How AI Helps Sales Teams Actually Work More Leads

This episode breaks down how an AI SDR agent built in Salesforce Agentforce helps sales teams handle more leads, improve outreach quality, and remove bottlenecks at the top of the funnel.

Eric Housh and Zack Terry walk through the technical design, including grounding strategies, brand voice control, observability, and prompt evaluation, along with key insights on AI performance and scalability.

Prefer to watch? Find the full episode on our website: https://loom.ly/C3j2IuU

Transcript

SPEAKER_00 0:09

AI in action is brought to you by Fast Slow Motion. Our team helps growing businesses put AI to work with practical, scalable solutions. To learn more about how we can help you implement AI in your business, visit FastSlow Motion.com/slash AI. Hey, and welcome back to AI in action. I'm Eric Hush, joined as always by the one and only Zach Terry, Director of AI at Fast Slow Motion. Zach, how are we doing today, man? So good. Just, you know, as one of my favorite colleagues would say, never better. Never better. It's always a good day when you can talk about wins that we're able to deliver for our clients, leveraging uh Salesforce and HubSpot AI solutions. This one's a fun one today because we run into this, we see this a lot. Company has a great product, company has a solid list of leads, but for whatever reason, their outreach just grinds to a halt. Yeah, maybe it's a restructure or the cost of having a human run that SDR outreach is just too expensive for the results they're getting. But the result is there's this massive gap of potential money, revenue just sitting there gathering dust.

SPEAKER_01 1:16

For sure. And I first of all, I mean, having too many leads, that's a good problem to have. I think it's much better than the alternative, right? But the issue is, of course, the volume, as that volume grows, you start to run out of resources to effectively reach out. And so typically, historically, what does that mean? It means you're either you're hiring more SDRs to come in and start reaching out at the top of the funnel. Or if you're not able to hire anyone, your team's starting to get spread pretty thin because they're having to use the same resources that they have across a much bigger volume, right? So that's sort of the baseline that we're working with here, right? And so this particular client that we worked with, they're actually an AI sales company themselves, which is really meta and kind of fun. Yeah. So you may be asking, well, why are they implementing an AI sales tool in Salesforce? And the answer is because they serve very distinct purposes. So what we built for them is a top-of-funnel tool to help with volume, what they provide is meant to address an entirely different part of the sales cycle, something that's farther into the funnel, mid-funnel or later on, right? And so uh their tool was not suited to do this exact problem, or they would have had to reconfigure it. It would have changed their positioning and their marketing. And uh ultimately they decided that that they wanted to go with Salesforce and an agent force. It's it's a lead nurturing agent. It used to be called an SDR agent, a little bit of a name change there. But that's ultimately the solution that we arrived at, which was creating this lead nurturing agent that could handle some of the volume at the top of the funnel here. So at this point, when they came to us, they had pulled back on their outbound team. And so they were not only dealing with this volume, but they had actually had fewer, had fewer people than they did before to deal with that volume. And so they really needed something that could help fill that gap. Um, because they are an AI company. If they also use something that was basic or generic or did not give good responses, if they used an AI tool, that would also reflect poorly on their product. And so we had to make sure that we built something that provided good quality messaging and was actually performant for them, because you know, you wouldn't want a company's CEO to show up in a taxi because his own car broke down, right? If if a car company CEO, you know, imagine like the CEO Ferrari showed up in a taxi to a meeting, that would look a little bit weird, right? And so we want to make sure that we are creating something that is a quality experience that you would expect to get, even though it's a different product than what they're actually selling, right? So they needed these SDR representatives at the top of the funnel. It needed to be as smart as the product that they were selling. It needed to have a really good outward face and present itself really well, designed to address the volume and help the dwindling BDR team address all these new leads that they had in the system.

SPEAKER_00 4:06

So you talk to this, but in terms of that before state in the olden days, right? The alternative to humans was just blasting a bunch of automated emails. It sounds like there was a little bit of brand risk here that made that not such a great idea. Was there anything else going on in terms of driving us toward an agentic solution versus the alternative?

SPEAKER_01 4:26

A few specific things. One, they like like we had mentioned before, they didn't really have a lot of people able to work the top of funnel. They had reduced their team there. So they were working with limited resources and a growing volume of leads to reach out to. That also meant that they probably weren't getting a lot of good market signals in because they weren't able to follow up with all these leads and start to get an understanding of what's working and what's not working. So they were essentially flying blind. And then they had a unique brand risk. As we discussed, they are an AI company themselves. And we didn't want to risk brand confusion between this outreach tool and the actual tool that they sold themselves. And so if that outreach tool sounded a little bit too rote, a little bit too robotic, people may assume that their actual software that they sold would sound mediocre too. And so we had to make sure that we created something that was uh really responsive to their tone and their messaging and reference their products and their services in the right way. Uh, and then of course, the the last thing is they they had a timeline. They had a deadline that they had to hit. They had an event that they wanted to make sure they were mentioning in this initial outreach for this agent, the big industry event, and they didn't have time to hire and train a new team to handle that. And so not only did they need something to cover the volume and have consistent messaging, but they also had a deadline that they had to hit really quickly so that they could get this messaging out to uh all of those prospects.

SPEAKER_00 5:52

So, from a technical standpoint, let's get into the nuts and bolts here. Uh, we're building this solution inside Salesforce using uh agent force solutions. Walk me through that grounding process. I think when a lot of people hear AI, they're probably thinking of Chat GPT and there's a risk of hallucinations and making things up. How do you stop?

SPEAKER_01 6:12

That is definitely a valid concern. And we've we've talked about this, I think, several times on this podcast. There, there's always a risk of hallucination, right? But there are tactics that you can take to minimize that risk. And so specifically, what we did with this AI SDR lead nurturing agent is we grounded the agent in their business knowledge and established a retrieval augmented generation pipeline. So you've heard us talk about this before. We've talked about this in the context of a service agent referencing knowledge articles, answering questions for customers, kind of doing that, you know, tier one customer support, deflecting service cases from getting created. I think we've also talked about this in terms of maybe an internal employee assistant that's able to answer questions from knowledge. This is kind of unique because what it's doing is it's it's not really grounding its first response in knowledge. The first outreach is really kind of grounded in the specific information about the lead and the messaging that they want to send. But what happens if that lead comes back with a question? That agent needs to be able to answer that question or it needs to be able to identify when it should escalate that to a human being. And so in this scenario, it has access to a database of information, structured and unstructured data that's in Data Cloud, and it can go and look up question or look up answers to questions in some of those PDF files that they have. It can answer questions about the products and services that they have. Let's say somebody wants to know if it can meet a certain criteria, then it's able to go in and reference those documents in order to provide an answer directly to the customer. And so that's, I think, is a big differentiator because it doesn't require the conversation to then be escalated or reassigned to a human being in order to answer a question from a prospect. It can be a seamless response and then it can attempt to uh push that person either towards booking a meeting or towards, in this case, and this was sort of a unique change that we made in the default behavior of this agent, they they wanted to send a specific URL. And so we customize the default behavior of this agent to instead of attempt to book a meeting with a human sales rep, we have a bunch of URLs that are stored on the prospect record in Salesforce and it's unique to each prospect. And so we're able to have that agent go and find that information from the CRM and inject that into the email to ensure that the right person is getting the right link at the right time. So that's another sort of interesting technical customization that we made. We also implemented uh what I'm calling a three reply rule. And what the client really wanted here, they didn't want an ongoing, endless back and forth conversation with this AI agent. We wanted to make sure that if there was, let's say, an ongoing conversation, that after three replies, the AI agent should be smart enough to understand that they probably need to get a human involved. And so after three replies have occurred, then the AI agent understands that it should then go and get a human being involved and reassign that conversation to a person. So it hands that off to a human. And then, of course, all of that we we need to know how it's working, right? Uh we can't be blind in terms of, hey, how are these agents answering questions and are they answering questions in the right way? And so, of course, we set up the observability stack, agent force observability and optimization. There's a whole supervisor view as well that allows uh the human sales reps, the supervisors to be able to go and look at what those agents are doing and how many conversations they're managing and whether those conversations are active or just started or they've already finished and what the result is. And so, of course, we wanted to make sure that we had a lot of reportability there so that they could not only assess the effectiveness of it, but also proactively go and and make changes, refine how that agent is working. Maybe that's tweaking some of the instructions or it's making sure it has the right context. But ultimately, it's it's about being able to understand when it's missing the mark and when it's hitting the mark so that you can continue to do the things that allow it to answer questions correctly and then make the tweaks you need if there's any gaps.

SPEAKER_00 10:13

Let's uh let's double-click on a couple of those because those sound like pretty wise design choices that that may be carrying a little bit more weight than than people realize at first. Start with the the what'd you call it, the three reply rule. Why why did we land on three there?

SPEAKER_01 10:27

So it's really about reading the intent of the prospect. So by the third reply, and this is business logic for for this client, right? But the prospect isn't really just clicking through something. They're they're engaged, they're having that conversation. And so at that point, you you really want a human to come and take over, someone who can actually move that conversation somewhere. We didn't want that agent to get pulled into an endless long exchange. You know, obviously the longer these conversations go on, the greater the probability of some kind of drift in that agent's behavior. And so we wanted to put some guardrails there to help minimize that potential. And it keeps it grounded as well. And so uh just from a practical matter, three turns has kind of a natural rhythm to it. And it doesn't really feel like a door is closing. It just feels kind of like, okay, we had this conversation. I need to get some more information, and we can get a human involved to come in and have a higher touch. So it really kind of came from the client's requirements, but I think they were grounded in uh really good and sound logic in terms of why they might want the AI to hand that conversation off after three replies.

SPEAKER_00 11:28

And go into the URL logic, what was happening there? So, based on the prospect record in Salesforce or the lead or the contact record, it's pulling different URLs. Why was that design made?

SPEAKER_01 11:40

So, this particular event, like many events that I'm sure you've been to, there's there's different sessions and different tracks. And so in instead of just one event where everyone is attending the same things, there are separate presentations going on at the same time, right? Something like Dreamforce, right? When you go to Dreamforce, there's a little overwhelming, yeah. So overwhelming, right? And this this wasn't that, but it was it was a large event with a lot of different tracks going on in that conference. And so um, for them, we wanted to make sure that we were sending them to the right event page for whatever persona they belong to so that we can market the correct event to them. So, yes, it's one event, everyone's gathering, but we want to make sure that they're getting the right materials for whichever track is gonna make the most sense or that we think is going to resonate most with what they might be interested in. And so uh we basically saw that they were storing some of that persona information on the lead, which is exactly what they should do. And so then we came up with a little mechanism to um identify those personas, create URLs specific to each one of those tracks that are aligned to those personas. And then that way it's really easy for us to just merge in the URL from the record into the generated email. So the AI is still generating the messaging around that URL, and that's specific to attributes about that lead, about that persona. And just as a part of that, we also have a unique link that we can send. And I say unique, it's not unique to the person, but it is unique to their persona. So let's imagine you have three personas, just you know, to keep it simple, and I have three different tracks at that event. I want to make sure that if somebody is, let's say, interested in, you know, let's say it's a CRM conference and I've got Salesforce on one side and I've got HubSpot on the other side. Well, I don't want to market the HubSpot track to the Salesforce customer, right? That's really the idea. We want to make sure that we are directing them to the most appropriate and most of interest to that prospect, whatever sort of, you know, information at that event we think they're gonna resonate with.

SPEAKER_00 13:40

It's funny when you said these events that could have multiple tracks, my mind immediately went to Dreamforce. Like, hey, maybe, maybe they need a solution like this. Yeah. Exactly. Exactly. Another thing you sound uh you said that I found interesting was their brand voice because this guy, this company sells AI. They can't afford to sound like a generic bot. That could cost them sk uh sales. What sort of prep goes into making sure that you get that you nail that brand voice and what do they do there?

SPEAKER_01 14:07

What's important here is they had already done a lot of that work. We talk all the time about establishing a foundation, having good defined processes, having good data. So they had already done a lot of work to understand, hey, this is who we are. This is how we talk about these things, this is what our brand voice sounds like, this is what our style looks like, that kind of work, right? And so in this case, we were able to take that documentation and ground the agent's responses in that documentation. And so when it's generating emails, it's essentially looking at these guidelines in order to determine how it is going to write the wording and the tone and the messaging behind those emails. And because they had already done the work to define it, we didn't have to spend weeks going through and figuring out what that actually should be. We were just able to adapt it, really as simple as uploading a few files and directing the, you know, the configuration and the plumbing for the AI and how it works and how it gets that information to reference that documentation with some specific instructions on, hey, here's where you're looking for that information and here's how to interpret it, and here's how to incorporate that into your messaging. Uh but I think I think that's the the the biggest aspect of that, which is they were ready, right? They were at the right place in time to have an AI come in and replicate that brand voice. Now, is it gonna do it perfectly every time? No. Is it something that you're gonna have to monitor and refine and tweak over time? Absolutely. But that's true with sort of any AI implementation, any technology implementation. It's always gonna require some tweaks. But the but the main thing here is they were ready to go. So we could hit the ground running really quickly and get good results uh without having to spend weeks defining what that should be before we're able to connect an AI agent to it.

SPEAKER_00 15:49

Yeah, having that documentation always critical. Uh, Zach, whenever we publish one of these episodes, I get contacted by a handful of folks that say, hey, that sounds like that that could work for me. Uh, what do I need to be thinking about? What are the prerequisites that make a solution like that happen, that make it possible? We talked about the brand voice, we talked about a couple of other things. Anything we haven't hit on yet that companies that are interested in pursuing a genet solutions like this need to be thinking about?

SPEAKER_01 16:16

I'd say there's a few things that you probably can't just manufacture once you're in the build. Um, this particular client, I mean, they already had their ICPs defined. It wasn't here's some companies that we think might be interested. It was real criteria around this is who this persona is, and this is how our brand resonates with that persona, and this is the messaging that we're gonna send. And so um, having that defined before trying to automate it and not just automate it, but also connect AI to it, where it's gonna be deriving and generating new content based off of that, having that well defined and written down and agreed upon before you have AI start generating content off of it is gonna get you a much better result than if you're starting from ground zero. Now, that doesn't mean that you you can't get started with a tool like this if you don't have that defined, but um, that's going to be a prerequisite to getting started because uh you're gonna get some pretty variable, maybe not great results if you're trying to generate personalized, targeted messaging and you haven't provided information about how you actually think about and craft that messaging. So that that's something that I think was really important here. And that's important, yes, for this specific instance where I'm I'm creating sales communication and I'm trying to hit prospects in a certain way at a certain time. But um, that concept applies to you know any sort of automation, a process or any solution that's going to embed AI into it, defining the processes, making sure that you have the data to support it, all of that is super important to the end result. Can you use AI without it? Absolutely. But as we've said before, it's going to reflect the state of that data back to you and then multiply it across those surfaces. And so it's really important that you you kind of have that stuff figured out beforehand, or you're able to have a conversation and arrive at whatever that process is going to be before you start to try to automate it.

SPEAKER_00 18:12

And is that process, is it as simple as opening up a Google Doc and saying, like, how do we talk to CFOs? How do we talk to COOs? How do we talk to CTOs? How does my best salesperson do this? Is it is it as easy as that that we just need to stop and write it down?

SPEAKER_01 18:28

Sometimes it can be. And and what's interesting is most companies have this stuff somewhere, but it could be a very senior sales leader who has just done it this way for years and they have the process down, but it's not written down. And everyone else has learned from them and kind of follows them, but we haven't really gotten it documented, right? One of the best ways to get this documented, and I have made this recommendation to a number of clients. I do this myself whenever I'm trying to get something out and documented. Open up an AI tool or use an AI dictation tool like Whisperflow. That's a great tool, but you don't even have to have an AI tool if you don't want to. All of these devices these days have dictation built in. I can open up my phone and start dictating, and it's going to transcribe when I'm writing your Mac or your Windows PC, probably has that same capability. Uh, but you can actually just describe it. You can just talk it out, right? Yeah, just talk about it. And it doesn't have to be perfect. You don't have to worry about formatting everything correctly, having complete sentences, any of that. Just talk about what you do. And then you can have an AI tool analyze it and start to put it into a format that makes sense. And my recommendation is not to just take that entire transcript and dump it in and take the first result that AI gives you back, it's to take all of that and use AI kind of back and forth with a conversation to start getting it formatted into the process that actually makes sense. And so there's a responsibility for you to read what is coming back, understand how it's formatted, make recommendations and tweaks and verify that the information is correct. But I can't think of a way to go from no documentation to very quickly having documentation. To me, this is kind of the fastest way to do that, right? Just talk about the process. AI can kind of analyze and synthesize it and get it formatted for you. And then maybe it's not the final version, but you've got 80, 90% of what you need at that point, and it makes it much quicker to actually get it documented from there.

SPEAKER_00 20:31

I love that. That's giving me ideas on this side of the microphone of ways we could implement similar solutions. Uh, let's talk quality of life. Once this thing's up and running, what actually changes for the sales team?

SPEAKER_01 20:42

Yeah. So ideally with with a tool like this where you're getting that volume of leads, hopefully you're continuing to get that volume of leads. But the main thing is it's no longer a bottleneck because now you have an autonomous system that can take on thousands of leads a day, right? If you had a thousand leads a day beforehand and only two SDRs, you're never going to tackle that. And you're going to have to figure out how to triage which ones you're going to prioritize. And so if even if you're going to reach out to a hundred a day, which is a lot, but I know there's some SDRs out there that do that, right? And more power to them. That's incredible. But a hundred out of a thousand with more coming in every day, I mean, that it's just, you're never going to get on top of it, right? So to have a tool like this that doesn't have those same constraints, it's able to understand who the lead is, get that context, go out, figure out, you know, in this case, which event they're going to be most likely to be interested in, crafting a message and then sending it at scale, all of a sudden that that bottleneck more or less goes away. Now, it doesn't mean that you you don't you can just kind of hand it off and ignore it. You've got to monitor it, you've got to make sure things are going well. You will find examples of perhaps some emails that that agent is writing that do not sound right, right? We've probably received those. I know I have, right? Where somebody took some old information off of LinkedIn and it has nothing to do with. Me anymore, right? But again, that's that's a data problem, right? You got to fix, fix that data problem first and make sure that that you have good data when you're doing that outreach. So um, yeah, I think that's the main thing, right? That that volume bottleneck sort of starts to disappear because you have a tool that can handle that autonomously when you implement something like a three-reply rule, which can be completely different from company to company. But in this case, you you have a little bit of filtering that can happen here. And so um that way you know when someone is extremely engaged and you can get somebody in the room with them really quickly so that you can do that triage, right? So that that problem where you had to figure out which 10 leads I'm gonna reach out to today, well, now the AI can reach out to a thousand of them and identify which 50 of them are most interested and assign them out to the right people. Um, so I think that's another sort of win that that you can get with this. Um, you know, also I think another way you can take this concept and supercharge it is, oh, when that gets handed off to a human sales rep, you should probably have an AI note-taker tool that's gathering that transcript. And so then all of a sudden you're able to take that information and that's context that you feed back into the CRM. And then maybe you have another agent that can reach back out at some point in the future if you know the timing isn't right now, but six months down the road. So you can kind of see how this all feeds into itself, right? And it's really all about getting all that context aggregated into one place so that you can get the best quality responses.

SPEAKER_00 23:28

I mean, obviously, Zach, we talk about this in just about every episode, keeping the human in the loop, uh, measuring, you know, what sort of responses we're getting, how effective the AI is. That was what you talked about earlier with that observability stack, making sure that we have those systems built in. Uh, what other things have you seen as kind of ancillary benefits from just having that layer of visibility into how the AI is operating and responding?

SPEAKER_01 23:54

Well, in this scenario, you can kind of see which leads are responding. You can see how many are being reached out to each day with the full observability suite. So, not just this supervisor view where you can say this lead was reached out to today and will be reached out to tomorrow. That's useful when you're looking at it at a glance, but with the full agent force optimization and observability tool set, you can go in and you can inspect every single conversation. And so that's where you can go in and you can say, okay, I think we may have said something a little weird to this prospect. Let's go pull up that conversation and audit it and make sure that things are looking right. And some of that, you're going to get quality scores as well. And so it makes it a little bit easier to go in and say, okay, I've got several leads that have a quality score of five or six, and I really need to be eight, nine, and ten. So you can go and look at what those responses look like, determine some commonalities, and then figure out how to tweak the instructions and tweak the configuration of the agent to hopefully make it respond better in those specific scenarios. A lot of that I would say you can also get with testing. There's a concept called prompt evaluation. And the idea there is you use AI to essentially generate scenarios that you think are likely to happen. And then you have AI also grade how the responses to each scenario back align to your goals. And so then you could take 100, 200, 500, 1000 examples and have AI generate scores for all of them. And at the end, you can kind of get one big aggregate score. And so you're using AI to help evaluate how your AI tool is performing. So you're getting a response back from AI to a specific prompt, and then you're using another model to actually evaluate, hey, did it hit X, Y, and Z in terms of how I would evaluate a good response? And then it scores it. And then you can use that sort of batch testing before you even launch the agent to figure out what tweaks you need to make so that you can be reasonably sure that you're going to get good results. But then of course, you got to do that on an ongoing basis. And that's where the observability platform comes into play, which is actual responses graded against some of that criteria so that you can see, hey, I'm trending really well, or today I'm not trending so well. And that allows you to see how you might be able to make some changes to improve continuously the quality of that output. And that's important. It is, it's continuous. These tools are not set it and forget it. You've always got to be monitoring, you've always got to be making sure that somebody is evaluating things so that when something isn't exactly right, you're able to go in and fix it. And it's no different than when you have, you know, a new team member. They're going to go and they're going to do some things and they're probably going to surprise you in how well they're handling some things. But then maybe they got on a call and it's like, well, you could have handled that objection a little bit better, and here's why. No different, right? It's, it's, it's really just making sure that we are, you know, when it comes to a team member, we want to make sure we're training them and giving them all the resources that they need to succeed in their job. And when it comes to AI, it's understanding, well, it didn't succeed and it probably didn't succeed for a specific reason. And maybe I didn't write the instructions the right way, or I didn't give it access to a tool that it needed to. And so in order for me to understand that, I've got to continuously monitor it and then go in and test it and make those tweaks. And hopefully you're you're continuously improving upon that configuration.

SPEAKER_00 27:20

Yeah, I think that's what's really resonating with me in this story, and it's not uncommon, right? The AI, the agent's only effective because the client already did the hard work. The client already did the documentation. They had already defined that up front. The AI is not building any strategy. The AI is executing the strategy that we tell it to with the tools that we give it.

SPEAKER_01 27:40

Yeah, I was on a call with our uh president and CEO, founder of our company just yesterday, where we were talking to uh a different client about how important business fundamentals are to the success of AI implementations. And I think what he said makes a lot of sense. These fundamentals are only going to increase in importance. They're already important, right? The businesses that succeed today have to have good fundamentals. But businesses that are going to implement AI, it's even more important that their fundamentals are sound because the businesses that will win in this era are not going to succeed without solid understanding of what they do and how they do it and really good business fundamentals in the background, right? Like you still have to scale, you still have to manage a team, you still have to understand how to actually treat the customer right. You know what I mean? And so it's like you you can't use these tools as an excuse or a shortcut. Um, ideally they're making your work more efficient, but I don't think you can be long-term successful if if you don't treat the fundamentals of the business as a prerequisite to automating on AI.

SPEAKER_00 28:53

Well said, my friend. Zach, this has been a fun one. Always enjoy our conversations together. Uh anybody out there listening, um, AI doesn't have to be overwhelming. Uh this these are the convers the conversations we love to have. If your pipeline sounds like this, you may be dealing with some similar issues, we'd love to just have a chat. Um, no strings attached. Fastslowmotion.com slash AI. Uh hit us up and we will take care of you. Until next time, see you, Zach. Bye, Eric.