Name: Workshop: How We Built An AI Sales Agent On Salesforce
Uploaded: 2026-10-20T16:30:56.285Z
Duration: 59 min 12 s
Description: Workshop: How We Built An AI Sales Agent On Salesforce

Transcript for "Workshop: How We Built An AI Sales Agent On Salesforce": Hello, and thank you for joining our webinar today with Salesforce, Ben, and Nutrient. We have a very exciting session today showing you how Nutrient built an AI sales agent on Salesforce. But before we get started, a little bit of housekeeping. The recording of the webinar will be shared with you after today's session, and we will have time for q and a at the end. So please feel free to put your questions in the q and a tab as we go along. A little bit for you about Salesforce, Ben. Our goal is to help people find their path to success. And on the next slide, we have some examples of the type of things that we offer. We're trying to share practical, relevant, and accurate content in all sorts of ways from articles, newsletters, events, YouTube, podcasts. There really is something for everyone, whatever your preferred way to learn. If you haven't explored everything that you can see on the screen right now, perhaps take a look after this webinar. But without further ado, let's get started. My name is Christine Marshall, and I'm the technical content lead here at Salesforce Ben. And I'm thrilled to, be joined today by Greg Lives and Richard Benoit from Nutrien. Welcome to both of you. it's so good to be here. Thank you so. Thank. you. much, Christine. Alright. And thank you to everyone who's joining us today as well. And so to kick things off, Richard and I are gonna treat today's session as a a fireside chat. You know, we're gonna keep it very conversational and fluid. So, you know, we'll first kick off with, you know, discussing all the challenges that led up to building the agent in the first place, you know, which we're hoping actually, resonates with most of you, you know, these these same challenges. Then we'll jump in and show you a quick demo and, you know, about how the agent works, what we did, how we set it up, followed by some tips and rules you should follow if you're gonna build your own. We'll even share at the end some of the results we've seen after the first twelve months. And then, finally, you know, before we wrap things up with a quick... You know, before we wrap things up, we'll end with a, just a a quick word around Nutrane. Alright. Oh, and just to repeat what Christine said, there will be a q and a at the end. So please stick around if you have any questions, you know, along the way that you'd like to ask. So... Alright. So with that said, let's jump right on into the presentation. Alright, Richard. So this project, this agent has been your baby, for a while now. Right? You know, I think a little bit over a year at this point. And we're talking about from the time that you came up with the idea all the way through the build, you really brought this to life, and and and I've seen it evolve. Right? So let's let's go back and start at the beginning. You know? So can you help set the stage and tell everyone what led you to create it? Yeah. And, apparently, I was trying to replace myself, in in our workflow. We have been around for about ten to twelve years. I've been with the company the entire time since we founded it, and pricing has always really flowed through me. And there was a really good part about that because I inherently have all this industry knowledge. I have a lot of muscle memory. And so, like, when we run a very bespoke pricing motion, so we had a lot of different ways that we did deal desk. And so with a really technical product, there's a long arm ramp. So, you know, traditionally, you know, with a lot of SaaS solutions, you have a per set See pricing, there's a discounting motion with AEs, but we worked a little bit for more of, like, we have a floor of what we wanna charge, but then we we go above that. But inherently, like, everything ran through me, and so that was a cumbersome... It took a lot of time. Vacations were never really on vacations. And so the idea was how could we inherently build something that could replace me as the gatekeeper for pricing, but still have all this knowledge that we do that we do it. So, like, a lot of it was gut based, but we were never really to do postmortems on it. And so, there's a lot of holes in it. And so, like, that was kind of the thing that I started out to see. Like, could an agent potentially provide. value Yeah. Yeah. Certainly, the way that it was current... Or the way it's set up in the past wasn't sustainable or scalable. Right? And you're only one person. So I think that... You know, here? I I. actually think that that's a that's a pretty common challenge or a pretty common scenario, I think, especially at companies, you know, our size. So... Alright. So we understand, you know, what was going on. Why don't you walk us through, like, okay, the the evolution? Like, where where did you... Where'd you start? Start with chat g b t where I think everyone did with chat based kind of interactions. Like, you know, I think ultimately was talking to it through a chat interface and, you know, letting it... Just giving a very high level, like, what would you do? Like, question this. Be a thought leadership partner. And so, like, the chat based interface is great, but, like, there's no memory, there's no rule book, there's no bible for it to really understand what you're asking for because AI so much is always guessing what you want. Right? And that's it's... From. a chat interface perspective, like, unless you give it all this context, it really struggles to continuously provide value. So then ChatGPT comes out with GPTs, right, which allows you to give it maybe its compass of, like, this is what I want you to do. You can continue to go back to the GPT, which is great, and then you can upload documents to it. But I think the limit's 20. And so, like, it got me closer, but there were still all these limitations of working with the GPT. Or if you're using, you know, Gemini, it's a gem. But there are so much. limitation, and it wasn't able to learn from itself. And I was still having to go from Salesforce to JAT GBT back to the deal desk. I'd have to pull in Gong manuscripts, like, like, it's still very manual. It was getting me closer, but it still was pretty cumbersome. But it was a good way to learn. Like, I had started the. process. And then Claude basically kind of really helped revolutionize this a little bit, especially with Claude codex. Like, the connectors. Okay. Okay. Now I can connect Salesforce. I can connect Gong. I can connect a lot of these systems, and it's able to pull this information from... But the automation of it was really, really struggling. So our previous founder, Peter Steinberger, which is our technical founder of our parent company, actually is the one who, came out with Cloudbots, which was, like, this big storm of, like, how you can actually take Claude and build these, like, self sustaining agents. And so, like, obviously, since he was our technical founder, we had, like, a direct access to him to kinda pick things off. I think he got pretty annoyed by the end of us kind of always asking him his question. But, like, that really opened up the ability that you could build an autonomous agent that could sit. And then the next question was just like, how do we do this that it could actually safeguard that we don't just give it access to our entire business model and go crazy? So, like, that was kind of the small progressions of how we got there. And we we we brought the sales team along the way. So, like, we build the GPT. We'd love everyone access to the GPT. So we build a a cloud agent. We give everyone access to the cloud. Okay. Now we have a fully autonomous agent. How do we give everyone access to that? So, like, the go to market team has actually kinda followed us with this journey, and we've gotten a lot of feedback. So, like, it it was cumbersome and we're still learning, but, like, it wasn't just like, oh, we can fix this. Like, we kinda have waited for the technology to catch up with us. Yeah. Absolutely. You know, it's it's sort of crazy that, you know, in an adjacent team to get you with, you know, with the the sales team. We we obviously work very closely with y'all, so we've seen it evolve. Why don't you walk us through? You know, this started off as, you know, really, really like a pricing bot. Right? And something to to help you out on that note. But then, like, what's it evolved to? Everything. Wow. So at first was just kind of a... It was a a pricing bot that was supposed to question reps, help them with ballparks, but as it became more powerful and has more information and more access to stuff in memory, right, because it's learning from itself to some aspect aspect. So I kept it to myself probably for six to twelve months and just basically yelled at it for for six to twelve months. Like, no. Don't do that. Yes. Do that. Okay. Maybe close. No. Why why are you doing that? So, like, there was a lot of internal q and a and and staging for me. But then once it... We were able to get to a confidence score and kind of this rule, so we started, okay, let's have a couple AEs tested out. Alright. Then let's have a couple CSMs tested out. K. Why don't we give it it to marketing and let them start using it for their stuff? So it really did kind of expand a little bit. And so what it ended up doing is, okay, we just went from, like, a new logo pricing pod, and I'm like, okay. What about renewals? Can it run our renewal playbook? Okay. What about upsells? And then, like, once it got that, you know, the ability for it to learn stuff is like, okay. Let's give it our design template. Could it do decks? Could it take a deal desk ticket and build an entire proposal deck on it? Okay. Cool. Alright. Now can it actually research companies and prep us for stuff? And then the last thing was just like, okay. Now we need us, like, a regulated audit circle so that it's coming back and going, is my work good? How are things happening? Because so much with the GPT and the Claude work is it's one directional. Right? Help me with this. Help me with this. Help me with this. But you never go back and say, hey. This is how it ended up. So it... Once we were able to kinda complete that loop, and we'll show it in the demo of how we were able to complete that loop, it became so much self aware and being able to kind of question itself and learn to just not. So... Yeah. I mean, Alright. right now, it does everything from prep meetings, do proposal decks, like, do pricing, account research, flags for upsell and expansion. So, like, it's pretty ingrained into everything at this point, which is cool, but it has this compass of what it all means, which is really the hardest part of what took so long before we were comfortable sending it out. Yeah. Well, that brings us to the big moment, where we introduce you to the agent. Alright. Everyone, meet Dealie Madele Face. And now, Richard, you know I've gotta ask. Where did the name come from? And it came up with its own name. As I was... When I was going through it, I asked it to name itself, and so it came up with Delia McDeal Face. And so it's kind of stuck. And so we have some sub agents we're working on, but it's been great. So Deli McDeal Face, their stickers, they're on all the comp... Like, it's a household name here at Nutrien. So, yeah, I I can't. take credit. for the day, but it's great. Yeah. But... And we love Dealey. Alright. So let's let's take a quick look at, at Dealey. Yeah. So awesome. So we'll jump in. We have anonymized some of the data in here, but this is what deal.ly basically does. So everything starts from Salesforce. Right? Ultimately, like, we've got an opportunity. You know, we run Medicare at Nutrien. So a lot of this information that we were initially giving deal.ly was, like, the why buys and stuff like that, the different contacts on the account, and looking at what what was inputted from the discovery calls and stuff like that. So Salesforce begins on, like, our... It's it's our buy bar. I think so. What we then did is we had a self automated deal desk that we put into Salesforce that would kick out to a workflow automation platform that we have that would prefill everything for the reps, which they loved. And so all... The only thing they were really responsible is giving out the rationale. What deal.ly would do once they got the ticket is look at basically the entire content of which it had access to. Gong, Gmail, workflow, what we've done in the past, Salesforce open account. But I also trained him to kind of flag things that were AI written because as you know, there's a lot of things happening now where, like, you can automate stuff written to Salesforce, which inherently, like, when you had AI on top of AI on top of AI, you start getting slop. And so the interface for deal is in Slack. So you can at him, basically, once a deal desk ticket is set from Salesforce, you can con... You can converse with him what your rationale is. And so he's basically gonna go back and forth in terms of saying, you can do this, you can do this. Oh, maybe we should talk to Richard about this. Dolly Mcdealface is a sub agent that looks for AI generated content in Salesforce and then questions the reps. Because, like, we use Gong internally, but Gong auto writes Medic to Salesforce. What we've seen though, unfortunately, is if AI is already... Always writing Medic, the reps don't really know it. So what Dolly does with Dealey is go, hey. I actually wanna question these AI written things and update them in Salesforce so we have more genuine information. And so Dolly and Dilley work together and basically then give a recommendation what the pricing bands are. And so what the cool part is, I've gotten to the point where there's some confidence with Dilley or he will go, wait a minute. I think you need to go talk to Rich about this. Like, this is gonna be... This is a big deal. I think you're way off. But apparently, then we have an approval history in Salesforce for what's approved. So, like, what's nice is, like, there's a loop now that when we run commissions at the end of every month, it looks at the approval amount from the agent, what was approved on the order form, and it flags it if the reps went outside the approval. So what we see sometimes is, like, reps love the discount to win. All our reps have a 10% back pocket discount for the price that's approved. But if the order form doesn't match what deal he approved, we have this audit system and we have to understand what it is. So, like, it's a full cycle there. I think the most difficult part was getting the Slack interface really to work Yeah. so that we can go back and forth with Salesforce. But, I mean, Salesforce being our record of system, like, the audibility to this stuff has made it so much easier that, like... And it's... The way we set up the deal desk process to say you have to have everything filled out in Salesforce to initiate that deal desk process has made our Salesforce data so much better, so much cleaner because the carrot is, hey. We have an agent that can get you pricing approved in, like, thirty seconds, but you have to fill out everything. And by the way, if it's AI slop, Dolly's gonna call you out and ask you to update it. So, like, it's really helped two things. It's helped the reps move more efficiently, but it's gotten us so much better data about the prospects, the company, the loss reasons, and stuff. And then if the reps go completely rogue, they get in trouble. Dealie calls them out. We get an audit. It's like, hey. By the way, Jonathan got a little out of hand. He he gave this for free. So, yeah, this is this is where we sit right now with it, Greg. Yeah. That's great. So you actually teed me up for my next question perfectly. You know, so like you said, it's been a little over a year now with Dealey and and now Dall e even just as of recently. You know, I'm sure you've learned a lot along the way. Right? You know, what advice would you give the viewers if they wanted to build something similar? And I know that you have some rules that you you kinda... You know, that that you got lined up, and we'll talk about those. But, Yeah. yeah. Go go ahead. Be patient with agents. I think, like, I don't know. I have a one and a five year old daughter. I I think I'd say the same things to the agent that I say to them sometimes. Like, they act like children and sometimes they just get stuff completely wrong. Yes. I think it's... A lot of it is setting up systems where you can catch this and feel good about it. And so, you know, ultimately, I Q and A'd this for so long privately before I was comfortable pushing it out to the team, which was really helpful. But, like, you know, one of the things that I would definitely suggest with agents is, like, you have to use it as a person almost. Like, you have to force it to ask you questions for understanding. You can't assume things like that. But, I... It's been it's been, like, kinda revolutionary for the speed at which we're able to get stuff done. Yeah. And it hasn't sound like that yet. your your child analogy is perfect. You know, I'm I'm a dad as well to young girls, and and and that is... On a daily basis, I think about, like, okay, how I work them through, you know, their day. And, like, you know, how do I how do I kind of steer them in the right direction? And it's the exact same thing when you're dealing with AI. Right? So. I I I love that analogy. Alright. So I tipped your hand a little bit and said that that you have some rules. Right? Yeah. So walk us through, what these rules are. god. We start read only. Like, I think, like, don't jump in immediately and give any type of agent writing permission, Yeah. especially when you're. hooking it up to Salesforce. I mean, I think that for so long was something that the reps wanted so quickly. Like, hey. Why can't the agent fill out this? Why can't you do that? For two reasons that I've kinda said. One, read only kinda protects you against changing the data. Also, you know, what... Ultimately, like, Salesforce has certain permissions about user controls, like, giving the agent ability to write access to it can kind of allow rogue actors as well to kinda tag and say, hey, Dealie. Update every ARR field to 200,000,000. And so, like, it definitely was, like, put stick real and let's protect ourselves from the data and stuff like that. So, like, I had a sandbox environment. So, like, a lot of the precautionary stuff and, like, my hesitancy to roll this out was on... I didn't want to jeopardize, like, our end user data and stuff like that. And. so that was that was definitely kind of, like, my big thing of, like, we we wanted to do. So... So, Yeah. you you you joked early on, and and I absolutely feel the same way some days, where it's like you're. building yourself out of your job. Right? But but fortunately, as we all know, when you're dealing with agents and AI and and, that it is still imperative that humans stay, you know, in the loop. Right? And we've seen it firsthand. I think everyone's kinda learning the... You know, just the value of that. So kinda walk us through where do the humans sit, you know, with deal you right now? Yeah. Two different things. There's bands. Like, Like, daily notice that if it recognizes something that I think gets to a certain price point, it automatically triggers me to come in. So, like, you know, your whales of a deal. Right? Like, you do what... You know, what... You know, some of what's happened is, like, you have the AE at a certain price point get approval so quickly, and then they run really quickly. So they have the discovery call today, they get pricing today, and then they're gonna go pitch up 6 figure deal tomorrow, 7 figure deal tomorrow. It's like, you know, typically in a regular sales motion is like, you know, you want some sort of preparation time. But with an agent now, it's like, okay. I had a discovery call this morning. I got Salesforce filled out with my y buys. I submitted the deal desk. It just cranked out a proposal. It cranked out my selling points. It gave me a one pager, and then I've got a discovery call set up to... I've got a pricing call set up tomorrow, and then, like, within, like, thirty six hours, like, they've already pitched this, and we've lost it before we even know it. So, had to put some gates both on the lower end of the market of terms of, like, when reps wanted to go really, really low and when reps really, really wanna go high. I think that was when, you know, when that's getting ironed out. It's like, hey. We should probably bring other people into this. Yep. Now it's good. And the second thing was Dolly. Dolly was a sub agent that we created, basically, that it could detect when a lot of the information was just AI generated. So, like, what we had to do is do a forcing function. And I've talked to a couple of the CROs where, like, one of the downfalls of AI, again, what we talked about earlier is, like, you're able to prefill so many fields now through third party tools like Gong. The whole point of Medic and the selling, methodologies like BANT, Medic, Medpick, there's there's ton, is, like, the reps understand the pain points, the decision makers, and stuff like that, and they intentionally go get this stuff out because they can run a deal process. With agents, they look at a transcript and then guess what they are, fill it in, and what we saw with the reps is they only... They came a little bit lazy about it. Like, they're not asking these pointy questions. So we'd sit in pipeline reviews like, hey. You. quoted this thing for a $100,000 yesterday. Like, when's... What's the timeline? I don't know. And they would go check Salesforce to see what it said because they're hoping the agent picked it up. So, like, having. the ability to still keep the human in the loop, especially in that discovery phase, Yeah. I... That was really important for us to do because, again, the agent's only gonna be as good as the information it's given and what its compass is. And so, like, right now, the world is changing so dramatically, especially in software that, like, agents are always good looking backwards because it has information. Like, all the models are trained on preexisting data. The models don't know what's gonna happen in the next six to twelve months, but we're in a place where most software providers are quoting and going into a business model that, like, is changing on outcomes and stuff, and, like, agents aren't really equipped to do that. So, like, where to put the human loop is definitely on, like, the data collection and then certain pricing things. And so, like, Yeah. that is where we've had to put some forcing functions, which is like a happy middle ground. I think, you know, reps... I mean, reps always, I mean, complain. You could have two fields in Salesforce, and reps are still gonna be like, that's too much. Just let me sell. But, yeah, that's a bounce. Yeah. That's good. And you were cert... You you touched on there for a minute too about, you know, how you gave, you know, Dilly and Dally, like, guardrails. Right? And kind of pointed them in the right direction. That actually tees up the the next rule, I think, perfectly. Right? So, expand a little bit more on this one. Yeah. So, I mean, again, with agents, what... Or just AI in general, AI is guessing what you want at all times, and it learns you over time about what it thinks you want. So, like, you know, where should I go to eat the dinner tonight? Like, okay. I've asked for six Italian places in the past two months, so it's gonna go like, okay. I think you should go to this. Like, it's always guessing. And then everyone's Salesforce instance looks different. Everyone has different terminology, whether it's subscription, you know, whether you sell a certain business model like ARR, GRR, all these things. Like, everyone has different things. Like, we're we're owned by a private equity company, so, obviously, they're asking for stuff. So, like, one of our wonderful things that our COO did very early on is he built this basically glossary of how Nutrien works, and that was feeding org charts, terminology, like, definitions of stuff. Like, we fed it board decks. We fed it so much stuff. And that became ultimately, like, the thing that we gave every agent or really everybody using even cloud. Like, you need to ingest this from Notion before you even start. Because when you're asking an agent or cloud to say, hey. Find me the biggest deal in Salesforce. Like, what does that mean? Is that TCV? Like, is that the total contract value? Is that the biggest... Just the biggest upfront payment? Is that the biggest usage fee? Like, it's going to guess what you want. And so if you're using all this stuff for your board reporting, your decision making, your comp structure, like, it needs to know what you mean. And so, like, Yep. we wrote out basically, like, this is Nutrien, and we feed it to everything. And from that basis, it understands when I go, tell me the biggest company. It's gonna go, that's gonna be ARR plus usage, and that's our biggest customer because that's how much annual revenue they give us. So, like, that helped us so much more, and then we consistently update it and things like that. But, like, a little trick that I've done with a lot of my AI, interactions and even dealy is it's prompted to always ask three to five questions for understanding. It's a quick thing you can do. If you have... You're. using chat g p t, you're using a g p t, using cloud, you can go in and give it operating structure. Before you answer me on anything, ask me three clarifying questions and then remember that. Because then it doesn't assume, oh, Rich, you wanna eat tonight? Like, oh, what are you in the mood for? Italian, steakhouse, you know, Chinese, what do you want? And that helps it be better, and then you tell it. to remember that. So, like, you give it this base, but then you give it all this information. Because, like, Salesforce itself as just a human sometime could be different. I don't know how many questions I ever heard about the rep. It's like, my ARR is off by 4¢. Why? Like, well, how much commission am I gonna get off of this? So, like, just giving it that information makes it so much more valuable, and then you can trust the information as well. Absolutely. Yeah. Yep. Yeah. No. I... That's that's one of those tips that it sort of ups your game right away when you start employing that. So a, 100%. the... We go back to the kid thing. You're... Like, at some point, you're a... Why do you ask so many questions? Just answer the question. Like, yeah. But, like, it's in this balance of teaching and stuff like that. So, Yes. that's. been a big tip on when training an agent. It it it can learn because you're... You're... It gets curious about... Especially with pricing and stuff like that. So, like, Yeah. what it really did with deal.ly is when it trained on my stuff, it started to pick up the questions I would always ask the reps, and now it does that when the reps submit the deal desk. It'll come back in that Slack that I'd be like, hey, Greg. It looks like you forgot this discovery question. Are you sure they wanna start tomorrow? They haven't even evaluate us. And then, like, what's the budget cycle? And they go get that or know that without me being in there, but it's learn what I typically. ask, which is good. Yeah. No. That's a that's a great guardrail. I mean, because, I mean, obviously, one of the first things you think about with automating pricing is like, oh, that's that's that's scary. Right? And so the more you question, the more, parameters and, obviously, governance that you employ, I think is is just critical. So, speaking of which, I think that that actually plays into your next point as well. Yeah. We talked a little bit about the audit logs. Like, I think the one thing with agents and just like children, I think I keep going back to, like, you have to watch them. Like, you can leave them in a room for a while by themselves, but, like, you need to go check-in on them every once in a while, make sure nothing's broken. So, like, I think the auditing for, like, two things we've done, you know, on the on the front line than just, like, what's been improved and stuff like that. But, really what changed by using an agent versus, like, a GPT or a chat based interface or even Claude was its ability to run a routine once a week and basically evaluate. Because in the early days, you know, we had a deal desk platform, which is our workflow product that we use here at Nutrien, and then I would download the ticket, stick it into the GPT, then it would give answers. I would do stuff, and then I go back into the workflow product. We'd hit submit. The reps would then update Salesforce independently. And so, like, I had, like, three different systems of information. But the hardest part of, like, does the... Is the agent working other than, like, productivity? It's like, what happened? And so the audit log when we asked daily to do once a week is like, okay. Now go see what happened. Okay. You quoted this. Where's the deal now? And that's the beauty of building it on Salesforce. Like, you quoted here, what did that do to your time to close? You quoted here, did it close, did you close lost? And so, like, we're getting so in tune with some signals on, like, if we quote this banner, time to close goes down to this because it's all built on Salesforce and there's a system of record. And so, like, daily is now coming every Sunday at 7PM because that's when I do my forecast roll up for our CEO, on Sunday nights before we have the forecast meeting on Mondays. Like, it's giving me more real time feedback. I'm like, hey. It looks like our average initial offer has jumped 30%, but our win rate stayed steady. Okay. That's a great signal to know that we might be undercharging or vice versa. Like, hey. We... Like, internal build is now your number one loss reasons from the past week, Yeah. and your pricing has stayed the same. So, like, once you have it built on Salesforce, you can have that loop on the stages and time to close. Like, the information that comes out is, like, it's it's it's the best thing. I mean, like, I mean, if I was at RevOps now, like, the amount of stuff I could pull out now is great. So, like, I think that's. really, really good. And then, like, Salesforce is just a great system of record because every time you click, you move everything, it's there. So you have to give it access to it. Yeah. Yeah. Which... And, look, you've teed up your final point, you know, extremely well. Right? And and, look, given the audience, I think this next step. isn't gonna come as a surprise to anyone. So tell us about where Salesforce lives in all of this. Yeah. I mean, it's... I mean, it's, I mean, it's basically the foundation of this. Right? I mean, like, ultimately, I I forget. We hired an amazing, sales op person several years ago and, like, we adopted... Like, if it's not in Salesforce, then it doesn't exist. And, like, that's where it... That's where we've kinda sat with, like, if there's information in Notion and things like that. So, like, the more information you can put in the system or record, the more intelligent your agents are gonna be in terms of reading it. You know, obviously, I have the agent point point towards Gong. It looks at Gong. It looks at Gmail. Like, it looks at the market, and it goes to LinkedIn. So, like, it has this stuff, but, like, you can give it a ranking of, like, what it what it what it believes and what it can do because, ultimately, the problem with tools for all is like, okay. You have an AI summary for Gong that then puts an AI brief into Salesforce, that then my agent that's built on AI is gonna read that. So you have three AI summaries before it even get... And, like, things get lost in translation. So, like, what we've tried to do so much in Salesforce and why Dolly exists now is how can we really get down to, like, what we heard and what we saw and not AI summary. So, like, Dolly is supposed to look for, like, the transcript instead of the AI summary to get more, like, from the horse's mouth. Yep. But, yeah, I mean, we put you have to make Salesforce your system a record. But, like, ultimately, it... I mean, it's been great, I think, in terms of doing that. So anytime we make a tool decision, anytime we make anything, it... Like, it's like, what is the connection to Salesforce? Yeah. Yeah. Yeah. That... That's that's great. Now, obviously, you know, we... Could you talk about, like, what that looks like? Like, what does that structure look like? You know? Yeah. So, essentially, like, what's really, really great is, like, you know, most of... Especially with, like, NCP servers now, and things like that. Like, a lot of things can write back to Salesforce and things like that. So, like, Gong pushes everything into Salesforce. Right? Like, you know, Gmail pushes everything into Salesforce. And so Dolly also is doing a loop right now to make sure things are getting synced because, like, it is parsed. Like, sometimes stuff shows up in Gong, doesn't make it Salesforce. Sometimes you have Gmail, doesn't make it back to Salesforce. Like, we're always looking at that. So, like, coming back to, like, giving it that kind of weighted pipeline, it's like, look in Salesforce. Okay. If it's not in Salesforce, go look in Gmail because that's more direct email. If it's not there, go look and Gong. Right? So, like, you. almost give. it this hierarchy of where it finds information, and we try to get to, like, the source of the the most accurate. And then, like, if you have to, go ask person than what. it is. And so, yeah, I think that's really good. So, that's been great. Yeah. Well well, thank you. So we'll... In a minute, we'll talk about some of the outcomes that we've seen. But, I I know the story well, and, there was a surprise along the way. Right? You know, we we went in thinking one thing, but I think, this experience kinda changed that. Right? Yeah. Absolutely. I mean, like, I think when we looked at... Really what I went into, I was like, you know you know, looking, you know, obviously, like, our CEO and our CTO, like, the things they always were talking to me is, like, our product is much more valuable than that. It's much more valuable than that. It's more for, like, we're we're selling it too low. The reps, like, they don't understand the product. They're not pitching value. Like, it was like that. And so, you know, once we had an agent kind of being, like, I guess, I wouldn't say... I mean, an agent is still subjective. Like, you know, it is... Like, it's, you know, it still has bias in what it's giving you, but, like, trying to remove that. bias. Like, it started picking up patterns of, like, when we lost and when we won and what combination of components or features or things. And so, like, it really started to pick up also where, like, you see a lot of the market shifting right now. Like, you know, I think a lot of people the past three or four years have tried to go, we're gonna go up market. We're gonna sell to the enterprise. We're gonna build these big packages. But, like, the way AI and token usage and things like that have really kind of exposed us. Like, it seems the market's now pushing more towards, like, low entry point and. then adoption and credit usage. I mean, like, you know, Salesforce, everyone knows is, like, we're we're waiting for the bills for, like, MCP usage. Gong has done this. Like, a lot of these tools are like, oh, by the way, you know, we might be moving away from a per seat basis to just a token credit. So, like, Dealey was able to kinda get along with that and and kinda make some suggestions with that weekly audit. Like... And so, like, it was just a prompt. Yeah. It was just changing the prompt. Hey, Dealey. On Sundays, when I'm present... When I'm getting ready for the forecast, do this, this, this, this. By the way, make one recommendation on what we think we should change so I can put it on the agenda for tomorrow. And that was really, really helpful. So... Apps... Yeah. And and I see we're getting some love and some reactions on. that one. That that right there, the quality I've noticed, those forecast meetings just becomes so much more, succinct and clear and focused when you... When everyone in in in the room comes in with that type of information. It's a game. changer. Yep. You have to be careful because, like, I think, like, pre AI, our forecast agenda was much shorter. Now it's like everyone does their huge roll up with everything, and so, like, you know, the access to information is so great, but, like, being able to distill it down. And so, like, I I mean, you know, we'll talk a little bit about, like, daily sprawl of agents and what can happen, but, like, daily sends me in, like, an 8AM. ping. I'm like, here are the five most important things you need to do as a CRO today. And then I went, hey. Can you tell all my managers this? Or what about the ICs? Yep. And so, like, it almost like they wake up 8AM. They get a Slack ping that stands in. It's like, hey. From Dealey, it's like, hey. Here are the five most important things you need to do today. And it. it ranges. It ranges like, oh, you need to do your forecast. Hey. I need to I need to get feedback to that one page to Greg. He's been pinging you twice. And so that's you, that's really the way. helpful. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. So Excellent. Alright. Well, look, you know, look, we've got a lot of people here today, and the the concept of building agents is obviously, you know, has become pretty commonplace. Right? But, you know, it almost never fails when I talk to my peers at other companies. I I find that it's kind of rare that they can actually tell me about the impact. Like, what is what what is the impact that the business outcome on a day to day basis been since you incorporated, you know, digital workflows. And so I I love the fact that you're you're gonna walk us through real quick, you know, the impact that Jilly's made. Right? So what what does that look like? What's it been over the last year? I mean, it's been dramatic for me specifically, and also for the reps. So, you know, deal... I asked deal... I asked deal... I need to take a look at it for for this calendar year, and so we had 629 deal desk requests from 2026. That's new logo, upsell, renewals, everything like that. And so, we've actually made it much more stricter on the reps that everything needs to go to the deal desk. That was the trade off. Right? Hey. We had some preexisting things that you guys could approve. Now we want all the information. We want all that information. But the trade off is you get almost real time feedback. And so the pricing review time is pretty instantaneous. Like, the only thing that kicks that up to an hour is when it gets flagged that I have to get involved. And so, like, our promise was always during a single business day. So what did I usually do? I'd go home, hang out with family, put the kids to bed, and then I get back on and I crank for deal desk. But the worst. part is, like, what if I have questions? Okay. Now we're back the next day. Okay. What if I have more questions? We're back the next day. So pricing would take three and four days. So, like, the review time now immediately goes like, okay. Approved. This is what's approved. Or, by the way, you need to set up a meeting with with Rich tomorrow. We discuss this. What's great about agents is what we just launched this week was now if Deely thinks there needs to be review since he accessed the Google Calendar, he go ahead and sets up the pricing review meeting for the next day, and and tells the rep what it's supposed to prep. So, like, we've even automated the meeting so it doesn't go to the next day. The escalation's easy. Right? Everything was escalated to me. I was the gatekeeper. Now I think I'm reviewing less than 10% of the deals, which is great. And then the pricing policy extension, it's actually crazy because now that they know that agents looking at a rule book and their commission plans and an audit, the... I don't know. I... My... A brilliant CRO that I worked for named Chris Wilson gave me one of the best quotes I ever had for being a a sales leader. Reps will only do what they know you're inspecting, not what you expect them to do. And so me inspecting stuff, I'm one person, but an agent is now expecting what they price, what it was closed. They can go look at the Gong thing. And so, like, the amount of, like, us staying close to the rule book of what we wanna price and actually run pricing experiments has been so much better because an agent flags variables pretty quickly. And all we have to do is say, hey. Dealie's looking at this every thirty days to see if you guys went outside your budget range, then we wrote it into the comp plan, and it was like, okay. Well, that just solved that. So I think that was... We get cleaner, better data, and, like, an agent is able to inspect. Now we still... We pick and choose our battles on when we're gonna do it, but, like, an agent's able to inspect kind of that body of work of, like, hey. It seems like so and so keeps going under their discount threshold, which why is their win rate higher than everybody, which on the scoreboard, you're only like, wow. Like, what's John doing? His win rate's amazing. Yeah. And then dealer goes, yeah. Because he's 30% under the price book and you just haven't seen it yet. I was like, oh, okay. That's... Thanks. Oh, Appreciate it. So... Yeah. yawn. Yeah. Yeah. So... That's great. The... These numbers are... I mean, they're pretty staggering. Right? I mean, Yeah. the... It speaks for it speaks for itself. So we we we we... Yeah. We started off talking about Deely, and then Deely got a sub agent, Dolly. And so, you know, before we wrap up, you know, I wanna hear about what you have planned next. Does does Dealie get another sibling? Or or, like, is there something else that you're you're looking at, you know, in terms of, like, leveraging an agent in day to day life? Yeah. I think the, the big thing that you have with agents is the more information you give them, the more cash you give them, the more generalized they become. Right? And. so it's the same thing as growing an org. Right? I mean, like, you have, you know, you have AEs tackling new logos. You have AMs for upsell expansion. You have CSMs for renewals and relationships. Like, those are specialized things, but then you have managers that do that, and then you have manager manager. Then you have a CRO. Like, you know, like, what you lose is that specialization. So Delia has become this massive, like, agent that can do everything. But what happens is, like, it takes him longer now because he has to crank through more information. He's more generalized. He gets confused more, right, because he's asked to do so many things. So, like, one of the things is Dolly came up as the first sub agent. So what we'll typically see and what I'll work on is I'll start building sub agents for specific task. And so, like, when those tasks come in, deal... Like, hey, Dolly. Can you do this? Hey, Frank. Can you do this? We'll probably... We'll keep it with the McDeal fit... The McDeal face family, but, like, definitely. But, like, ultimately, the sub agents allow you to do so much more, and it works quicker because I think it... You can hit a big... You go to cloud code action, like, hey. Run this evaluation from you. It's spinning off sub agents to go quicker. And also token usage will go down if you use sub agents because they don't have to crawl through so much stuff. So I know we did. you mentioned that. Yeah. That's that's a great point. Absolutely. Yeah. And I know everybody on call is thinking it too, so I'm. glad you called that out. That's great. Alright. Well, you know, up to this point, everything we've talked about has been around, you know, how we solve the set of problems, you know, internally with the internal tool. But, I'd be remiss if I didn't take a minute just to show you real quick how Nutrien can help organizations like yours with your own challenges. So at Nutrien, our work spans the entire document workflow. You know, we're talking about extracting and structuring data. We're talking about generating and converting documents, governing and routing work flows, and and even supporting reviews and approvals. So on top of that, we deploy those solutions wherever you need it. So whether you're talking about in the cloud or self hosted or air gap, we've got you covered. Now, you know... And while... Look. Nutrients, we we represent the entire document later... Layer to your workflow. You know, I do wanna, however, put a a spotlight on one of our tailored solutions because, you know, this audience is very, very relevant. And we also believe that, you know, you shouldn't have to settle for multiple disconnected tools to to manage document related tasks within Salesforce. Right? So documents for Salesforce is a fully native solution, you know, that allows you to create, edit, redact, even sign documents without ever leaving Salesforce. There's no third party tools required, no downloads, no switching tabs, just everything you need to work with documents. But like I said, that's just one of our solutions. Over the last fifteen years, Nutrient has established itself as the industry's leading document platform for both agents and humans. We provide the document processing, governance, and human interfaces that enterprises like you need to move faster and to lower operating costs. And today, more than 3,000 companies worldwide trust Nutrient Technology in their document workflows. If you'd like to learn more about how we can help your teams, you can always find us at nutrien.io. Alright. Well, that moves us to our q and a. So let's see here. We do have some coming in. Christine, thanks for coming back out here on stage. Let's see. Looks like the first one is, do you do you use any other agents within your org? If so, can you share what they do? Yeah. I can take this. We have a couple different agents. You know, Daily is obviously the one that's... People use the most for the go to market. We have a agnostic company agent that's there for, like, anyone to use. So it's it's it's a Slack handle. We call it Nuri, but which is, like, I think it was short for Nutrien. It was, like, a offhand, but, like, Nuri is just for a general thing and that anyone can access to. And so it has access to a lot of things. So people ask a lot of questions, customer questions, and things like that. And so, like, if engineering wants to know about something, they can be like, hey. What are the top five customers that are doing this? So we have, like, a a generalist agent, which, you know, I tried to do that generalist agent generalist agent first here and realized it wasn't specific enough. It was just... It's too much information, which is why Dealey came off. Then we have Coach as an agent. It's for enablement and onboarding, which is awesome. So, basically, it does daily onboarding sessions with the reps via Slack, and so it's like a five to ten minute, like, today's your quick... You know, I think we adopted something in terms of, like, one of the language, applications that help you learn, like, a new language where, like, here's your daily thing that you just gotta knock out. So we have Coach, which is an enablement agent that basically helps people with onboarding and things like that. It can run role plays, and then deal and all its sub agents. So those are the three big ones that are running. I think, you know, I think marketing and all the other teams are building sub agents as well for their own, their own usage. We also have a support agent as well that we're we're considering rolling out for first level support. So, like, yeah, we've been we've been there quite a bit. Yeah. Yeah. And I would even chime in that, the product marketing team, we are, very close to rolling out our own agent for the organization. It's gonna live in Slack, where it gives regular updates from a competitive standpoint, from a release standpoint. Really all the the main functionality, you know, that you come to expect with with product marketing. We're gonna put that in the product marketing channel. So, we... You know, Nutrien, obviously, with our pedigree, every employee is a builder. And so that is... That's just a given. And so we're always looking for ways to improve, you know, our efficiencies. So... Alright. Great. It's a great answer. Alright. Oh, this is a good one. Do you have any horror stories with Dilly? Yeah. We've lost deals. Like, it's been bad. Like, I mean, like, it is, like, it is hallucinated, and we've quoted way too big and way too low. And so I think, I forget. I don't know how it got through. I think someone came in through with a Gmail email from a contact sales form, and it never recognized the company and the adoption. And so it flagged it as like a SMB startup package. And then, like, all of a sudden, we have an error horror channel on Slack, and it's like, Fortune 100 company closes for $700 a year. We're like, not great. Oh. So we had to explain that one to our CEO. So, like, we do have some horror stories where, like, it does hallucinate. It does things like that, and. it's gotten things wrong. Also, like, I'm sure we've all logged the wrong email to Salesforce before, like, accidentally. So, like, if the data is not there or it logs a call to the wrong thing, like, we, like, we work with a lot of companies that... Like, we work with DocuSign as a customer, but DocuSign comes up a ton in, like, email threads. So, like, it really struggles to, like, delineate, like, am I talking to DocuSign or customer, or are you just saying we're sending a DocuSign envelope? And so, like, there has been some stuff where it comes back just wildly confused. And so, like, building some sort of catch all to be like, hey. Just make sure you understand this or ask clarifying questions. That's that's been some horror stories. But we have definitely lost some deals, and we've won some deals at prices that got me in a little bit of trouble. But... Yeah. So... Yeah. Well, you learn from those, and the governance gets better after those. Right? Yeah. Yeah. Yeah. Excellent. Let's see. If you could... Okay. If you could do it all over again, what would you change? What would I change? Probably would've just done it quicker. Like, I think... I mean, like, I mean, some of this was waiting for the technology to catch up too. You know, I see some. questions in the q and a about Cloudforce and, things like that, but I wish I would have got to it quicker, but some of it was waiting for agents to become a thing. I think we were going with the technology. I also did it in my spare time where, like, I wish I would have, like, just taken a week and built it and then did it. I didn't. So, like, it slowly came over and over and over again. And at one point when you're like, you have this GPT that's working so wonderfully for you, you're like, oh, I do not wanna go and rebuild it. Oh, I don't wanna. go and rebuild it. So, yeah, just probably should have done it quicker. Yeah. Yeah. I see, by the way, I see a popular con... I I see a question in here that, we kinda joked about earlier. Let's see. Tim asked, are you are you paying for API tokens, or are you using flex credit tokens? It's a great question. So there's two things we're doing, to try and offset token usage. Right? One thing is, like, we had a... We built a CLI that basically tells the agents to do certain, like, certain, models for certain tasks. So, like, you can tell the agent because you don't wanna be operating at the most expensive model for everything. Some things is retrieval, some things is just reading a transcript. So, like, one of the things we did is we built... We used a CLI to basically say for certain tasks, use these models which saves token usage. Right? Like... So, like, hey. If you're just reading a transcript, you do not need the best. model. Like, just go to DNS. But for the thinking, it does. And so, like, that was something that we we worked with pretty heavily, and we've used it for all ages. That really helps with token usage. Also, like, the more you can balance putting stuff that it remembers and puts it in local memory, the more... The less it has to retrieve. Also, like, when you go look at Salesforce, that is a very broad statement. Right? There's so much you can go look at. It might be looking at the account, 14 different opportunities, and so, like, pointing it more directionally to look at was it... And what was great is because when you launch the deal that's from Salesforce, there's the ID that comes with it. And so, like, just telling the agent, like, make sure you're looking the opportunity that matches this ID. Look at these deals specifically. Ignore all the CS stuff because we're not there yet. And so, like, based on the opportunity type, we told it to look at certain things and, like, that just, like, these little mix of, like, save cost here, save cost here, like, that's... That really adds up over time when you build it at scale. But surprisingly, yeah, it's it's going really well. Yeah. So a question here. I see a few people, and, and I especially wanna call out Cesar who's been very, very active with questions. That's awesome. I love all of these. And... But one of them seeing come up multiple times is around, like, data grooming. Right? Did you do any. cleaning? Did you... What... You know, so his question is, do you have, have to do any data or content cleanup before connecting to the agent? I think we. know the. answer. So go. ahead. Pre Dolly, yes. Because Dolly now catches it. Right? And she's trying to, like, make sure it's that. But, like, I essentially had to walk into our CSOs... CEOs off and said, hey. I wanna hit enter on a dating groom, basically, audit for our Salesforce for the past three years and basically flag everything, and I need to give it read only permit. I need to give it write permissions. And so we tested it in a sandbox for some... Like, clone some opportunities, put them in a sandbox, and see what the results were. But, ultimately, like, you know, the the thing with AI, y'all, is, like, you can also ask it how much you think it's going to cost. Right? And so that's where we kinda got into, like, how can we... This... That whole exercise is where we went and try to figure out how can we make this cheaper. That's the best thing about, like, agents and stuff. Like, you can have a conversation with us. Like, how can you become cheaper for me? Like, how can you be better? And so, like... And so it gave us a lot of the ideas to cut down cost, which I don't know if Anthropic is needing to do this. Like, I'm sure they want us to use more tokens, but, like, it's a cheat code. But, like, ultimately, like, I was able to say, hey. Hey, Jonathan, who's our CEO. I was like, I need $1,700 to do a data grooming thing. But he's like, yeah. That's fine. And so, like... And basically that did that. And so we did it in a sandbox on about 10 opportunities. I liked what I saw. They were closed one opportunities. The response I got from our CEO was like, I don't trust that anyways, so we might as well just update it. So I think that was the kind of the thing because it was all AI generated for, like, two years from Gong that, like, it was so, like, generalistic that, like, we liked it. It was much better. But, yeah, we hit enter and let it run for a while, and it it auto updated, which was good, which inherently made the agent more more... But the the way around that, Greg, is obviously, like, you give it multiple sources. So, like, Yeah. Salesforce is the ultimate truth, but please check the transcript for what the person. actually said. By the way, please check Gmail as well. So, like, you give it almost a thing. And then, like, ask a question if you're not sure. Like like, children. Like, you know what. I mean? Like, go ask mom, go ask. dad, go ask your brother and sister, and then ask questions if you still don't understand. It's like children. Yeah. Exactly. Alright. We have time for maybe one or two more. Okay. Cesar asked about, about time to value. He was asking, did it take, you know, a year to build and test before sharing widely or, you know, basically, or was it shared before that? It was immediately valuable to me. I mean, I built this sitting at a... I built this in, like, at a lunch at an SKL, like, two years ago. Like, I think it was just, like, I got an idea and we did it, but it was immediately valuable to me. It helped me move quicker. And so, like, I think, you know, if if we do this again, Greg, I'd probably, like, give the time the value based on the difference. It's like a GPT versus Claude versus a full blown agent. Like, you can see where. it really does. And so, like, it definitely was helpful there. But the time to value for me was immediately, and so like that. Yep. So, but what was fun is, like, I had daily run our SKO sessions, back in February in Charleston. So, like, some of the workshops and exercises, like, he was able to run, like, Jeopardy and run things like that. And so it was cool as everyone's in Slack, like, playing Jeopardy or they're playing, like, a use case guessing game. And so, like, it really... Oh, like, directly, yes, for the deal desk, but, like, him having that knowledge, I was able to do a lot. And so, like, we just onboarded a new enterprise rep, and the first thing we said is go download Dealy. And, like, he self onboarded himself in, like, two to three weeks. And that, Yeah. like... Okay. That's like... There's so many side effects that you don't even know that, like, you walk to me all the time in the office like, hey. Could you do this? I'm like, I don't know. Ask Dealy. Fun fact, Dealey built the demo today that we showed early. We went and said, hey. Can you build a demo of yourself? And then I think Claudia came back. He's like, I don't really like the subtitles. And so yesterday, I was like, I don't know. I'll ask Dealey to remove the subtitles. And so we did that. So, like, there's so many side effects of these agents when they have this history and knowledge that it it it it seeps into other things. Like, the time to value, I think, is almost immediate. Yes. Yeah. Yeah. It's, it's really amazing to see how it's evolved over over time. That that is for sure. Alright. Then let's see. The next question, could you do this with Claude now, a k a Claude Force? Yeah. Definitely. I think the thing with Claude and things like that is, like, what you're ultimately going to do is, like, what is the interface you want people to interact with? We use Slack, obviously, because Slack Slack's owned by Salesforce. Right? So, I mean, like, there's a lot of synergies already there, what you can do. And Slack's our communication tool, so it's an easier interface, I think, for them to do that. Claude, you have, like, to download an agent then to do all the connectors and stuff. It's a little bit more com... Cumbersome and, like, there's more there's more on the front end for the people you're giving it to to do it. So, like, I think what a lot of people end up resorting to is you build a skill and you do a skill and... But, like, every time you update the skill, everyone... And so, like, I have found that, like, the chat interfaces are more comfortable to people because I think most people are used to using AI in a chat interface. I mean, like, a lot of people are still using ChattyPT just to rewrite emails. So, like, Slack has been the best interface from us to just be, like, at dealy, at dolly, at neri, at coach. And it's also in public channels, so everyone learns from it. We're like, someone comes to me and is like, hey. Deli is messing up in Claude. I was like, alright. Bring me your bring me your computer and things like that. So it's just much easier to diagnose or, like, something goes wrong. Or I can jump into Slack and say, Deli, you're wrong. Blah blah blah blah blah. Or when I get tagged from Delia and Slack, it's much easier for me to come into the thread. Right? Hey. This got flagged. You need to talk to Rich. Meeting set up for tomorrow at twelve, and I get a Slack notification. So that's my preference that we do. it, and Slack makes it real kind of easy to build these interfaces now. So... Yeah. Yeah. And and, honestly, you you touched on something earlier that I just wanna echo again because this is this is sort of an under spoken, benefit that came out of this, and that is the enablement, the onboarding. It is been a game changer in terms of product knowledge, in terms of organizational structure, in terms of, like, how we work deals. It... It's just... Because we have, you know, multiple agents now that are just given to new employees, and their ramping is just... Is shrunk dramatically just since we started doing that. And it's it's something we're gonna continue to build out and improve because the payoff has just been immense. So... Well, that looks like... Christine, is that... I think that might get us all the questions there. Yeah. I think that takes us pretty much up to time. It's been so nice to see so much engagement and great questions and, honestly, such a brilliant session. It's so nice to hear from people who have actually used AI instead of it being the art of the possible, someone who's actually done it and gone through the trenches and, iterated. And it's so fascinating to see those statistics about how it's making change and the improvements and the adoption that you've experienced. So thank you so much for sharing that with with us, with the audience. But, yeah, Yeah. thank you very much. to to you both, to, especially especially Richard from Yes. Yes. Yeah. No. It... This was awesome. And, yes, thank you, Richard. Thank you, Christine, for having us. We we enjoyed, you know, obviously, sharing, If, you know, if this can help, even just a single person on the call today that, you know, then that's a good thing. So, yeah. Thank you everyone again. On behalf of Richard and Nutrient and and and Salesforce, Ben, thanks a lot, and we'll see you soon.