Same. Hey, everybody. I'm Shannon Jensen. I'm a success architect at Salesforce, and I specialize in agent force. I'm from Austin, Texas. I've been at Salesforce for about four years, and then before that, I worked in digital advertising setting up data pipelines for ad tracking. It's a pleasure to be here to discuss AgentForce today. And next, I'll pass it over to John to introduce himself. Thanks, Shannon. Good morning. Good evening. Good afternoon, everybody. My name is John Pettafore. I'm fortunate enough to spend the last fifteen years in the Salesforce ecosystem. I was a customer for a few years when I joined a partner. I ended up then moving to Salesforce for the past four years, where for the last year, actually got to work alongside Shannon and her amazing team in customer success. And then recently, I joined the Absolute in the leadership capacity to help kind of move through this kind of disruptive moment in the partner world in all things AI and agent force. That's fantastic. Thank you, John. And I know you're gonna tell us what's on the agenda for today. Yeah. Absolutely. So I'm gonna talk to you here. You know, I think we wanted to bring a bit of a diverse cast here with not just ourselves and our implementation experience at the absolute, but also bringing Salesforce and and Shannon's perspective in terms of just, like, really what makes this successful. So my time last year working inside the machine at Salesforce, I really got to see up close with a lot of different Salesforce and agent force implementations, and we just kinda noticed some themes that we really wanted to figure out some different ways in approaching, which we'll talk you through a bit today. From an agenda perspective, we kinda got some structured content here we hope to get through in the next twenty to thirty minutes minutes, leaving plenty of time for q and a at the end. We're gonna start talking a bit about the human side of this. You know, this technology is incredibly novel in our space and how how that how we think about interacting with that is a lot more akin to working with humans than it is technologies. We're gonna talk about the discrete data and knowledge problem behind these technologies. We're also gonna talk about agent force chat GBT and that whole conflict that seems to be coming up inside of organizations consistently and comparing those technologies. And then I think, really, one of the most sort of substantial pieces we've seen is just really kind of turning the implementation approach of technology on its head a bit, where we talk a lot less about building these things and having structured binary outputs in UAT, and then moving into sort of a more refinement model that we'll go into. And then at the end, like I said, we'll talk through the q and a post side too. So I'm gonna get us kicked off here and talk about the human side of things. So, you know, I see a nice little chat window here. We'd love to just kind of throw a show of hands or something. You know, how many people have tested an AI tool, said it gets something wrong, and then immediately thought, you know, this this isn't ready. You know what I mean? It's like, oh, this isn't working. You're sharing screenshots. You just you see it do the wrong thing, and and it kinda goes from there. You know? I think on on the other side, it's like, okay. So same question. How many times have you seen a new hire join your company, and they've made a mistake in the first month, and you're like, great. Let's fire them immediately. Right? Like, I'm not gonna ask you. Hopefully, no one puts their hands up there. I'm gonna assume all these thumbs up are from from the previous question, but you get what I'm kinda talking about there. I think it's a bit of a double standard we're looking at here in these these technologies, you know, that that there's some it is nondeterministic. You know, when we see AI get eight out of 10 things right, we say it's not working. You know, we we share the screenshots, so we kinda move on. However, you know, if a human's getting eight out of 10 things right or even six out of 10 things right in their early days, you know, that that's going to be a good thing. Great. You're doing well. Let's perform. Let's do well. And I think this is one of the fundamental things when you're thinking about deploying enterprise AI in general. But definitely, you know, into agent force being a good example of that. I don't think success means AI never makes a mistake because we would never hold a human to that standard. I think the success means AI is handling a meaningful volume of work at a quality level that meets or exceeds our threshold, and it's improving, which is the exact kind of vernacular you would use with a human that you are hiring, into your team. Now I just wanted to talk a bit about AI being a part of the solution, not a whole solution. There's so much hype out there right now and all of these different things. And, you know, we could have seven more podcasts on, you know, why I I personally don't believe in the SaaSpocalypse and why I still believe Salesforce is a phenomenal platform for doing these things. But, you know, I do wanna acknowledge that I think there's a real disruptive moment happening right now with AI. And I wanna talk a bit about examples of why I think you know, I wanna give an example of what we're doing here at dAbsolute as a way to think about the same thing. So so when I joined the organization six months ago, you know, we were all very much talking about, hey. Like, there's a thing here with AI, and we definitely think we can do consulting and professional services different. And when we started picking at that, that wasn't just giving everyone a bunch of licenses. You know, when we built an agent that was able to help us build better solutions for our customers, we realized that we probably need some context from the upstream discovery agents that we would end up building. When we realized that we needed data to be shared, we realized we probably need some governance in place. Then when we built an agent that was able to, you know, basically look at the suite of tooling that was able to deploy software for customers in a week, we realized we really needed to think about how we work with customers and what that impact would mean. You know? That that led us to at our business more holistically and say, okay. If we're gonna deploy AI, cool. Great. There's a strategy. What does that actually mean? So, you know, what you see here is, like, what we started to unpick, which was okay. Yes. AI tooling will be a part of this. And I'm keen to say this isn't just, you know, one thing, one chat GPT. It's not anthropic. It's it's a multitude of different things done in a variety of different ways. But it's also then a methodology change, shift, and impact that we're seeing happen with customers where we're able to move in a much different fashion than we have previously, and we'll talk a little bit about that later when we're talking about AgingGuard. But that's an example of how that's implementing. Just because these tools exist doesn't mean that everyone's able to use them. I can show you just as many great uses of AI as I can, terrible uses of AI, and that means that we need to spend time getting to know and train our team on those things. And if we can move faster for our customers, like, significantly faster, 70% faster, better quality results than ever before, you know, what does that mean for pricing and, like, the velocity of how things are doing? You know, I I wanna use this as an example of what we're doing internally, and I wanna I wanna encourage everyone to, like, take a step back, realize that no one's really gone through enterprise AI implementations at any meaningful volume together, and that we're all learning this at the same time. Take a step back, look at your business, and think about how you can bring AI in a fashion that's thoughtful to your organization. I really believe organizations that are gonna succeed aren't deploying the flashiest AI. It's not the coolest, the most tokens, the list, the crazy. I think they're actually embedding this technology into how they actually work. When we went through this whole process, we really started to see that, like, how important data is at the core of all this, and I'm excited to have Shannon take over and and and talk you through a bit of her experience there. Thank you, John. Okay. So now we're gonna talk about the knowledge problem in enterprise AI, which is currently one of the biggest risks to AI success. And we're gonna focus on a real world scenario from a fictitious company called Astro Airlines. So Astro airline airlines launched their first customer service agent last week. Now the c suite and the development team, everyone's holding their breath. They're monitoring the conversations in real time. And to everyone's dismay, the agent is failing to answer many of the questions that they expected it would. When it does answer, the response are great, but the team was really expecting a higher deflection rate. So here's an example of one of those requests. A passenger reaches out to the agent because they're trying to rebook a flight, and they don't wanna pay the difference with a credit card. They wanna pay with some combination of loyalty points and credit vouchers. So the passenger asked the agent, how can I rebook my multi city ticket and pay with points and two vouchers? I've got a digital voucher. I've got a physical one. And the agent, it searches, it pauses, it searches some more, and then finally it responds, you know, I'm sorry. I can't process your your request. Please contact support. So perplexed, the team really wonders what's gone wrong. Do they need stronger models? Do they need more compute power? You know, do they need, you know, to to tune the prompts further with prompt engineering? They followed all of the technical best practices carefully. So why is this happening, you know, and how can it be fixed? And, the answer here is actually not a technical one. The agent failed to answer due to what we call the institutional knowledge problem. So what does this mean? Well, it means that the person who knows how to solve these complex multi city change requests with multiple payment types is one is that one veteran customer service rep on the 4th Floor. You know the one, the person who solves all of the hardest support escalations just entirely from memory. Every organization has that person. So what the passenger actually experienced, it wasn't a glitch. This is by design. It failed to answer due to an institutional knowledge problem. So humans are naturally adept at improvising. We take an incomplete set of rules. We apply our past experiences. We talk to a colleague over the cubicle wall, and we just kind of bridge the gap. But enterprise AI cannot improvise in the same way. It requires explicit instructions and structured information to take decisive action. If the solution to a request was never written down anywhere, then that knowledge is essentially invisible to the new AI agent. It can be even worse if things are written down but conflict with one another because they're poorly maintained. That leads to agent confusion. Now it's confusing for human service reps as well when you have two managers giving contradictory instructions for the exact same task. But as humans, we're much better at dealing with that ambiguity. We don't tell the enterprise AI, if you can't find an answer, just do your best to respond. Instead, we say, follow your instructions explicitly. Make sure you find the answer in your approved grounding data before you draft a response. And before responding, always double check your response to make sure you followed all the instructions, and that the response is grounded in the company's, you know, official data and policies. So in this way, your new AI agent acts as a ruthless auditor of your company's documented knowledge. Undocumented or poorly documented processes lead to gaps in coverage. Each time the agent fails to answer, it's like holding up a mirror to reveal those missing pieces or conflicting documents. And this isn't an AI failure. It's a feature. It is organizational debt that the AI is bringing to the surface. And ultimately, if acted upon, this is actually a gift. Next slide, please. So this is where it gets really interesting. So earlier I mentioned that humans can improvise, that enterprise AI cannot. And anyone who's used a conversational LLM like chat GPT might think that's false. So let me be clear. Agent force and enterprise AI at large are not the same thing as a general conversational LLM. So let's say the airline executive who sponsored the project gets frustrated. He pulls out his phone, opens chat GPT, types in that exact same query, and immediately, the app generate generates step by step instructions, how to enter the voucher IDs, combine them with points, update the ticket. The response is beautifully formatted. The LLM sounds confident. It sounds authoritative, and it's cheerful. You know, and the executive is is furious. How did a free consumer app come up with the answers in three seconds, you know, when my team's been working on the corporate agent for three months? And if foundational models are so smart, why can't Astra Airlines just connect their agent to the Internet, give it strict instructions to act like an airline customer service representative, and call it a day? But the thing is that answer didn't come from Astra's official policy. It feels like magic, but here's what actually happened behind the scenes. The model synthesized information from its training data, which is basically the entire public Internet. And in this case, it pulled information from a handful of Reddit threads, a popular points enthusiast travel blog, and the transcripts from some YouTube videos, where they're exposing loopholes and airline reservation systems to bypass the traditional expiration dates. If the airline were to put a general conversational LLM in front of their customers, the core issue, it becomes liability. The company can't be, financially and legal legally responsible for policies invented by a Reddit user named points guru nineteen eighty seven. That's just not how it works. At the end of the day, a conversational LLM is not designed to run your business. A conversational LLM is like a very talented improv actor and its objective is to predict the next most logical sounding word based on the entire Internet. Anacuro carries zero liability for the correctness of the information that it provides. Now if you contrast that with AgentForce, AgentForce is a business orchestration layer. It's stateful, meaning it remembers the conversation and it knows the context. It's natively tied to your company's knowledge base, and your CRM and data cloud. So when a prompt comes into agent force, it triggers a sequence of actions. And these actions do in fact interact, you know, with those conversational models, but with very, very strict checks and balances built in. For example, it only searches across, you know, validated and legally approved internal data repositories, verifies the customer's permissions, it might trigger some flows in Salesforce to make API calls, check the passenger record, validate the credit credit voucher expiration dates, and it all operates entirely within a closed loop. So this is the the critical distinction. The measure of success, for general LLMs is conversational polish. But the measure of success for an orchestration engine like AgentForce is, did it do its job? Okay. So now we've clarified that these are just distinctly different technologies. But if enterprise systems only draw from approved knowledge and the company has gaps due to undocumented knowledge or poorly documented knowledge, what happens? Astra Airlines, they can't just freeze their digital transformation for eighteen months to write manuals for every possible edge case. So next, let's let's talk about what to do about it. So this is where organizations have to adopt a triage mentality. The way out is a fundamental shift in how a company views documentation. Companies have to stop viewing knowledge as a one time prod project and start treating knowledge as a product. It needs an owner, needs validation steps, needs accountability, and it needs a maintenance cadence. To handle any overwhelming volume, customers can apply the eighty twenty rule, look at historical support data, and focus strictly on documenting the 20% of topics that drive 80% of the routine volume. You can analyze historical cases. You can talk to those veteran employees. You might find that the highest volume of requests usually come from incredibly mundane mundane and straightforward tasks, things like standard baggage policies, simple flight rebookings, password resets. So really master documenting those high volume processes first. That's where a small amount of effort will have a substantial impact to start. But what about the edge cases? Well, at a baseline, AI agents must always have the ability to gracefully escalate to a human. That might be a transfer to a live chat agent. That might be an opportunity to create a case for someone to get back to them. And historically, that escalation has been viewed as a failure, but that is outdated thinking because escalation is now your diagnostic tool. Every time the AI hits a boundary and routes a ticket, the organization needs to ask, is this a technology gap or is this a knowledge gap? And more times than we'd like to admit, it's just a knowledge gap. It's easier to blame the software than it is to look inward at our own organizational rigor and documentation, but facing that reality is exactly how you escape this institutional knowledge problem. And in this new AI driven world, when a request is escalated, the human employee still solves the edge case problems, but they don't stop there. Then they immediately leverage the AI that's embedded directly into their sales or service console. So instead of spending thirty minutes typing up notes about the interaction, they use Einstein AI to read those transcripts, document and classify the resolution. And for any escalations which are recurring, they can use generative AI AI in the platform to draft brand new knowledge articles detailing the exact steps, the customer service rep took to resolve it. And now there's nothing wrong with leveraging conversational AI to help draft those articles or to research how the conversational model might have answered the question based on its training data, but the information has to be validated. A real employee needs to review it and give the enterprise agent the official stamp of approval to use it. And through this orchestration platform, your employees' expertise is captured, it's structured, and then it's fed right back into the AI's approved knowledge base. Each edge case that's resolved and documented actually expands the boundary of what your agent can confidently address. So that's the the concept of compounding refinement. The agent gets smarter and more capable over time because it's learning your actual legally approved business operations directly from your best employees, really one resolved edge case at a time. It's not relying on Internet folklore. And that's where the payoff for doing the hard work is really massive. When embedded in your ecosystem, AgentForce brings true action oriented power to the enterprise. And now I'll pass it over to Zane to talk about how all this comes together when dAbsolute builds agents. Perfect. Thanks, Shannon. That was great. So now the next question is, how can a company like Astra Airlines set themselves up for success with implementing AI on the Salesforce platform? Well, it's a different approach than a typical software development model, and we learned this through the this approach by using it, in helping a handful of our customers. And so I'm going to use one of our higher education customers as an example where we took their customer service agent from less than a 10% good response rate to a 82.5% good response rate just as of last week. And just so everyone understands, when I say good response rate, that means that the support reps felt that when they looked at the quality of the AI's responses back to their students, the content was so good that they would not change a thing about it. Okay? So the first thing Astra Airlines needs to needs to do is create a baseline of expectation. They need to ask themselves these four questions before they even attempt to start to build out an agent. First would be, what does success look like? How will we measure performance and adoption? And how will we collect feedback from our users? What are we willing to live with, and what requires immediate action? Now when I say that last point, I mean, what are the critical high severity situations that we should test for to make sure that the guardrails that need to be built into this AI agent are working before we go into production so that there's no major risk to the business or to the end user? And at the same time, what are the things that we are okay with knowing that this is not going to be perfect? K? That last point is going to be extremely important for Astro Airlines as it's the baseline that is going to help the company make sure they do not fall into analysis paralysis once they get into testing. Once they have their baseline, ASTRO now needs to look at changing their actual implementation approach. Like I said earlier, the old software implementation approach doesn't work in this AI world. The typical approach being 80% build, 20% refine. The 80% is us going through discovery, build, and UAT. The 20% is the refinements we make once we are in production. The approach that ASTRO should take when implementing AI agents is 40% build and 60% refine. That means that ASTRO should not aim for perfection, but rather, before going into production, rather they should build an AI agent, test all of the guardrails that are necessary, and align and align with what will, require immediate action based on their documented baseline and then go into production as quickly as possible. Now the reason for that, unlike traditional software where we have code that will behave the same way over and over, no matter, how many times it's run, AI has various layers of technology. We have an LLM. We have Salesforce. We have data. We have humans. And so there are so many dynamic layers within this that the number of testing scenarios that you could go through is going to feel almost infinite. So the best thing that you can do is to go into production as quickly as possible and get as much real world data to understand where the AI agent needs to, needs improvement so that you are able to extract as much value as quickly as possible. Most of the times where I've seen organizations not follow this approach, they still go through this lengthy refinement period regardless if if they had a long UAT because they did they did not consider the various different scenarios users could interact with it. So next, once, ASTRO has their AI agent in production, they're going to require a refinement model that allows them to optimize the agent week to week. And that refinement model would look something like this. First, they need to monitor the AI agent's performance on a daily or weekly basis, see how it's engaging with end users, what kind of topics are being leveraged, what are the actions that it's taking, how quickly is it able to respond, etcetera. Then they need to monitor implement a human feedback loop to gather users' positive and negative feedback. And then three, they need to categorize their findings into four main buckets. First is data. Is there anything about our database that's feeding into this AI agent that needs to be improved? Agent optimization. If the data exists but the agent didn't use it correctly, what do we need to improve so that it provides a better output in the future? User training. This the agent maybe is working fine, but users aren't providing the right feedback. And so maybe we need to train our users and educate them so that they can give more appropriate feedback in the future. And technical limitations. Data is fine. The agent is optimized. The user user's feedback is fair. The last thing to look at then is is there a technical limitation? Is something being asked of the technology that it's not capable of doing just yet? In which case, ASTRO can think of a workaround or educate the rest of the organization on the limitation so that they can continue to build trust around this technology. From there, once we have that, they're going to take all of this information, and they're going to have a weekly, biweekly review with the necessary stakeholders in the organization so that everyone understands what's happening and what needs to happen in order, to improve the AI agent. Now as a bonus tip, I would encourage ASTRO to not just track the agent's performance week to week. We also wanna compare the agent's performance over a long term horizon so that we can actually see the trend in the AI's performance. The same way as we track our employees and the people that we work with to see how are they performing over a over a long term horizon. So just remember that one bad week doesn't mean failure, and one good week doesn't mean success. If ASTRO follows this approach and model, they will start to realize faster value with Agent Force. And with that, I'm going to pass it on to my colleague, John. Thanks, A. Appreciate you. Couple slides left, gang. I think I just I've seen personally hundreds of agent force implementation, Shannon, Zay. Like, I got we're in a lot of lot of these categories between the three of us. And and these last two slides are just really, really important to me in terms of this whole approach of just really rethinking how we do it. And it's, again, not only is this technology novel to us, but we're starting to discover how we implement it is novel. And, you know, I think there's a bit of, like, humility I personally and we all have all had to go through as we've gone through that, and I just encourage you all to to feel okay walking that path. Couple slides left. So at the absolute, if you if you want us to do this for you, we can. So we we've created an entire offering called agent guard, which initially I thought would be something we'd offer as an option to our customers who are implementing agent force. We now include it in we include three months of agent guard with every single one of our agent force deployments because we truly believe this is what is needed for for these deployments to be successful. So all of those things you heard Zane talk about, we take care of that for the customer that helps us move in the right direction. You know, we monitor, we analyze, we optimize, we stabilize, and then we report back to you in an ongoing basis. So reach out to us if you think we can help you there. But I think what is more important to myself, Shannon, and Zane is whoever is doing your agent force implementation, whether that's yourself, whether that's another partner, You know, ask yourselves what happens after go live. Who's watching this tuning this and making sure it gets better over time? Come back to how we opened this webinar talking about an employee. Just because an employee has a great first couple of weeks doesn't mean you leave them alone. Right? Just because a company has a bad first couple of weeks doesn't mean you fire them instantly. Right? This consistent ability to help, like, govern and and go and grow these people is important. If your answer when you're talking about what you're doing is we'll just hand it off. Like, there's a hard cutover that's not gonna be appropriate. Right? I think the post go live investment is really not optional. And this is where enterprise AI earns trust or loses it. This is not an agent for specific thing. This is not this is when you are deploying AI in the enterprise. This is important. Takeaways. Right? Number one, check your expectations. So define success with data, not feelings. You know, we talked about the Zay talked about some very specific questions you can ask yourself before going through these implementations. Second is knowledge is the foundation. Right? I love the whole conversation about it being a gift in servicing that organizational debt for you. You know, invest in that knowledge as you're setting AI up to fail, and then build it and then grow it. This is no longer ship and forget. This is truly treating AI as a living system. We all know that there's so many different things that can tweak and change in the this ecosystem of enterprise AI tooling from the model updating, from agent force releases, from your data changing, from prompts changing. But, like, you need to consistently be thinking about this thing as a living, breathing organism. We are here for you whether you wanna chat through it or not. So please, myself and Zayn, love to have this conversation with you whether you're exploring a partnership or not. I wanna make sure that you're getting it right. We all succeed when we all succeed. We want everyone to have a great experience with AgentForce. Enterprise AI has been a lot of, I'll use the word fun over the last year and a half for everyone trying to navigate this in a business. But, really, I think we really wanna make sure that everyone is starting to see and feel it. We know internally that you can have amazing results with agent force. You just have to apply the right approach. You wanna see us physically? We're gonna be in a variety of different places globally. Well, this is obviously heavily skewed towards The Americas and a little stop into Europe. But any other exciting places you love us to come, let us know. And with that, I'll pass it over and invite Christine to come back on stage alongside my friends Shannon and Zane, and we'll get into q and a. Thank you so much, John. That was fantastic. And we've had a real flurry of questions come through now at the end. Let's start with a question from Sean who says, when we implement enterprise level employee agent, how do we limit the agent to answer social questions such as who will win the World Cup? Who wants to take that one? So I think I'll probably go first. And then, Shannon, maybe if you wanna take your I'd love love to know your thoughts as well. But so when you're implementing a agent on the AgentForest platform, being able to articulate to the agent what its its specific roles are, and what its instruction set is, what are the access in regards to data that it has. So you can control with an agent force and say, you know, we don't wanna actually give it web access. We want to give it access specifically to our enterprise data that's sitting within our Salesforce platform. And so if things like, for example, who's going to win the World Cup, if that information is not within your CRM database, or your enterprise databases, that is you know, that's not something that the AI is gonna be able to respond to. So having a clear scope, role, instructions, and giving the agent access to only the data that it's that is required in order for it to be able to do its job, that will allow you to then eliminate all the other different kinds of, you know, scenarios that someone could co go to the AI agent for that it's not within its particular role set. Thanks, Zane. Anything you want to add to that, Shannon? No. That that was a great answer from Zane. It's kind of a funny question because it says, like, how do we limit the agent to answer social questions? Like, who will win the World Cup? I'm trying to understand if if they're asking for the agent to answer like that or for the agent not to. Generally, an enterprise AI is not gonna answer who will win the World Cup because answering a question like that is not gonna be within the scope. Every agent has has topics. Every topic has a scope of what it's there to help to do. So if it's just asking conversational questions, generally, an enterprise agent is gonna be coded, you know, where if it sees something like that, it it will kind of kick it off as small talk and and not engage and say, I'm here to answer questions about x, y, and z. Is there something that I can answer? If you do want the agent to answer questions like who will win the World Cup, that's possible too. In that scenario, you give the agent a scope and you give it instructions, as far as, you know, you're here to answer questions about pop culture. You can answer any questions. You can give the agent a web retriever to search the open web. Technically, it's possible, but it's a risk to the company has to kinda weigh the risk of of what the agent might say and who might say that too. If it's an internal facing employee, then you could definitely work through that with the design. But the the agent's scope and the topics and the way it kind of, like, operates under the confines of that are really gonna give you flexibility to to do either option. Thank you, Shannon. We have a question here about managing client expectations with the build less, refine more recommendation. And how how do you manage that, especially if you're providing a solution for them and then it doesn't meet their expectations at launch? Do you have any tips on that? So the way that I'm reading this particular question, it sounds like you're trying to implement a solution or a customer's coming to you or a business is coming to you to implement agent force, and you're looking for advice on how to manage expectations around using this approach. I think the the first part of that answer is going back to that step zero or phase zero that I was describing earlier. So asking those four questions and setting what the baseline is. It's important to actually have a really candid conversation to understand what is the customer's goals and objectives, Yep. why, do they have these goals, and, you know, being realistic around what is the technology, going to look like using this approach in the beginning phases, but also being you know, understanding the fact that this will get create the results. This will create the goals that they're looking for. It just it's a different approach in terms of how we get there, but you are going to get there. And it's going to be a lot more of an efficient way to get there. I would say also as a piece of advice is when if you're looking at implementing AI within your business as it's your first time launching an agent, the most, the best recommended thing is implement something that is going to be internally facing first. The reason for that is because you're gonna have a lot more control over the, expectations, the change management, the communication, etcetera, with your employees and internal users, it's a lot easier to have that control and have that communication than to have it with external users, with partners, or customers. So you'll be able to build that muscle around what it's like to actually implement this technology within the business and be able to do it the right way without having to take on so much risk as a, launching this externally first. So That's great. We have another question that's slightly similar, and I think something that plays on a lot of people's minds when it comes to AI, and that is around good data. So one of our attendees has asked if you have any best practices to ensure commitment from the business in creating and maintaining good data so that leveraging AI makes sense and provides benefits. So I can take that one because I talked about data, and and that's a great question. So I would say start small first. If you're if your organ if you're if you have if you need help from a data perspective in your organization, which I would go as far as to say that's most organizations, Start small, so decide on your use case. A lot of AI projects stall out because they wanna create an agent that has, I'll say, like, godlike power. Like, I wanted to know how to answer any question, and I wanna know to know how to take any action in Salesforce. But, really, the agents that are bringing, like, a lot of value for customers are agents that have a specific purpose, and they're grounded on a specific set of data. So focus on what are those things that take up a lot of a lot of people's time that that can be automated with an agent. Evaluate what data you have there. And then agents are modular by design like everything else in Salesforce. You can limit the scope of what the agent will reference or talk about to certain objects or even certain fields in the object. So really have a very defined use case, evaluate the data that you have, you know, clean it up, and then test and test and test it. We have in in addition to, like, observability, kinda real time observability features, we also have testing center. And so you can create the agent, load up your data, have AI generate a bunch of test questions for you, and then test it a 100 times and and see what the agent does. That will let you know where you have some gaps in coverage to really work on to really work on that. And then I would also say, from a data perspective, there are many features built into Salesforce that will that will help you, keep your data clean and optimize your data and capture good data. Every time you work a case, you can create something called a case wrap up. And so it identifies, it uses an LLM to read that case and identify what was the issue, what was the resolution. And so as you're embedding AI in your platform to kinda work with and clean your data, that will also improve it over time. You also can do that in sales console, finish a call with a customer, have it write up the notes. So I would say kind of a combination of both of those. So start small, Alright. make sure you've done an excellent job of of of getting that clean, test the agent, and then, you know, as you develop your AI practice, leverage those AI features built into the platform that will sort of produce and cleanse and summarize and categorize your data for you as well. I love this answer, and especially just highlighting something really important that you said about agents having a modular design. So they don't have to have access to everything straight away. I think that's such a great point for people to remember is that you can start small, fix something really important, just use that data before moving on. That's a great takeaway. So another question that we've had come in is how much time do you anticipate spending reviewing an agent's output to ensure accuracy and effectiveness? What's the balance that supports growth and the concepts of iteration without dragging the overall process down? Yeah. I'm I'm curious to know what that last statement, what that what that means to the person who asked the question. But in terms of how much time do you anticipate spending reviewing an agent's output, ensuring accuracy, I think that it's very dependent on the type of AI and the use case. But on average, what we've seen when we're like, if we're talking about implementing AI agents, for example, I would roughly say, like, expect to see like, the turnaround that you're looking for is, like, 90. Like, usually, it would take like, you can implement an AI agent within in less than a month. But the time that it takes you to go through the process that we described, which is looking monitoring the AI agent, getting feedback, and then looking at how do we optimize this and making those optimizations. And then once you've made those optimizations, you need time for the AI agent to be reinteracted with and see did the behavior change? Did we get a better output? And so that all takes time. So roughly saying that the average is about ninety days. That's great. Thank you. Another question that we've had, which is two good questions. The first part being, before rolling out agent force, do you follow a structured AI readiness exercise across data processes and governance? Then the second part of the question, which I actually think is is the even better question, is what criteria would you use or recommend to identify high impact quick win use cases? I I think that's gonna be so important to people who show value pretty quickly when they're using things like agent force. I might jump in and take this one at a high level. You know, I think and we made the we did the same thing here at the Absolute where, initially, we were talking very much about, let's let's assess. You know, we're a consulting company. Let's let's figure out with you how ready you are for AI. We're gonna look through your data. We're gonna do all these things. And, you know, the answer is always gonna be the same. It's gonna be like, you're not very ready for AI because no one was ready for AI. We have not been building ourselves to be ready to go for this. I think that it's just a bit of a reframe around, like, where are you the readiest or moreover, like, where will you have the best impact immediately? But, again, this is where I would say what we did at the Absolute is we said we see the opportunity. Like, everyone in their head understands now this technology enough where they look at their business, their process, your organization, and you go, I see where this tool can help us the most. Right? You know, we gave examples through airlines. I walked you through what we're doing as a professional services. These are two very different organizations, and we could see where this technology applies to us. I think what's important is to bring everyone to the table and go, okay. Where do we want to impact, like, go after this? And where we go after it, let's go after it. It's important. It's gonna make a meaningful impact to our business. It's gonna help us to have the humans focus on what they need to focus on, and we can use this tooling appropriately over here. What that'll do is that'll set you on a path to say, okay. If there's value in us doing that, we're going to need to do some things to make it. So if you are not ready, which would mean the data and the governance and those things aren't in place today, you know, to use Shannon's example, we we rely on Terry on the 4th Floor to execute all of this for us. Okay. Well, like, that's been a problem forever. Let's fix that. That means we should build the data. We should instill the governance, and we will do that as part of this. Right? That's what becomes part of this overall implementation. I think too often people are looking at readiness, and then it's causing paralysis. Like, I can't. I shouldn't. Right? And then secondarily, like, when you go downstream and some themes that I've heard here about, like, you know, how much situation, how do I get customers to be a lot more comfortable, how do I have my business feel more comfortable with doing this, This is why that post go live is just it's just not optional. You know what I mean? And this is a fundamental shift in deploying technology where for the my entire career, I have trained and told people that don't worry. By the end of UAT, user acceptance testing, you will accept it, and it will be good exactly like you need it to be. And rest assured, you know, within twenty four hours of us going live, it'll operate exactly like that. This technology fundamentally doesn't operate that way, and so we have to take ourselves and our businesses through a bit of an evolution there that says we'll never be fully ready. So let's look at where there's value. Let's make sure that we understand this value. Let's make sure we bought in. Let's make sure we understand that what we've all been playing within our own homes with ChatGPT is not going to be what we do here for all the reasons we spoke about. Let's also make sure we sit down and ask ourselves those questions to say, what is success going to look like in this refined area? And then let's also commit as an organization to doing the work post go live. But like I typically, within the first, like, ninety days, you're able to get to a place where you are feeling a lot more complete in terms of the output that you're looking for and sort of walk along walk along that path. So that would be my kinda high level component through to that. Yes. Like a mind a mind shift as well as a. technology shift. So we have another question here about how do we implement CICD processes for agent force deployments? Any any guidance on that? It's flawless. Right, Zane? Everything's just super interoperable. AgentForce is really stable. Like, there are no tweaks and quirks with it. It's never changing. Right? So it's, really. easy. You you it's really easy. In fact, you just ask the agent to deploy itself, and it's good to go. That's actually a really good question. And to be honest, that's probably more so my technical team will be able to answer that far better than I could. So we can do this in kinda two different ways. Shannon, if you if you have an approach that you have in mind, we can go with that. Otherwise, actually, what I might do is actually might come back to you with a full on proper response because I feel like I just would not do that justice on this, particular webinar going off the going off the top of my head. So That's no problem. Did was there anything you wanted to add, Shannon? I. mean, no. AgentCourse has the same development pipelines for programmatic development as other Salesforce tools. But I think it definitely with kinda, like, the shift in mentality, there is kind of a kind of a re sort sort of a reset to the order in which things happen. So, I mean, I definitely it's the same story as as any other software development process from you're gonna wanna build and test agents in sandbox and deploy those to production. You know, we fully support that. Now as far as, like, the like, the kind of ongoing iterative changing something today, changing something tomorrow changes, I think that's where it's a little bit different. It's not a, you know, big deployment project and then everybody says that was great and moves on. So I think folks have to be kind of thoughtful about, you know, really about, like, how often you're gonna refresh the agent and and sort of the more iterative, you know, always working on something, always deploying something. So I imagine that the the deploy the deployments will be more frequent, sort of faster. You might be on a weekly or a monthly schedule. But, you know, once the agent is ready to go from sandbox to production, we really recommend the exact same sort of, like, you know, technical process in that. So I am not sure if the question is really more around, like, the technical of how is that done. But I would say, really, the big difference is just kind of the cadence will be a lot faster because you're gonna be recognizing optimizations a lot faster in the agent process because it's always kinda like that test and learn approach. Yeah. That makes sense. Well, let me ask a question that concerns me and I'm sure worries other people, and that's around consumption licensing. It's something that can be quite hard to understand and that people get a little bit nervous about. Do you have any thoughts or tips on how best to manage that? Yeah. I can take the first go at this. I think consumption license is a real it feels like a monumental shift. But, again, I would just reframe it differently to be all this means is you just have to be intentional with the value of your spend. So when you have licensing that instead of just being you know, we we for years, if you've been in the Salesforce ecosystem for a while, you know, you're used to seat based licensing. And we can say we've adopted Salesforce, and we're using the value in this license, but that use could be, okay. I've set up two flows, five custom fields, and someone's logging in every day. Okay. By that value, yes, I've gotten used, and I'm using it every day. We all know that you could do way more typically with that seat license. And so the the incremental value shifts is heavily there, but it's always been abstracted because you just get used to the OpEx spend every year on, like, your licensing. As long as it you know, no one's complaining, it's kind of fine. I think just globally, we're gonna see the shift to consumption licensing, and I think it's important, actually, for a couple of reasons. I think it's gonna enable us to really sit there and just think about the value of what we're trying to do. Right? And I think everything we've spoken about today means you could confidently go and sign up for some consumption licensing because you're going to do the work of monitoring it, it, and making sure that there's value in all of those kinds of fun things along the way. Also, technology is moving at a rate that is consuming power at a rate that's pretty unprecedented. And I actually think it's on us to to really make sure that we're ensuring that we're using that resource wisely, and I think this is forcing us to do that in a manner that's that's kind of more appropriate and and at scale. So there's good reasoning behind wanting to move towards it. Always, if you're worried, start small. You know, Salesforce are incredible. They have a whole variety of things they can do to help you to, like, make that transition and feel confident with it. But like I say, really, at its core, it's just it's a value based conversation where you should be able to say, this is the value this is providing. And, like, if I do more of it, it's costing me more. If I do less, it's costing me less, etcetera. That's great. do wanna quickly add just one more thing to that. That Salesforce actually has some wonderful tools, one of them being agent force observability. And so you're able to leverage that piece of technology to get an insight in terms of, like, let's say, one particular AI agent, what are the the topics, the actions, and the things that it's being asked up to do. And based off of that, you can see patterns. And based off of those patterns, you can figure out, okay. How can I, set up the agent so that way it is more cost effective based on how it's being used historically and, you know, make it do certain actions in conjunction with other actions to save, you know, on specific, to like, saving on tokens or saving on credits here and there? So you can you can leverage the technology in a way where you're not just getting insight into the agent's performance, but you can also get insight in terms of, like, how it's being used and where it can be optimized from a cost perspective as well. Yeah. That's fantastic. Thank you, Zane. So I think we have time for just one more question. And something that's been going around my head that was mentioned that's really stuck there is this concept of readiness paralysis. I think that's so so relevant. So I wondered if perhaps we could finish with top tip from each of you about what would be your recommendation for someone who's just about to get started. What would be your number one thing that you would say? Here's how you can kind of take that first leap off. Maybe we can start with John? Yeah. So we've started this whole process internally where we're starting to think about different tiers of AI inside of our organization. There's a sandbox tier, which is where we have allowed everyone to have access to a variety of these amazing tools out there where we say, go play. Go get understand how these things can help you. Understand how this technology works. It's simultaneously the most complex technology ever put together in the world of software, but at the same time has the easiest, you know, interface, and you can ask it, like, how to work with it. So go play with it. And then what we'll do is as you understand and you go, I see where this fits into my business, we'll promote it to a second tier where we're governing it in a much more structured fashion where it's touching enterprise data, etcetera. So pilots and playing is important. I see it's incredible to me still the amount of people who are just not using this technology in any fashion. And so so just start by encouraging your business leaders, all of the people in your life, to just play with this technology and get an understanding of it. Thank you, John. Shannon, how about from your perspective? What would be your your number one tip for someone looking to get started? I would say to get started, pick an out of the box use case. There are a lot of things that agents just do by design. They're they're low code. You set them up. You test them out. I think you get you'll get you gain a lot of confidence in how an agent works just from getting your hands on it. Agents can be fully customized to do, you know, a 100 things. But in order to kinda, like, just learn how to get started with an agent and feel comfortable with an agent and really get excited about it, I think people get really excited when, you know, they kinda see the power of this unlocked. And we have a lot of, like, really streamlined, set it up in an afternoon. You've got something working by the end of the day for a lot of those kinda out of the box, you know, AI and agent force features. I would say start with one of those to get familiar with it. You'll feel far less overwhelmed, after an experience like that. That's great. It gives you an idea of what's possible, doesn't it? Which can be really hard to picture with something like AI. And Zane, what about you? What would your top tip be? I think Shannon basically took the words right out of my mouth. But I would say, you know, depending on your maturity and your comfortability, like, start with an internal use case and start enabling your internal team to start leveraging this technology. Start creating that shift in mindset around this is not something that is, here to take away work. Rather, it is here to be assistive in helping us do more do our work more meaningfully and more impactfully. So I think that's the the the biggest way that you can make that transition is just giving this technology, whether it's using generative AI all the way to agentic AI, create that assistive capability internally within your organization and go from there. That's fantastic. Thank you so much. Well, with that, sadly, it is time for us to wrap up. But I wanna say a huge, huge thank you to Shannon, to Zane, to John for a fantastic presentation and to the Absolute for sponsoring this event. Also, thanks to all of our attendees for joining us. Don't forget to follow us and like and subscribe and stay up to date because there are plenty more amazing webinars with great panelists that we'll be putting on. With that, thank you once again for joining us, and we hope to see you next time. Thank you, everyone. Bye. Thanks, everybody.