Gerry Marletta on this episode
Gerry Marletta, SVP of Revenue Operations Systems at Union Square Consulting, joins Eddie Reynolds to decode why 97% of AI SDR implementations fail and map the technical and operational system required to make AI-powered outbound actually work. The system is the same problem that plagued Outreach and Salesloft years ago, and Gmail and Salesforce before that: the tool is not the constraint, the engine is.
Building an AI-native outbound engine means designing five distinct agents (targeting, signals, enrichment, scoring, research) that work in concert, with Salesforce as the single source of truth, not chasing a magic AI tool. Most companies spin up an AI SDR, let it write emails, and see nothing. The real work is upstream: narrowing ICP ruthlessly, capacity planning so reps work only high-probability accounts, triggering on behavioral signals alongside demographic data, and scoring holistically so false alarms don't drown out real opportunities. The research and messaging layer—handled by Claude for personalization, n8n for orchestration—is what converts good targeting into actual replies.
In this episode, Gerry and Eddie walk through the catch-22 of enrichment (how do you decide which accounts to enrich before you've enriched them), why separate agents beat monolithic ones, when to hand a step to an LLM versus an orchestration tool, what a grading agent actually audits in rep emails, and which account updates a machine should never be allowed to make unilaterally. They close by framing the GTM council meeting as the feedback loop that keeps the entire engine improving quarter over quarter.
The lines worth sharing
Only 3% of those companies that had implemented AI SDRs had any revenue to show for it. It's not the tool, it's the system, the strategy and the process.
Eddie Reynolds · 4:18
You really do need to make sure you have a feedback loop for all of this data. As you source and act on data, measure success and attribute wins or losses to data quality, messaging, or other elements.
Gerry Marletta · 14:46
Not every signal is going to be created equal. Scoring a signal in isolation is risky; you need to think holistically in combination with account attributes.
Gerry Marletta · 19:30
Web research with Claude has been a real game changer. It's not the only capability, but it 100X what any one individual could do on their own.
Gerry Marletta · 46:17
Common questions from this episode
Why are most AI SDR implementations failing to drive revenue?
The tool is not the problem; the system is. Companies lack proper ICP definition, capacity planning, signal orchestration, and enrichment workflows before they deploy AI. Only 3% of AI SDR buyers saw revenue because they skipped the strategic engine design.
How do you decide which accounts to enrich when you have a huge list?
Use free or low-cost signals first (Claude web research, binary attributes like international presence, title searches in Zoom Info free tier) to narrow the list. Only then spend credits on field-level enrichment for your refined target set.
Should I build one big agent or multiple smaller agents?
Build separate agents (targeting, signals, enrichment, scoring, research) for modularity. Smaller agents are easier to debug, iterate, and reuse across workflows. They also allow sub-loops where one agent can call another (e.g., scoring calling enrichment for premium accounts).
When should I use Claude vs. an orchestration tool like n8n?
Use n8n for repeatable, fixed workflows (form-to-email-to-CRM pipelines). Use Claude for research-driven decisions, content generation, and judgment calls where ambiguity exists. LLMs excel at synthesis; orchestration tools excel at reliability and cost.
What makes a sales email actually good in an AI-native engine?
Keep the template structure fixed (greeting, CTA, signature) and use research to fill the personalization sections. Two-times higher response rates come from genuine account-specific insights, not over-scripted templates or false urgency.
How do I know which account and contact updates should be automated vs. human-reviewed?
Never let a machine unilaterally change critical fields like status, stage, or close date without human approval. Enrich and score, yes; commit to CRM actions, no. Build in human gates for decisions that affect deal velocity or manager coaching.
SEO meta description
Gerry Marletta reveals why 97% of AI SDR deals fail and maps the five-agent system required to build an AI-native outbound engine that actually drives revenue.
Target keywords
AI SDR revenue
AI-native outbound engine
Gerry Marletta
targeting agent
signals and triggers outbound
account scoring methodology
Salesforce source of truth
Claude vs n8n
personalized sales email research
outbound enrichment waterfall
go to market automation
Read the complete conversation
Auto-generated from the episode audio and lightly cleaned. Expand to read or search inside.
Read the full transcript · 93 KB · Gerry Marletta
GERRY0:00there's obviously something to be said about having an agent go out in the concept of,
GERRY0:05self-starting,
GERRY0:06but a lot of the tools that we're talking about here, whether it's zoom info or Sixth Sense, they also have ways of proactively pushing
GERRY0:12those signals
GERRY0:13into your systems as well.
GERRY0:15there's always good old fashioned web research, which just like we might have a beta or an AI, go out and look to see if there's any of these kinds
GERRY0:24signals or things in the press, in the media. Now we're using Claude, for example, to do that, but
GERRY0:29Claude and or Salesforce in conjunction with Zoom Info and six and some of these other tools can catch these signals as they occur.
GERRY0:34Those platforms are already doing that, going out and keeping an eye out for those things so they can be proactively pushed to you. So you're not waiting for Claude to go and find that information?
EDDIE0:52actionable playbooks from the trenches working as go to market strategy and rev ops consultants for our clients here at Union Square. Consulting and candid conversations with revenue leaders in the market that have been there. Now let's get into it.
EDDIE1:07welcome to another episode of Go to Market Science. Today we're going to talk about building an AI native outbound engine. And our guest today is our very own Jerry Malta, our SVP of Go to Market Architecture.
EDDIE1:18Jerry, I just changed your title on the fly just now. Thanks for joining us. Happy to be here and appreciate the new title. Yeah. No problem. It's so funny because like, spending so much time thinking about the best, you know, wording to describe the work that we do. And it's like constantly evolving. And so maybe that's your new title.
EDDIE1:34At least it is for this podcast. I think it's an x an hour scripter. Yeah. So anyway, let's get into this, the elevator pitch here. What we want to discuss today is how to build an AI native outbound engine. And what we've learned building this for ourselves and our clients. We're going to discuss designing the engine, meaning like not just a technology, but the process and the system and how it's all going to work.
EDDIE2:00Targeting triggers, signals, enrichment, scoring, segmentation, the research process, and of course, the actual messaging. But in addition to this, we're going to talk about how to actually make this work, how to run outbound, even driven by AI that's still human and personalized, how to analyze the data to improve it, how to leverage the data for management to provide coaching to reps to improve, and then how to build this into a continuously improving engine that we iterate upon again and again and again.
EDDIE2:33So it gets better and better quarter after quarter. So Jerry, thank you for joining me today. I'm so excited to have you helping me fill in the technical gaps that I need help with. As we talk through this concept that you and I have been, you know, workshopping for this podcast the last couple of weeks, and I'm happy.
GERRY2:52To be here, happy to help. And, you know, I love talking about the technical stuff. So I'll take any opportunity to do so with
EDDIE2:58Cool. So I'll just share a caveat for our audience here. So we're also writing an article that you'll be able to find. It'll be in the show notes here on our newsletter, and it'll probably come out before this podcast is published. But in this article and in this podcast, we're going to be sharing an example here. And it's going to try to be a cookie cutter example.
EDDIE3:20But what this is going to look like for each company is going to be very unique and different. A company that's targeting 300 named accounts in the fortune 500 is going to be very, very different from a company selling to a million plus SMB companies across the United States. Right. Like identifying 300 named accounts in the fortune 500.
EDDIE3:38This is extremely public and easily accessible information. Identifying every new business that started yesterday is significantly harder, and the approach is going to be very different. And what's important for each is very different. So I just want to lay that out there. Also to ease your concerns, Darya, that you don't need to provide a one size fits all answer to all of my questions as we go through this, but we will try to like, kind of strike somewhere in the middle and talk about what this engine would look like for the average company.
EDDIE4:07That's not necessarily on one extreme or the other. So, Jerry, I'll start and I'll talk about the first couple pieces, and then I'll start to layer in questions for you as we go along.
GERRY4:17Sounds good to me.
EDDIE4:18Awesome. So the very first piece here, and I want to preface this by saying, like, anybody can just go buy an AI, SDR and spin it up, and the AiSd is going to write messaging for you. A lot of these tools can help you add contacts and you're off to the races. Unfortunately, this isn't working for anybody. Sasser did a survey of a number of companies that had implemented AiSd Is, and only 3% of those companies that had implemented HDR had any revenue whatsoever to show for it.
EDDIE4:46And the reason for this is the same reason that companies that used outreach and sales loft a few years ago didn't have any revenue to show for their outbound efforts, or the companies before them that were using Gmail and Salesforce. It's not the tool, it's the system, the strategy and the process. And we need to start by designing the engine.
EDDIE5:03And that starts with our ICP and our buyer personas. We've talked pretty exhaustively about that, but we need to get our ICP and our buyer personas defined as narrowly as possible, with really granular details and data that we can get from various sources out there, including looking at the company's website, and use that to narrow in our targeting list so that we can get the best possible prospects.
EDDIE5:29By definition, or ICP, are the companies that are the easiest to sell to, the easiest to serve, the ones that are happiest, the most profitable, easiest to retain and expand. And so if we're targeting any company that's not squarely in the center of our ICP, then by pure definition, we're going to have diminishing returns. Once we've figured out who that ICP and those buyer personas are, we need to outline the sequencing process.
EDDIE5:52What has worked for us in the past to reach out to folks is at five emails. Is it six? Is seven? Is it this many voicemails? Is it LinkedIn DMs? Is it sending actual snail mail to people? What are the things that have worked in our own business for our customer, not someone else's most effectively? And what sequencing process do we want to use going forward as the next iteration?
EDDIE6:14And then what is the research and messaging process look like? Study after study shows that companies that do personalized messaging, that actually research the accounts they're targeting and reflect that research in their messaging have significantly higher, in fact, on average, two x higher response rates. And that's on top of all the other things I mentioned, where these customers are faster and easier to land and they pay more money.
EDDIE6:38It's a significant uptick in the revenue that we can produce.
EDDIE6:41So the first piece that we want to go want to fix here is designing the actual engine before we start implementing technology.
EDDIE6:48Jerry, I'm kind of like on a soapbox here, but anything you want to add or clarify on that piece before I move on to capacity planning?
GERRY6:54No, I think you've probably have. Or we will in future conversations, talk about how we can leverage technology to make
GERRY7:02documenting, defining, building all of those other elements you just said as sort of precursors to this, to make that even easier and smoother. It doesn't have to be necessarily all done by hand.
GERRY7:12But again, that's probably for another episode to dive into
EDDIE7:15Yeah, and we did a newsletter on this as well. I forget the exact name of it, but it's something like the AI driven, continuously improving ICP engine and overlaps with this somewhat in the sense that we need to get all of our data collected together, and then we can put this into AI and ask like, what are our best?
EDDIE7:31Like, we'll first identify our best customers and then ask, like, what describes our best customers and what data points differentiate our best customers from the ones that we didn't win, from our worst customers that churn? And how do we build a better ICP and buyer persona definition, right? I don't want to go into great detail on that here, because that was and is its own separate piece of content that goes into great detail.
EDDIE7:55But there's a lot of overlap here as well. So we'll touch on this. And for folks that want to check it out, you can find the link to that article in the show notes here as well. The next piece though that I want to tackle here is the capacity planning. And the reason this is so important is and I'll just speak from personal experience, every sales job I've ever had in my entire life, dating back to like the first B2B sales job when I was 22.
EDDIE8:18I've basically been given like a list of thousands and thousands of accounts, and I could maybe work like a few hundred of them at the most in a year. So the problem with this is that, like, it basically puts your rep in a situation where they've got 1 or 2 choices. They can either just start calling and they make it through, let's say 200 out of 2000.
EDDIE8:38And that means that they're going to get 10% of the best accounts and 90% of the worst, and they're going to miss 90% of the best accounts just by pure math. Right. Or they can try to become a data analyst and do what the rev ops team should be doing and trying to do all the work we're talking about here, instead of actually selling.
EDDIE8:56Neither of these situations is a good situation, and so we are so adamant that it's important to do a capacity plan so that we can use the ICP and bio to narrow our list down to a set of workable accounts that our team is actually going to work, that they're actually going to message, and not just once, but however many times we need to message them through our defined sequencing process to maximize the conversion rate of those accounts or contacts.
EDDIE9:21Anything you want to add on that? I mean, I think it's very self-explanatory, Jerry, but any other thoughts that you want to add before I go to the next piece?
GERRY9:27Yeah, I think we just don't want people to feel trapped by what those numbers may be.
GERRY9:32You can always
GERRY9:33refine what the size of the account. Sorry, the the volume, if you will, of those target accounts should be as the capacity of your team changes, you can always expand or shrink the size of your team. Like these are these are things that are going to shift over time, just like the earlier elements.
GERRY9:50In the first step we were talking through around the ICP and who are the best customers to be trying to sell to and what messaging is working, what isn't. These are all things that should not be like a set it and forget it. You should be revisiting and reevaluating all of these elements as a more macro level of concept of continuous improvement.
EDDIE10:08Yeah, absolutely. And we're going to get into scoring later as well. And so like I don't want to be too hard and fast with this. Right. Like let's say that you give her up 2000 accounts and they can only cover 200. But you've scored and graded all of them, and they can easily see who the top 200 are.
EDDIE10:22And they start at the top and start working through them. Okay, I could make some arguments for why you wouldn't want to assign the other 1800 accounts to that rep. Maybe account number 201 is actually a decent account, and that should go to a different rep and actually get worked. But it's not nearly as bad as a situation where they're just staring at 2000 accounts and there's limited data, and they can't tell the difference between one and the other.
GERRY10:43Yeah, I don't know where to start.
EDDIE10:44So let's now go into our targeting agent. And the first two pieces were human led. And this is going to be the first one. This is like where we can actually build an AI agent. Right. And so what we need to do with this targeting agent is to select our data sources for where we are going to find the accounts and the contacts that we want to add to our targets, and build a waterfall, because we may want to use multiple different data sources.
EDDIE11:09Now, this is not a new concept that just came about with AI. But Jerry, I'd love if you want to just touch on that a little bit about the best way to do this and what you've seen work best, or what you've seen not work so well in this regard.
GERRY11:21Yeah. You know, there's a few factors to consider this. It's what tools you have available, what tools you have budget for, and then what kind of budget you have for spending on doing this research. But what we've seen work best is to start with using clay as sort of the the overarching control mechanism for this step. Because the nice thing with clay is you can pick and choose on a per data point element what, what data source you want to use.
GERRY11:51Beyond that. In other words, we could choose to use zoom info as a key element for identifying accounts to begin with. You can also feed into clay the sort of the definition, the outline of your ICP, and who your personas are. So you're not necessarily dealing with the old school box of filters. You've got to click around and hope you can select all the right combinations of things.
GERRY12:15You can use plain language to define exactly what kind of accounts you're looking for, where you want to start, where you want your spend to go, which data sources for your own particular industry you trust more than others. And again, you try to be as cost conscious as you need to or not need to, depending on those sources.
EDDIE12:33Yeah, I think the thing that's most important for me is just like, what is the quality of that underlying data? And every company that I've worked in, every company I've worked with has a different set of data that's important to them and different data sources that are more reliable for that data. Right. And so we can say, okay, clay is great, zoom info is great.
EDDIE12:51But I was on the phone the other day with a company that sells to school districts, and there's like they sell specifically facilities management. And it's like, okay, these are particular data points that are very valuable for their outbound that other companies don't care about. When I worked in private equity, we wanted to see how much assets institutional investors were managing.
EDDIE13:12I don't believe that you can get that in clay or zoom info. So there are these like really unique data points that are so, so important for an organization that it's really difficult for me to have this like one size fits all. Just use this tool and then use this tool and then use this tool. It's like, no, like if I need to know the assets that they manage or I need to know how many facilities managers they have or whatever it is, where is the most reliable place to get that particular data point?
EDDIE13:35Then what's the next data point? Do you have any more thoughts on like how best to do waterfall? Jerry I think for me it just comes down to like, well, let's start data point by data point and what's the most accurate source? And then where do we go next if we can't find it there anything more you'd add on that.
GERRY13:50Well I'm going to start with the most, you know, identify what the data points are that you need in order to execute this overall outbound process. Trying to come up with a good example of something that's probably not as important for company level information. You may not need to know where all of their subsidiary geographic locations are, so don't start with trying to identify that data point.
GERRY14:14Start with something that's much more higher level, whether it's annual revenue or industry. Some of those other factors, you want to start with those first. And you really do need to make sure you have a bit of a feedback loop for all of this data. As you as you source the data and as you act on this data, you kind of want to measure the success and be able to attribute successes or losses in your process due to the data, due to the messaging, due to some other element of the process, and if it's due to the data, then you may need to evaluate using a different data source.
GERRY14:46Your business and the folks on your team may not have the experience or the exposure to each of these data sources to know which ones are best. So you kind of you might have to experiment to start, but experiment small to start if you can. Don't try to find 20,000 accounts using some small, you know, data source that you've never really worked with before.
GERRY15:05That's a real gamble, and you're going to be burning a lot of resources on that unnecessarily.
EDDIE15:10Yeah. And I think like the location thing is a really good example because it could be both extremely valuable and important for one, you know, team selling motion, etc. and not for another. I mean, I remember a client that we were working with also sold to school districts. We don't specialize in this, but these are just the things that are coming to my mind.
EDDIE15:27And they said, you know, like, we'll have this school district here. And they bought in and they all know the people at the school district like five miles down the road. So that's really critical data for us to have. Right. And then of course, there's a million examples where we don't really care where they're located. It doesn't matter.
EDDIE15:43And that's why I think it really comes down to like sitting down and thinking about like, well, what's your process? I mean, if your process is, well, we researched and the first thing we look for is, are they five miles down the road from the school district that we have as a client that needs to get baked into this entire engine?
EDDIE15:59They can't see you nodding. Jerry, you've got to use verbal cuz my.
GERRY16:02Friend, I know this is a both an audio and visual medium. Yes, I agree, I was going to try to come up with some other examples of what may be relevant for one organization and not for another, but we've probably delivered that message here the best we could.
EDDIE16:17I think we've exhausted it. We can move on to the next thing. So now we're going to talk about triggers and signaling agents. Right. Or signals agents agent. So we've got our targeting agent. And the idea here is that we're using AI to orchestrate and to automate the collection of a number of accounts we want to target. Now, depending on the business model, this might be done once a year.
EDDIE16:38It might be done once a month. It just really depends on how quickly it makes sense for us to iterate. But now we're going to go out to the world and we're going to look for other triggers and signals that are relevant for our go to market motion. Common examples would be, you know, new executive hires, fundraising, a past champion that you had as a customer or in a deal that is now in a different organization, obviously, third party intense signals, right?
GERRY17:09M&A events.
EDDIE17:10M&A events, data breaches. Like if we're talking about data security, you know, that there's like a hack or something like that. Like there's a million examples, right. The challenge I've always had with triggers and signals is like the old school way to do this is you go in as a sales rep and you build some kind of an alert, and then all of a sudden your inbox starts getting flooded with all these messages.
EDDIE17:32I literally saw one of these messages from zoom info today, and then you're like, well, what do I do with this? Like, I've got this list of accounts that I've got to work, and then I've got these other like set of messages that's like flooding my inbox. Like, who do I call next? And so the idea here is, is that we want to build an agent that's going to go in and grab these signals and add them to these accounts and possibly identify accounts that are still in our ICP, but for whatever reason, maybe didn't come into our targeting account list so that we have this like holistic set of data that we can then
EDDIE18:04take to the next step, which would be enriching the data and then ultimately scoring the accounts so that our reps know who to call first and that you would add on the triggers or signals, or just like the best way to build this and orchestrate this.
GERRY18:20I think the
GERRY18:22important part on this, we talked about this a little bit on a side conversation, is the idea of signals that may fire for organizations that we haven't already identified as being important to us, and really just thinking about that generally in the concept of like scoring a signal in and of itself isn't necessarily enough to warrant picking up the phone or sending an email.
GERRY18:47It needs to be thought of holistically in the context of the company itself. The timing, depending on what the what the particular signal is, as well as any individual contacts within the organization that are involved in that signal. So if we were thinking about it, I know we're skipping ahead and just talk about the concept of scoring. If we were to score a certain kind of signal a certain way in and of itself, you want to be careful that that doesn't elevate an organization too high on your list or keep it too low.
GERRY19:18You really have to think about it holistically in the combination of these other attributes about the organization.
GERRY19:24Yeah, I think that's just a really important thing, because what I have seen, I've seen organizations that when we do get to scoring,
GERRY19:30certain indicators, score certain way, and all of a sudden we're blowing up as lists of who they need to reach out to, and they take a look at these companies, they look at these contacts and they say, this isn't really all that valuable.
GERRY19:43These are sort of false alarms. So again, like we keep saying here, this concept of continuous feedback and always having a feedback loop in place to refine these things. And that is one place that we're utilizing AI technology. AI tools can really help. You know, again, the old school way of thinking about demographic information from a graphic information and the behavioral information independent of each other was kind of the only way you had to do it.
GERRY20:09Now we can start to leverage Llms to think more realistically in applying scoring values. So not every signal is going to be created equal. Kind of what I'm getting at.
EDDIE20:19So let's talk about how to actually build this from a technical perspective just for a moment. Because, you know, I always spend so much time on sort of like the strategy and the thought here. But I know people do want to know, like what tools to use. And I'm always a little bit hesitant to share this because it's like, well, you know, by the time this podcast publishes, maybe like ChatGPT has inched ahead of Claude and now it's a better model for this, or that's just what your organization likes using.
EDDIE20:43It's like, fine, I don't care about the tools, but people want to hear about the tools. It's just like tools a little bit. And specifically how to orchestrate all of this. Right. So
EDDIE20:52I think like the infographic that we built for this, the major tools that we're looking at is first and foremost like cloud and Salesforce, Salesforce being like the ultimate system of record for this right now.
EDDIE21:04There's other tools like outreach that we can use to actually do the outreach. There's tools like zoom Info and Clay that we've already mentioned that provide the data. But at the end of the day and tell me if I'm missing anything here, Jerry. Yeah, by and large, each of these agents should be going out and finding data from various sources and then updating Salesforce with that data so that sales forces are our single source of truth, so to speak, specifically for the accounts that we are identifying and the outbound activities and results that we are doing and seeing with these accounts.
GERRY21:38Correct. I would agree with you on
GERRY21:39The only thing I'll add to that is there's there's obviously something to be said about having an agent go out in the concept of, of self-starting, but a lot of the tools that we're talking about here, whether it's zoom info or Sixth Sense, they also have ways of proactively pushing those signals into your systems as well.
GERRY22:00So, you know, there's there's always good old fashioned web research, which just like we might have a beta or an AI, go out and look to see if there's any of these kinds of signals or things in the press, in the media. Now we're using Claude, for example, to do that, but Claude and or Salesforce in conjunction with Zoom Info and six and some of these other tools can catch these signals as they occur.
GERRY22:22Those platforms are already doing that, going out and keeping an eye out for those things so they can be proactively pushed to you. So you're not waiting for Claude to go and find that information?
EDDIE22:32Yes, absolutely. And that's exactly what I'm thinking here when I describe this, you know, triggers or signals agent. But for you, the best practice is to take that data, these triggers and signals, and add it into the accounts in Salesforce so that we have the account, we have the trigger, the signal, etc., all in one place that we can then feed into the next agent to enrich and then score these accounts so that Salesforce becomes our central source of truth for all of this data that we're collecting and analyzing on this AI driven outbound engine.
GERRY23:03Correct. You don't want to rely on your AI tool like Claude to remember and air quotes, if you will, to remember all this information. It needs to live somewhere and it needs to live somewhere, you know, sort of programmatically database in a way that, you know, things don't drift over time. So you want to aggregate it all in one place.
GERRY23:21CRM is a wonderful place to hold all of that so that when we get to later steps in this process, whatever tools you're using can reference them and reference them all the same way and interpret the information all the same way.
EDDIE23:33Awesome. And then you and I have had this debate internally, and it sounds like there's not necessarily like a single one size fits all answer. But as I was putting this together, I was really thinking about each of these being separate and distinct agents for one primary reason being just that. Like if it's not working correctly, it's easier to fix a single agent.
EDDIE23:52So if we say, okay, this is our triggers and signal agent, but it's not the same agent that goes in like sources, the original target accounts that fit our ICP and that trigger and signal agent suddenly stops working as well. For whatever reason, it's easier to fix that than to fix a much larger agent that's trying to do multiple things.
GERRY24:11Yeah, and I go back and forth a little bit on this that it's sort of a
GERRY24:14don't want to say older, but it's more of a
GERRY24:17programmers and developers point of view. If you always want to have your chunks of code, if you will
GERRY24:23be.
GERRY24:25I guess it's sort of the messy concept in a way, but you want to have them be these independent bits of code that are doing certain kinds of work so that when it comes to be time to create some new process where, you know, a portion of that is going to rely on something you did somewhere else, you're not rewriting all that code.
GERRY24:40And this other thing you can say, hey, go look at this. Go look at this. I don't know, enrichment agent and reuse it in this inbound process that we're building.
GERRY24:48Obviously, it's a lot easier to do that because you're not rewriting what you did on the enrichment side again.
GERRY24:53that also is going back to the days when we were handwriting code itself, which in and of itself is more time consuming.
GERRY24:59A lot of what we're talking about here, when you really get into the weeds of building the agents and leveraging skills and things like that, a lot of it is we're using plain English to define all of this stuff, obviously.
EDDIE25:09But even then, isn't it harder to tell? Like Claude, for example, hey, go analyze this, this massive thing that's trying to do five things at once versus, hey, just go and analyze this triggering agent and make this change, make this update, tell me why it's doing this or that.
GERRY25:23Yes, the concept is still applicable. I'm just saying in terms of the administrative overhead, it's not nearly as heavy as it used to be, which is what encouraged that original mindset of having these discrete blocks of work. It still is beneficial here, though, because again, you can you can repurpose these discrete agents and have one reference another, for example.
GERRY25:43So yeah, there's there's definitely still a benefit and a reason to do that. And also again, as you as you refine the process and therefore you're refining the instructions for these agents. Again, it is easier to open up a ten line document versus a 1000 line document that covers all of these steps on one place.
EDDIE26:00awesome. And then so I think my next question here is, is like, what do you see as the best practice for using Claude versus an orchestration tool like Natan to, to orchestrate this entire flow.
GERRY26:14Really comes down to when and where you need
GERRY26:17a thinking partner with how to approach
GERRY26:21a situation or a need, versus how and when you need. Given a certain set of inputs that could change. You want the same steps to be done and executed, and then the output to be fed back to you. So by that I mean, let's just take an example of somebody fills out a form on your website.
GERRY26:44I'm going to I'm going to assume just for the sake of this example, none of your systems are actually integrated or connected to each other. So you've got forms on your website. You have HubSpot, you have Salesforce. Let's just say you have zoom info, and then let's say you have outreach or Apollo or some mechanism that you can really easily, automatically send an email.
GERRY27:04Let's just say you want to get an email out to somebody who fills out a form on your website right away to connect all of those elements that I just stated out loud, I wouldn't use. Claude, you don't need a machine to try to remember based on a set of instructions and reinterpret every time the order of operations and what exactly to do you can very easily have, and which was going to work great for the orchestration piece to say.
GERRY27:27When the form is filled out, first take the information and put it into HubSpot. Then go and find this person on zoom info and enrich the data in HubSpot. Then take that record in HubSpot. Reconcile it with Salesforce. Then, depending on the information we have, you're going to pick template A, B, or C, and you're going to draft up an email and you're going to notify the AI that it was assigned to.
GERRY27:48Maybe you have another step in there of of some lead writing or lead assignment leveraging. And although I wouldn't necessarily recommend it for lead routing, you give the notification says review this draft and send it if it looks good to go. That's going to be the same process every single time you run into some risk using an LM of it, doing some of those steps a little different every time, maybe not enriching all of the fields the right way from the right sources.
GERRY28:14You're just open yourself up for a room for error, and then I don't want to keep coming back to this. But there's also just like the cost, obviously, something like nine is not going to really be so much usage based. Whereas depending on how you're leveraging an LM tool, you may have some usage based element or some limits within a certain time period that you want to try to stay within.
EDDIE28:33Yeah, that's super helpful. And like has probably everybody knows like I haven't tinkered around on these tools in my own hands and like ten plus years thanks to having you here. But like ten years ago when I was building all this stuff out with Zapier, which is another tool that we could consider for this, it was quite easy to do this, and all of this could be done without AI.
EDDIE28:54And the example that you run through, the only place where you run into, like maybe an AI use case is drafting the email, right? And so everything else is extremely programmatic. And then if you say, well, let's draft the email, and the email should say these exact things. Hi. Insert name. Thanks for reaching out. Would you like to book some time tomorrow?
EDDIE29:11Click this link. That's something that like an orchestration tool does really well. But then if you want to go and do research and personalize that email, that's where we then need to use an LLM. Correct me if I'm wrong in any regard there.
GERRY29:22No. That's right. And even on the email itself, elements of composing the email are going to be a blend of like programmatic versus a genetic right. The programmatic part is going to be a greeting, an opening paragraph, a customized block of text, and then a signature that doesn't change. I never want let's just say this is how I'm thinking about it.
GERRY29:43I never want that structure to change. What goes into each of those can be ejected. We can have an LLM determine what the greeting should be based on the person and the role. The the block of personalized text should very much be based on aggregating all of the research. So it's a combination of both of those elements just in building the email itself.
EDDIE30:02Yeah. And I've got some strong opinions on this. And we're going to get into that when we get into our messaging piece.
GERRY30:06Yes please do. Because the example I just gave is garbage and would raise my own personal red flag if I received an email structure that way, very obviously put together by a machine.
EDDIE30:15So next we're going to get into the enrichment agent. And you and I had a lot of debate over whether or not this should be a separate agent or it should be part of one of the agents we just mentioned. But the idea that I have here is that. And again, this is like a one size fits all that will not fit all companies.
EDDIE30:31But you've gone out and you've built your target list or you've had you've built an agent to go build your target list. However frequently you need to do that, you then have a separate agent that is working with tools that are constantly monitoring signals and triggers in the market, and pushing that information into Salesforce to update those accounts and or create new accounts.
EDDIE30:50But now we may have a set of accounts that are not fully enriched, and before we score them, we need to enrich them. But here's the problem that you highlighted when we were talking about this. How do you know which accounts to enrich before you've enriched them? And this is where you and I had this debate. And the thinking that I had is, is very case dependent.
EDDIE31:13But essentially what tools do we have? What do we need to pay for them? Like what data can we get access to without spending too much money? And then is that enough data to then decide which accounts to enrich?
GERRY31:26Yeah. Like a.
EDDIE31:27Pretty.
GERRY31:28Basically, yeah.
EDDIE31:29So as an example, let's say that our sales team has let's say we have like 50 salespeople and they can each cover 200 accounts. And so we can cover 10,000 accounts. Right. And let's say our total addressable market is 100,000 plus companies. Right. So we're looking for less than 10% of the market here.
EDDIE31:45If we go out and grab a list of 100,000 companies, is there some data point on that list that we can use to say, okay, here's the 10,000 companies that we want to enrich, or if our goal is to end up with 10,000 companies, maybe we want to enrich 15,000 or 20,000 companies.
EDDIE32:03Love more. I would love to hear more of your thoughts on this, Jerry, because when you and I were arguing over this, I thought it was a great conversation.
GERRY32:09Yeah, the argument stemmed from this idea of being a catch 22.
GERRY32:13If all I have are a list of 10,000 company names, and I can't afford to enrich 10,000 companies with all the data points that I need, how do I know which ten, which 100 I can afford to enrich? So yeah, where we landed was this concept of like, well, use what you can start with what's cheap or free.
GERRY32:33That is. And that's not just dollars, but that's human time as well. And I'm just sitting here and I was trying to think of an example. And, you know, one, one thing maybe, you know, a really important factor for your organization may be a presence in the US or an international presence. That is something that you might be able to use something like Lord to just go and find you're not paying for credits in clay.
GERRY32:55You're not leveraging zoom info to find all the subsidiary locations, which in theory would probably be far more time and costs consumptive. Not sure if that's cost consuming, time consuming to get all that information, you don't really need to know all of the locations. It's more of a binary certain international presence or not. That's something that you can leverage cloud for and find out very quickly, even across a wide set of accounts.
GERRY33:17So that may help you refine that list of 10,000. And, you know, bring it down to 1000, let's say. And then if there are other data points that, again, you don't have to use a more programmatic tool like clay and zoom info and Apollo. Again, where can you leverage an LLM to try to just do web research, do some research on?
GERRY33:38I am hesitant to say LinkedIn because, you know, we try to be careful when we think about using machines in LinkedIn these days. But again, what's what's readily and freely available for you to try to to narrow that list down more and more, and then you unleash the hounds, and then you start enriching all of this specific field level detail and pulling the context that you need for those accounts.
EDDIE33:57Yeah. And I mean, we can just talk about like exactly what we did internally. I mean, we spent a lot of time working with private equity and venture capital firms. And last I checked, there's about 10,000 of them out in the world. Right? Unfortunately, a good chunk of those 10,000 have very little money, if possibly none. It's like one guy that goes and, you know, writes a check for $20,000 into some startup or what have you.
EDDIE34:19I mean, there's nobody, like, standing in the way preventing you from starting a venture capital firm tomorrow with 20 bucks and, you know, two dogs in a garage. But the folks that we work with don't fit that description. Right. And it's not difficult at all to literally just go to Google and say, give me like the top 300 private equity and venture capital firms in the world based on assets under management, or to ask Claude to do that, even if you didn't have any of the data sources we mentioned, which we do.
EDDIE34:46And so the approach that we took is looking at kind of like the top thousand accounts and then trying to filter that down to the few hundred that we can work, because the other 9000, we might miss 1 or 2 here and there. But by and large, like, you know, if a PE or VC firm is number 9000 on the list because they've got $200,000 in assets under management, they're just not going to be a fit for us as a partner.
EDDIE35:11And we don't need to go and spend a lot of like mental energy thinking about how we might enrich that data set or analyze it because it's just not valid. Right? On the surface.
GERRY35:20Something I learned, both in doing this exercise and in working with other customers, is that
GERRY35:26should also be aware of,
GERRY35:28even with a tool like zoom info,
GERRY35:29there's a certain level of detail that is available for free, depending on what kind of information you need to know to do this. Early scoring and ranking
GERRY35:38and prioritization,
GERRY35:39that may be some of those free data points.
GERRY35:41for example, you know, what's really beneficial for us is to understand what PE and VC firms have. Anybody in a go to market advisory role
GERRY35:47just doing a title search of individuals within those organizations,
GERRY35:52are some titles that are very clear that that is the function and the role within the organization that that person does. So that's a that's a nice sort of binary.
GERRY35:59Yes. And then obviously doing a count of how many of those are can also help further drive the prioritization for us. So you don't always necessarily need to be spending money with zoom info, for example, to get some real high level information that's useful for you. It's when you start to get into the more micro level phone email address.
GERRY36:18I color anything crazy like that where you're actually spending those credits.
EDDIE36:22Yeah. And the interesting thing about that, I mean, it's very unique to us is that some of those people have very clear titles and others do not. Right. And then that's where we run into this issue, where we've got somebody with this title and based on the title alone, we don't know what they do. And by extension, we don't know if that firm is somebody that we have a real opportunity to work with, because we don't work well with folks that don't care about go to market, unsurprisingly.
EDDIE36:51Right? Yeah. Not to say that like investors in PE firms don't care about go to market, but they care about it a lot less than somebody who's like a go to market operating partner. And so these are things that are really important to us. It's not sure all of our secret sauce, our own auto market. But okay, so we send this into the enrichment agent.
EDDIE37:06And we essentially need to again define the data, build the waterfall, integrate the different tools that we're using, and then automate this. I want to.
GERRY37:13Agree with you for one second. Of course, I think a really important piece that we sort of touched on at the top of the call, at the top of the conversation, the reason why we are breaking this out, also as a separate agent in this whole process here is because we're really thinking about this whole loop under the concept of new accounts, new companies and new contacts, as well as who's already in your database.
GERRY37:38The concept of the enrichment agent is not just focused exclusively on new companies and new contacts in those companies that you're targeting. This is also something that should be running against your existing data. You know, as most people listening this probably know, data goes stale x percent over every year. So like you don't want your data to go stale, whether it's something as simple as the person has moved on from the organization to just title changes or contact info changes, like you want to keep that up to date, let's not forget that we still have existing data in our systems that is always going stale.
RACHAEL38:11pause. Everything we talk about on this show. Diagnosing go to market ops. Prioritizing projects for revenue. Impact processes. Metrics. Insights. Building a predictable go to market engine. We've built frameworks for all of it. They're free and undated on our website. Union square consulting frameworks. The link will also be in the show notes, so make sure you check that out.
RACHAEL38:33All right. Back to the episode.
EDDIE38:35Okay, so now we're done with the enrichment agent. We have created our target accounts. We've added triggers and signals to those accounts, as well as potentially identifying even more accounts that come through those triggers and signals. We've sent it to the enrichment agent, enriched it with all the important data that we need, especially for scoring. And I think like the concept here is like really at any point along this entire journey, we could always choose to enrich our data,
GERRY39:00Oh, yeah. Absolutely. I mean, because you and unless you have a completely empty database, you always have some data
EDDIE39:05No, what I mean is, is like, if you think about each of these agents, in theory, any of these agents could go back to the enrichment agent and say, hey, we now need to enrich with this other signal, especially if that signal is expensive for us to get right. Let's say maybe we get down to like our absolute top tier one accounts, and we need to see x, y, z data point.
EDDIE39:23And it's just especially with clay, like some of those data points are much more expensive to get than others.
GERRY39:28Yeah. Yeah. There's I mean, this goes back to the conversation about why you break these out into separate agents, separate skills, if you will, things like that. And the sort of draft visual I'm looking at here of ours, you know, you could add a layer of these sort of like sub loops that are occurring where this one agent is still at times referring to or calling the enrichment agent.
GERRY39:46So yeah, there's there's sort of loops within loops and
EDDIE39:50Cool. All right. So now we go into our scoring agent right. And so we need a methodology to essentially create an account score so that we can stack rank these accounts and then assign them to reps and have reps figure out who to call first or build out our sequences accordingly. And so we first need to score the accounts.
EDDIE40:08We probably need to score the contacts within those accounts as well. And then of course we need to build this as like an automation so that this agent can continue to run. Now in order to do this, obviously it's really important that we build a scoring methodology. And Jerry, this is something you have a lot more experience at than I do.
EDDIE40:26I'm curious like what you've seen with account scoring and previous to this lead scoring, I've always heard you say that like you need to start with a hypothesis and you need to iterate on it because it's pretty much impossible to get it perfect the first time. Correct?
GERRY40:39Correct. So you need to have a. If you don't have a passionate feeling about it, you need to just pick something and stick with it around. Demographic. From a graphic information about an organization and individuals in the organization, what things are more important than others nowadays. I would also suggest having a conversation with an LM to try to get some feedback on your hypothesis.
GERRY41:05Have it poke holes in it, see if there's any flaws that you may not be taking up, particularly within your industry. And based on your ICP and the personas that you put together, you don't want to lose sight of that. The really important part, the part that actually, I think swings the needle more drastically in either way, and is why we always encourage having this concept of, again, a feedback loop when it comes to scoring is the behavioral and the sort of trigger and signal based elements of scoring itself.
GERRY41:32Because as I was mentioning earlier, those those could blow up. I don't want to say blow up in boxes. It's going to blow up your team's prioritization, who they should focus on sometimes in the wrong way. Sometimes it might work against them, and nobody is crossing a certain threshold or hitting the top of their list. They've exhausted the 20 or so organizations and contacts they should be reaching out to, and others waiting for something to get scored a little higher.
GERRY41:55And they're not. So you really need to have some sort of a feedback mechanism internally for your reps and your bars, and to talk with folks within the organization that are responsible for putting together the scoring. And again, you can start to leverage your LM tools, you, Claude and others to also be a part of that feedback loop and evaluate things that might escape reps.
GERRY42:18Providing feedback a getting reps to provide feedback on scoring is like pulling teeth. B there are indicators that are a little bit harder to track at scale, even if something as simple as like clicks and opens and unsub's and all of those, you want to take those all into, into aggregate when you're thinking about how you scored somebody and then again, not looking at each behavioral element of the scoring scoring methodology in a bubble, it needs to be thought of in combination with other things that are occurring.
GERRY42:50So sometimes when we think about like a HubSpot scoring mechanism, you may award a certain amount of points again for a form fill or a website visit that may or may not be enough in and of itself, but when you combine it with 2 or 3 other things that happen within a small window of time, maybe that score should be much higher than the sum of those parts.
GERRY43:10And that's where leveraging again, an LLM can be better than relying solely on the more old fashioned, if you will, scoring tools that exist out of the box with something like HubSpot.
EDDIE43:19Yeah. I mean, the nice thing about an LLM is that it can essentially look at multiple variables at the same time and try to try to compare those, which is just mathematically more difficult. Right. And so
EDDIE43:30so we now like below the or after the scoring agent we had in our visual that this hands off to a contact agent. Now I can see lots of exceptions to this. But the idea is is that like the first thing that we're trying to do is figure out which accounts are going to be the most valuable that we want to focus on before we then go and try the much harder task of identifying and enriching and keeping current the information we have about contacts.
GERRY44:01the contact piece is actually more again of this concept of like a sub loop if you will, or areas where certain steps within this process we may we may be
GERRY44:12looking into individuals, whether it's some of the signals may be job changes or new hires.
GERRY44:18At that point, we may be skipping some of these other agents and going right to enriching the contact in and of itself. But barring that, in that concept, let's just say we now have our accounts. We have our accounts stack ranked. We know which ones we want to focus on. So we're making sure that we're not overspending any of our resources.
GERRY44:36Now, this is the point in time when we go and try to find all of the right people. And this is where we lean on our the definition of our buyer personas to really guide who we're looking for. And again, having just done this for ourselves fairly recently, you know, titles in and of itself sometimes aren't clear enough.
GERRY44:53So what's really been helpful nowadays is being able to leverage Claud for doing web research. So not relying on LinkedIn solely not relying on just titles that are available in zoom info. But looking at the our team page for executives can sometimes give some more background or some more color, or some variations on titles that are in these databases and can help make a decision about whether somebody is relevant or not within an organization.
EDDIE45:17Yeah. I just want to say like this exercise that that you spearheaded has just been game changing for us, because I think that a year or two ago, looking at some of our target accounts, the information that I had access to with the time and tools that we had at the time just pales in comparison to today. And a lot of it is just because the surface level information, the firm graphic, technocratic, the job titles, etc. didn't tell us much for us personally about the most valuable accounts for us to work, and it really highlighted that we had been working accounts, sometimes accounts that like we got to know pretty well without realizing that at
EDDIE45:56the end of the day, they really weren't a good fit for us and missing some that really were. And so this to me was just really eye opening to be able to just say, okay, even with zoom info, even with Clay, even with all the tools, we don't necessarily have some of these data points that are so critical for us.
EDDIE46:11And we could just send Claude to the website to go find
GERRY46:14these tools are powerful. The web research part is really
GERRY46:17for me, that's been the real game changer. I mean, it's not the only one, but that is the one that kind of, you know, 100 X is what any one individual could really do on their own in any amount of time.
EDDIE46:30Yeah. And I used to always think of this as like, okay, like let's say let's go back in time ten years and we say, okay, we build out all this automation or we're doing a lot of this manually. And Reb Ops is using all the tools that we mentioned to do this manually instead of building AI agents. And then at the end of the day, we like serve up a list to an account executive.
EDDIE46:46Well, what is the first thing that you do as an account executive or an STR? If you're trained properly? Well, you go in and try to tier your accounts by going to the website and poking around and seeing what you can see and like. It's pretty amazing that we can now just talk a lot to do that.
GERRY46:58Yeah, yeah, as long as you can verbalize and document what a lot of that decision making logic is. Again, it's nice because programmatically you may be really hard to say, well, it's this or this or this, but it could be ANC and A and C are much better than just A or just see again that's something that you can if you just sort of write down your thoughts on that.
GERRY47:21Claude or another LLM can think the same way you are, but you know much faster and make the same decisions at scale.
EDDIE47:29Get into the research agent. Right. Which is also core to what a lot of these ACR agents will do, as well as like the traditional tools like, like outreach that are building these capabilities. But what's so important here, and I think this is where a lot of companies fall off implementing an AI is by not thinking about all the things that we're mentioning, getting the right targets, and then really thinking about what is a good email look like and what do we need to do in the research process to write a good email.
EDDIE47:55And I'll talk about the research agent and the messaging agent together, because I think they're very, very like dependent upon each other, right. The way that I think about this is, is if you want to write a really, really good sales email, you need to start with a template and then personalize it. And I say this because there are elements of a good sales email that just persist, regardless of what you have found in your research.
EDDIE48:24For example, hi, Mame is a perfectly fine opener that doesn't necessarily need us to like, overthink that the CTA. Like, do you have five, you know, 30 minutes next week or would you like to hear more or whatever? The research that we're doing doesn't necessarily change what that CTA is, right? And so when I think of a really good email, it's usually like high name the first paragraph that opens with some research that you did.
EDDIE48:51I see that you do this, I, I've uncovered this. I liked your podcast where you talked about this, etc. and then tries to tie it into some kind of a problem. And so as a result, I imagine you might be facing this issue. All we're trying to do in this email is to show the person like, hey, we really know you.
EDDIE49:10We have either done our homework on your website and or built this giant machine that we're talking about here on this podcast today to assure you that we are reaching out to you because we really, genuinely, like, know what your business does and what you do, and we think we can help you and we have a good reason for it.
EDDIE49:27This is why this works so much better, in my opinion. As somebody who receives these emails every day, the first filter I'm looking at is like, why should I believe that you can help me? You say, you can help me and you want 30 minutes, but how do I know I don't just get on the phone and it's a complete waste of time, and you're going to tell me about all the wonderful things you did for Pepsi.
EDDIE49:44And I'm like, what does that have to do with me? I run a small professional services company, and I don't think you know anything about my business. And so that piece in the beginning of the email, the more personalized it is. Not only is it my opinion, but studies show again and again that this has dramatically improved response rates.
EDDIE50:02But the second piece, where you're usually like highlighting a solution to the said problem, for me, that's something that like really should come from marketing because it's so difficult to sit down and articulate here is like a solution to a common problem. How to describe that concisely and in a compelling manner. I don't want a sales rep or an AI engine writing that if we are going to reach out to all these people and tell them about how we can help them improve their outbound engine, I'd rather just like, carefully write that down and then use that again and again and again, then to try to have the human or the AI agent reinvent the
EDDIE50:42wheel in every email. So now what that leaves us with is we have these templates that say, well, here's problem number one and here's solution, here's problem number two, and here's the solution, etc.. And now all we have to do is ask the AI agent and or the human to go do some research on that middle part that says like, hey, I've done research.
EDDIE51:02I know who you are and what you do, and you might have this problem. I know that, Jerry. This isn't your like area of passion, but if you have any thoughts on this, happy to happy to hear you out.
GERRY51:15I don't think I can really add to this is just being on the receiving end of this kind of messaging. I am agreeing with you and sort of shaking my head quietly like, yeah, if I just get something that looks completely non personalized, I'm going to ignore it entirely. If I get something that seems to have an understanding of who I am and what I do and who I work with, give me, give it a little bit more of my attention and I'm not going to question where they got my information from at this point.
GERRY51:38It's all out there and we all know it. We've kind of accepted that. But yeah, it's really important for the effort to be shown, even if we can assume at this point that it's a combination of a machine and a human that is putting in the effort, you've got to get through, there's just going to be more and more barriers to getting people's attention to begin with, as we as we evolve.
EDDIE51:56Yeah. And I'll say like, another failing point that I see is I get a lot of these AI emails. I think everybody's getting these at this point where it or maybe not everybody, because like, I put a lot of content on in the world that makes it easy for these AI agents, find something to latch on to. But then the email will say something around something I said on a podcast or in a newsletter.
EDDIE52:15And it has nothing to do with the problem that I supposedly am facing or that solution. And so it's like, okay. And then it's also just like written in AI speak, and I'm just looking at it. It's like, congratulations that you sent your AI to the podcast. Grab the transcript that we worked really hard to make sure it's available to AI engine so that we can improve our.
EDDIE52:33You then leveraged all that information to send me an email. What does that have to do with me or my company or anything? Like you're not making a point here. And I think that this is what makes things really hard for sales reps and where we want to have an opportunity to improve. This is to say, what is a good email look like?
EDDIE52:50What does that template look like, what research needs to be done in order to do that, and then serve that up to a rep on a silver platter so that they can then go in and say like, oh, like, here's the podcast. I click on the link, I go listen to it for three minutes. It's that middle section where it's something like relevant to what we do, and then I maybe tweak the wording just a little bit and then it's, hey, you might be facing this problem.
EDDIE53:11Here's the solution. And it's a really well written at all ties together. And then I send that out without spending an hour listening to a podcast before sending one single email. And now I'm balancing this efficiency and effectiveness.
GERRY53:24Well, and I'm curious, Eddie, like a good there are you're saying there's sort of a set of standards for a good email. And I obviously recognize that and agree with that. But I'm curious, isn't there also something to be said about persona based and problem based elements of what make a good email, in other words, who you're targeting and what the problem they may be experiencing at a given point in time in the fiscal year, even, let's say, like that's going to dictate how that email should be crafted a bit as well in terms of what a land in August versus in December.
GERRY53:54Let's say.
EDDIE53:55That's such a great point. You made me realize, I think we skipped over our segmentation agent.
GERRY53:59No, no, I assume we were going to sort of circle back to it because you've got to do the research to then think about
GERRY54:04where in what segment the individual lands and depending on what kind of things you're, you're trying to promote and how what messaging you're trying to get out, which is you're working with your marketing organization to focus on that.
GERRY54:14Right now, if your organization is launching a new product, which is solving for a new problem, well, you may need to reset your audience based on the research you've already done on certain individuals.
EDDIE54:23I think those are valid points, but it wasn't my intention to skip over the segmentation agent. So we'll go back to the segmentation agent, which sits between the contact agent and the research agent. Right. And so to your point, if I want to reach out to CFOs and B2B tech companies, the problems that they are facing is going to be very different from the problems that crows and B2B tech companies are facing.
EDDIE54:46It's going to be different than the problems that cross in professional services organizations are facing. It's going to be very different from problems that CFOs and professional services organizations are facing. Now, this isn't always just, you know, title and industry. I think it's really dependent upon the product a company is selling. But whatever it is that's relevant for that, that product and that customer, we want to build these segments so we can make it easier on both the AI and the humans to say, okay, here's the set of problems that people in this segment face.
EDDIE55:21Let's go do the research and see if we can find something that ties to that so that we can have a really strong email.
GERRY55:26yeah. And working with product and marketing to rerun that segmentation segmentation agent as needed. Because again today you may submit a certain way. You may need to refine that segmentation, make it broader, make it more micro, more macro over time.
EDDIE55:41Absolutely. And we're going to get into like the back end where we like do the analysis here. But basically we've got our our set of agents that identify the accounts, identify triggers and signals, do enrichment, then score those accounts, add contacts to those accounts, and then segment the contacts in those accounts into various segments where they have similar problems and solutions.
EDDIE56:11We then send that to the research agent to do the research, and then send it to the messaging agent to draft the messaging, so that the sales rep can then either press send or look at the research, do their own, like, you know, extension of the research and tweak the messaging. So once we do this, once we've built this messaging agent, we're now off to the races.
EDDIE56:34We're ready to execute. Right. This messaging agent can push into outreach. And like if we're using clause to do the messaging or outreach has its own AI capabilities to do messaging, as do tools like artisan that are out there. You can use Sales Loft Apollo. There's lots of different tools that you can use for this, but basically by the time we have this messaging agent ready to go, we're ready to execute.
EDDIE56:57And the way that I'm thinking about this is and this is highly dependent upon the price point you're selling at and how important it is to personalize messaging versus automate things. You could take a tier one approach where you're going to personalize every single message. You could look at tier twos and say, all right, we're going to have these AI drafted, but the rep is going to review them and approve them.
EDDIE57:17And then tier three could be fully AI written. You can say tier four would be we just have like an old school template insert merge fields, send all. You could consider that as well. But it's just as we go down the list of different options, we get increasingly diminishing returns. So for me, if the price point is high enough, I just want to personalize everything.
EDDIE57:38But I also don't want reps to have to like start their research from zero. But this is case dependent upon the organization. Anything you want to add to that? Like I think that this this step is a very human step. This is not automated by definition unless we want it to be. But anything you want to add to that, Jerry.
EDDIE57:56No, I.
GERRY57:57Nothing major. Just I definitely strongly agree that this needs to have a very large human in the loop presence. That is the differentiator at this point in time really, is how well a human reviewed and audited and edited the messaging before it went out and did it thoughtfully.
EDDIE58:14Well, and the funny thing about these tools is, like, I've been playing around an artisan for a little bit, and there's nothing that prevents you from doing all of these things manually. Even in like an ACR tool like artisan, you can go and find your own contacts through arduous manual effort. Load them into artisan, create your sequence, write your own templates, and then go in there and do the research and handwrite and personalize each and every email and hit send.
EDDIE58:38Obviously, that's not what that tool is designed for, but there's nothing that prevents you from doing that. The only reason not to do that is it's just a lot more work and it's a lot less efficient. So like the point of this whole podcast in this, you know, this this outline that we're writing is to say like, well, what's think about how we can automate these different steps while still giving them the opportunity to like ultimately send a personalized email.
EDDIE59:01let's move on to the next piece okay. So the next set we're going to go through data capture CRM update, the AI data analyst management review the human data analyst and then finally go to Market Council meeting. And what are all of these things? All of these things are ultimately just trying to ask one question. Did we actually reach out to all these people like we were supposed to?
EDDIE59:25Did we do it in the right way? Did it work? What could we improve? What could reps improve, and how could we improve this machine? And so we are going to start with the data capture. Right. How do we know if something worked or didn't work if we don't have any data. So this is all kind of obvious right.
EDDIE59:46We want to capture these emails. If we're using a tool like outreach or artisan like that's just going to sync with Salesforce, it's pretty straightforward. As long as you know how to set up that integration things our team does quite often. If we can get call recordings from tools like Otter or Gong or Attention, then we want to get those transcripts.
EDDIE60:05If we want to use a tool to go and grab our LinkedIn DM's, a little bit of risk there with like the terms of service of LinkedIn. But that's another option. We can look at our calendars and find out what meetings are booked. Jerry, is there any way is there any tool that tells you what meetings are actually held?
EDDIE60:20I guess you could just cross-reference against the recordings, right?
GERRY60:22Yeah, that's that's what we do, is we leverage the transcription to determine if it was held and who who joined as well from the invite list or if somebody joined that wasn't on the invite list.
EDDIE60:31Yeah. So we want to build this agent to capture this data so that we then enable the human analyst and the AI analyst and management to look at the state and figure out what's working, what's not working, and what we could tweak anything that you want to add to that, Jerry.
EDDIE60:44Now.
EDDIE60:46That's fine. Yeah. So the next the next piece and this is highly related is the CRM update agent. Now why is this a separate topic from the data capture? Well, I think for me is it's just about interpreting that data. Right. So if we want to just like go and grab like calendar invites like that's fine. That's automation that doesn't require any type of LLM.
EDDIE61:07But if we want to say like, you know, what are the medic fields on the opportunity that we can grab from Otter or from Gong? That's an exercise where we need to leverage some kind of an AI to interpret the transcript or interpret the email to then help us update those fields, which then tells us more information that we can use to figure out what's working and not working.
EDDIE61:35As an example, we might say like, wow, we've booked a bunch of meetings with companies that like refuse to introduce us to their decision maker, or we've booked a bunch of meetings with companies that are struggling financially, and we don't want to do that. So we need to like, redo our targeting agent.
GERRY61:51Yeah, you may you may have other updates. A lot of this is going to surface from transcriptions and conversations. But you know, maybe there is key data points about an individual. They need to be updated. Somebody's got a promotion. It's not in a database. Your enrichment agent is not going to see it. But it was on a call.
GERRY62:04You need to update that something as simple as next steps on an opportunity. It's helpful to have, you know, leverage an LLM to do that. You may want to use a clod or similar to determine when to overwrite versus not overwrite certain data points as well versus data. That's just new records being created, history records, activity records, things like that.
GERRY62:26So there's a lot that can go into the thought process of how you want to make changes to the CRM. Based on all of this data.
EDDIE62:34We can also send these these updates to slack and ask for approval from reps as well. Right.
GERRY62:39That's the human in the loop part that will you know, we've mentioned a few times here. Yeah, we use slack as our, our our surface if you will, to interact sometimes not just in like the cloud desktop or web app where you select a surface. A lot of these decisions.
EDDIE62:53Yeah. And I think that this is really important because there are certain fields where it's like, look, look, let's just update the field. Some of these title changed like that's relatively straightforward.
EDDIE63:02There's other fields that are maybe you want it like mostly automated. Like if we have a blank pick field and clod can go and grow or attention or momentum or whatever tool is going to go and grab from the transcript what that is, that's probably great.
EDDIE63:17We want to limit the admin on the sales rep, but like qualifying a deal in the pipeline, disqualifying and marking a close lost, possibly even moving it from stage to stage. I think you could make a debate one way or the other on that. These are things we want to really think carefully about before we automate it. And also again, it depends on like the price point.
EDDIE63:36You know, I could totally see a situation where a company is like a plague motion and they're selling things for like, you know, thousands of dollars per year. And the volume of opportunities is just so high that reps are never going to update this stuff, and we might as well automate it. But do we want to do that with our enterprise reps that are selling like million dollar deals?
EDDIE63:54Then all of a sudden deal goes out of pipeline and they stop working it because the AI agent decided like, no, absolutely not.
GERRY63:59No, you know, machine should ever be really well, I should say. And machine should never be close loss seeing a deal closed one. Sure. If that signature through any platform, that's probably okay. But yeah, we want to be careful with how definitive some of the automation is in in opportunity updates particular.
EDDIE64:17you know for someone magazine subscriptions for 50 bucks a year like or 2012 bucks a year. Right. Which is something I did in college for like a day. I think I could make a strong argument for why we just want to automate everything.
EDDIE64:27All right. So now we push this into our AI data analyst. Right. So what's important here is the AI data analyst has a really rich and deep set of data.
EDDIE64:36And the first thing that I want an AI data analyst looking at is did we execute if we said we needed to send a series of 10 or 15 emails, voicemails and LinkedIn DMs to this account or to this contact in this account, did we do it all? Did we try and send three emails? And then the rep gave up and marked it like no response and moved on to the next one.
EDDIE65:00So for me, that's the most important thing because at the end of the day, like if we think about go to market as an experiment, if we are not executing, whether it's outbound or inbound or what have you, like if we say all of these leads converted, but the reps only reached out once or twice, we don't know if the leads are good or not.
EDDIE65:18Like we haven't run that experiment to validate, but if we reach out to 10 or 15 times or whatever it is that we have determined is the right number in the right way, then we can start to look at, okay, did this work? Did this not? And the next piece of course, I want to look at is messaging.
EDDIE65:32And I think it is fascinating. Now granted this can take up a lot of credits, but I think it's fascinating that we now have the ability to leverage an LLM to actually look at the messaging and say, was this messaging good? This is something that we have built internally and it's helped us onboard new sales reps. It's been incredible in the sense that, like, it's impossible for a human to go look through like hundreds and hundreds of emails and read each one to coach a rep on how to do better messaging.
EDDIE65:58But an AI can absolutely do that.
EDDIE66:00Any thoughts on that? Jerry, I know this is something I was closer to know than you.
GERRY66:04Yeah, I know you've been doing that a bit recently. No, I don't think I have anything specific to add to that. The only the
GERRY66:11You know, you could also include in this the concept of a coaching agent that can be reviewing and providing first level sort of triage and feedback to reps.
GERRY66:19But again, that's kind of going to go hand in hand with a well-defined process of what that means when the human delivers the feedback to another human versus when a machine should. But that's sort of a sub loop that you could have in this process.
EDDIE66:33That's a really great point. I might even add this to the graphic before we publish it, but as of now we have the AI data analyst feeding the the the insights directly to management for them to review. So I think one of the use cases of AI coaching reps is, is one of the most interesting and strongest use cases, like folks are talking about right now.
EDDIE66:53So love that point. Maybe I'll update the graphic, but as of now, at the time of this recording, the next piece is the management review. And the idea here is, is that management goes in and reviews the AI, analyzes and then starts coaching reps. Let me give you a tangible example from what we're doing internally. We have this agent that then goes that goes in and looks at actual sales messaging and scores each and every message across multiple criteria like is this personalized?
EDDIE67:19Did we do our homework, is there a clear CTA, etc., etc. all the best practices from all the folks out there like John Barrows to tell you how to how to do good sales emails and then it stack ranks them from best or worst, and then it puts it in a PDF that I can actually, like, skim through and read really fast.
EDDIE67:35And then it shares all of this, this coaching feedback. So I could literally just send the PDF to a sales rep and say, hey, this is how you scored today. And I've been doing this and I've seen dramatically fast improvement because of this, because it's just very, very clear feedback. And the AI is not perfect. Like there's a lot of nuance in here.
EDDIE67:57And so I'll say, hey, you know, like, don't worry about this too much. Don't worry about that too much. But there are certain things that are like really clear. Like, hey, a sales email should only be so many words long. If it's a novel, it's probably not going to get a response. And I've just been blown away at how valuable this is because I've tried to do this manually in the past and it's virtually impossible.
EDDIE68:15Yeah. So. Oh, sorry. Go ahead. I don't know if you had anything to say on that or not.
GERRY68:19No, I was just saying, you know, one thing that we have done as well, when we're thinking about sort of the messaging, scoring is also like highlighting specific things,
GERRY68:27not just sort of overall feedback about the messaging itself, but, you know, being able to surface specific words or not. Maybe not words, keywords, but phrases that can be more meaningful for a rep than just saying, hey, this was a good email.
GERRY68:38This was a bad email. The other thing I wanted to say, because you're talking about scoring, and we talked about scoring a lot throughout all of this. When we're letting an AI agent do any sort of scoring, particularly when you're engaging with other folks on your team, you need to be able to, I think, give some background as to what drove the score.
GERRY68:55If it's just an arbitrary 1 to 100, you know, depending on what kind of feedback, if it's good or bad, if it's negative feedback, let's say that somebody is on the receiving end of they're going to be highly critical of that score. And if you tell them that a machine helps score it, they're going to be really critical of it.
GERRY69:08So it's really helpful if you can work with the tool that you're using to do the scoring, to come up with a little bit of a rubric that can be shared. Hey, you know, a 1 to 20 is because of X, Y, and Z where things that were missed, whereas a perfect score is because you did all 12 of these things.
GERRY69:22That gives you a little more credibility when you're leveraging a score that a machine came up with.
EDDIE69:26That's a really great point. And that's exactly what we've done with this internal agent that I'm talking about here, where it says, okay, and I don't remember all the criteria, but it basically says like, here's a score on your subject line, on your homework, on the problem, identified the solution, identified the CTA, the number of words the email is and the reading level.
EDDIE69:46And so it gets really specific of like, well, you know, if the reading level is you have to have a PhD to understand this email, you're going to get dinged on that, right? If it's written at a fifth grade level where somebody that's very busy can skim through it and understand it like that, easy. If it's over 120 words long, you're going to get dinged, right?
EDDIE70:05If there, it looks like you didn't do any homework. It's just like, hi name. We do this. You're going to get dinged on your homework and your personalization. But to your point, Jerry, like the way that I was thinking about building this was not only leveraging best practices, but also being very clear about here's what's good about this email, and here's what's not so good about this email for those best practices.
EDDIE70:27And even then, so far, the current iteration of this agent is it's overly critical. So my message to sales is like.
GERRY70:36Take this with a grain of salt.
EDDIE70:38Take this is a grain of salt. But here's certain things like don't write an email that's, you know, a thousand words long. They're not going to read it. Yeah. And there's also things where well, this is what perfect looks like. But there are times where you try to do a research on somebody and there's just only so much you can say.
EDDIE70:55So anyway, I don't want to belabor this too much, but really great points, Jerry. Yeah. So the next piece is and you could argue that the AI analyst should send this to the human analyst before or at the same time as sending it to management. But the one thing about management here is it's like management is so important for execution.
EDDIE71:13If the analyst basically just says, hey, look like the problem here, like this isn't working. And the problem is, is that we're sending like 1 or 2 emails to each contact and then giving up. That's kind of the end of the story. There's not much more to say there. And that's where management really needs to step in and drive execution.
EDDIE71:29Sending it to human analyst. Just reiterate the same thing. Like, hey guys, like I don't know what's working and what's not because we're not actually doing the thing, but assuming that we are, I think it's really important to have a human actually look at this in conjunction with Claude, where we can ask a lot of questions, we can go deeper.
EDDIE71:47But the purpose of this is, is to prep for our final step, which is our go to market council meeting. And so the idea here is, is that the human analyst will go in, look at all of these things, everything throughout this engine, who we're targeting, what signals we're using, how we're doing segmentation, the emails that are going out, the research that's being done, the number of meetings booked, pipeline, etc., and look and see what's working, what's not slicing and dicing in different ways, which also the AI analysts can do and say, well, how is this working by segment?
EDDIE72:19How is this working rep by rep? How is this working team by team, product by product, whatever way you can slice and dice this. And this is where AI becomes really, really powerful. Now all of a sudden you're surfacing up these insights, then having a human look at it and then preparing for a go to market council meeting.
EDDIE72:35And this is the final step where we say, okay, every month or every week or whatever we want to do, we get all the go to market leader leaders together. This is usually run by rev ops or the the analyst whomever to come and say, hey, these are the insights that we have uncovered by doing this analysis. These are the things that are working really well with inbound.
EDDIE72:56These segments are converting really well. This messaging is converting really well. We're getting a lot of meetings after like the eighth attempt, which you keep doing these things. Here are the things that are not working so well. And then we discuss it and we might say, you know what, let's double down on this segment. Let's update our targeting agents so we can find more accounts that are and contacts that are in these segments.
EDDIE73:18Let's update our research process to identify these things. Let's update our triggers and signals. Because these particular triggers and signals are the things that are resulting in us getting more meetings, pipeline and close one business. And there's no one size fits all to this approach. But what I think is so powerful about the go to market Council meeting, and something we do internally, is that it's this forcing mechanism that requires the humans to a get in the room and be prepared for that conversation and see discuss what to do about it, and actually ideally update the roadmap and then go back to this engine and redesign the whole engine so that it's continuously improving.
EDDIE73:55Now, Jerry, this is something that we do internally. And you run this like any other thoughts on this. Like what's been your experience being forced to do this every single week for however long we've been doing it now?
GERRY74:06No, I mean, I will double down on the what we mentioned earlier about management having by in on this. It really has to be a top down emphasis on like the importance in reviewing this data and gathering the data and enforcing the process and then making decisions based on this. If it's not coming from the top down, rarely do you have reps who are just as hungry for this and will want to self improve based off of this.
GERRY74:35I just don't see that all that often. So I think having this, this sort of brain trust of folks who are going to be able to not only discuss and analyze and plan, they have to then be willing to go and implement an action on it. And that's where we often come into play, is because we see a lot of organizations that want to do these things, but don't have the bandwidth or don't have the capacity or the experience of how to juggle the changes that need to come out of out of the council meeting, alongside other priorities that some of these leaders have in their day to day.
EDDIE75:10Yeah, I think that that's the issue is every time we do this internally, it just spits off more work where we got to go update this dashboard. We have to update this workflow. We need to change this process.
EDDIE75:20And it's frustrating, but it's usually always coming from a place of well we either are looking at the data and it's not accurate, and we need to do something different to make it accurate or it is accurate, but it's not the data that we want to see.
EDDIE75:33So we need to go look at the process of the execution to go change that number into the number that we want to see, not on paper but in reality. And and that's not easy to do.
GERRY75:45Yeah.
EDDIE75:45But we are way over our typical time for a podcast. Jerry. What a what a dance topic. Any parting words on this?
GERRY75:53Keep a human in the loop, and don't treat this as like a set of and forget it way of approaching how you're going to run outbound. Every single element of what we talked about today warrants revisiting on some regular iteration or interval, I should say, whether it's weekly or monthly or quarterly. You know, if you feel like you're at a point where your business isn't evolving, the market's not evolving, your messaging isn't evolving, your team's not evolving, then you must have already figured it all out.
GERRY76:21And you don't even need to be listening to this podcast. It's so good for you.
EDDIE76:23Yeah, I couldn't agree with that more. Well, thank you for taking the time to do this podcast with me today, Jerry. And thank you to all the people listening.
GERRY76:31Thanks for having me.
EDDIE76:32Thanks for listening to the episode. If this resonated. Please give us a five star rating and a follow. It helps us reach more people, and you get our latest and greatest content without having to search for it. And if you're looking for hands on help and go to market strategy and or rev ops, please reach out to us.
EDDIE76:47We help our clients with everything from annual planning to improving processes and go to market, implementing systems to support those processes and go to market AI. We're always happy to offer a free consultation to help you identify the best opportunities to improve your go to market engine, with or without our help. You can find us at Union Square consulting.com and the info will be in our show notes.
