#6: How to use an MCP to act on live GTM data (and the play that shows how)

(upbeat music) - All right, everybody, welcome back to the Revenue Architects podcast, where we get into the nitty gritty details of what we've seen, break down with AI in the tech space and in GTM specifically, and what you can actually do to fix it. I'm your host, Florentutalia, the GTM engineer in residence here at ZoomInfo, and here's my co-host, John Lloyd. Welcome back, John, how you doing? - Good to be back. - Great, you were just overseas, were you not? - I was indeed. - All right, so you were in Spain when they won the World Cup in New Jersey, right? Was it just a constant party over there? - Yeah, I was, I'm in Europe in the summer, it's always one of my favorites. So it was out in Spain when they won the World Cup, got to Portugal for some weddings, which is absolutely beautiful. So you can't really beat Europe in in July, and I'm feeling good back in action here, and definitely feeling energized. - That's amazing. Well, ready to get going. Let's talk about MCP and maybe some hype or over hype today. - Yeah, 100%. I think this is gonna be an interesting episode. I know MCP is a term that comes up quite a bit, whether it's on LinkedIn or just in the space. I think some people have a very good idea of what it is, other people kind of just nod, and you're not really sure. If they actually understand what it is, so we're gonna get a bit into what MCP is, why it is helpful in GTM, and where some of the shortcomings are, and then we're gonna get into some place to show you exactly how we would use MCP to add zoom in forward or just in general with some of our customers, and John's gonna be walking you through that. But for the data point today, as you all know, we asked 50 senior GTM leaders, which AI use cases in GTM are either the most overhyped right now, and which are genuinely working. So on AI agents specifically, querying GTM data via MCP, we had 52% said that they were not sure or undecided if it was helpful. And I believe only 16% said it genuinely was working. So pretty much like nothing else in this survey left more leaders on the fence, and I think it's important to make sure that we get this framing right, and give you some different examples of where MCP works, where we've seen, again, some of the shortcomings, and then we'll get as tactical as possible here. Anything else you wanna add there, John? - No, I think the hesitancy about MCP is coming from people that haven't tried to use it yet, I think. And in my experience, there's just so much going on. There's the MCP connection, there's the API connection, people are confused about which one to use, whether they shouldn't use either one of them and just go to their SaaS platform, and use the AI that they have there. It's just, I think it's a lot of noise that's hitting them, and people don't know what to focus on and try to scale. So I agree that I've seen the data point, I've seen the evidence of that data point. I also think that it's telling that some of these other data points, going through the research, we found that 62% of people said that AI account research is genuinely working for them. I would expect that number to be a little bit higher by now. So a lot of these areas where some folks would think would be, I guess, a little bit more mature, are just like, still, people are hesitant about them. People aren't diving in and using them as much. I think it's because evolution is moving so quickly in the market with AI. - And I think, when you think about where the world is going, I think it's pretty clear to me that, in terms of an interface or how we interact, with AI has obviously changed over the last little while, but to me, it's pretty clear now that the interface is going to be this like chat type of interface, chat GPT, you know, cloud, whatever that might be. And I do believe that it is going to move more and more towards voice, so where we're not even necessarily going to be in front of a computer, and you can just be, you know, on a plane or at the airport going between business meeting and the business meeting and just interacting with that AI agent or via MCP, essentially, enqueering whatever data you need. And obviously the theme of this podcast has been the context layer and making sure that that, the data layer beneath the AI agent and where it's pulling from is super clean and has all the context it needs, identity resolution, we talked a little bit about in the past two. So we'll get a bit more into that, but I do generally believe that, instead of going into a UI for any type of software to look at GTM data or to help you find information, this is the reality that we're going to be living moving forward. So if you haven't queried data or used in the AI agent via MCP, I think it's something that every single person will probably do by the end of this year. No questions asked. - Yeah. - I'll give you an example too. Like the first time, when it actually works, it is quite a big wow moment. And John, I don't know if you had a specific moment where this kind of like clicked for you, but I remember when I was at my last company, a common room, one of the best use cases we had where we were paying AI via MCP was when we were building out account research agents and the depth of agents after we had very specific prompts. The depth of the research that we got within like 30 seconds compared to a human having to go and try to query this data in four different platforms, was such like a game changer for us and saved us so much time, especially on the SD on the EE side. So I'm really excited if you haven't had that, eureka moment, I think you will have it as soon as you get your data in order. - I think my eureka moment was realizing that through and this goes for any folks who use like conversation recording software like Gong or Corus, was that I could pull and basically summarize like transcripts from calls over the past three or four months and just basically, and this is kind of like the play that we'll go through today is identifying flag accounts that are maybe having some pain points that have been resolved since then or just really just like scale and beef up your data points for like re-engagement calls and re-engagement emails and just like giving yourself the time back that you would have taken even going through the recording software itself or asking the AI to summarize the transcript, like you can do that at scale across the board and that was my eureka moment. It's just so much easier and you can talk to it. - Yeah, 100%. You're prompting, that's all you're doing. - Yeah, and I think maybe like one last point here before we move on, I think one of the reasons that a lot of GTM leaders are still somewhat undecided on this topic is because again, they don't trust the underlying data, so sure you get all this information, you can, you know, through one voice prompt or like you're writing a quick sentence and funny enough, this actually just happened late last week. I saw it on LinkedIn, but Clays VP of Sales pretty much was querying their database for some data and then the user gems, I think it was one of their co-founders actually like called it out and pretty much said like, hey, of the five contacts that your MCP, that your AI agent pulled via MCP, like three of them or two of them don't even work at the company anymore. Again, this is like a data foundation problem, so it's just funny to see that like these different platforms in our space or the companies that are building these context layers and the AI agents that GTM teams are using still don't necessarily get it right. And I think there's a few different reasons why I think that happened in that specific example and to the VP of Sales credit, he came back and kind of addressed it and showed how he fixed it within 10 minutes, but to me it goes back to two things which is the main problem as to why people don't deeply trust AI, one, sometimes with prompts, like you have to get very specific in what you're looking for, so even if you miss one or two sentences, it's like, oh, like I'm looking for a VP of RevOps in this specific category, the system can either hallucinate or it doesn't have good enough instructions to know by, or to reason by itself what contacts it should give you in the case that that VP of Sales mentioned, but two, and I think this is where, obviously I'm gonna be a bit biased with Zoom info, but as someone that's worth in this GTM text space now for almost three years, that identity resolution and making sure that you have the same accounts in contacts tied together across different data points and platforms is gonna be super important. So we talk about the context layer identity resolution. That's what makes the queries naturally helpful. You need that underlying layer. There's no question about it. We've talked about a lot of this podcast. - Yep, I think some of that hesitancy comes from just maybe distrust in their first party data versus. their inability to use the AI or their lack of knowledge and how to use it, because MCPs are pretty simple to use. Once you connect it to whatever LLM you're using, it's just like you're off to the races. Maybe just for a bit of background, just to define what an MCP is because I know we use it a lot and I think probably half the audience knows quite well what it is, but in the most basic sense, so at MCP, it stands for model context protocol, so it's just a standardized interface layer essentially that lets AI applications connect to external data and then you're pretty much just querying tools through a universal language. And one of the best analogies I've seen for this so far is think of an MCP as like a USB-C port for AI. Yeah. So if you all remember before USB-C, whether you had an iPhone or whether you had like a Samsung phone, you needed different cords to plug into our adapters. So now with USB-C, you actually just have like a universal standard, so every single phone can plug in with that. So MCP pretty much is doing the exact same thing for AI, so instead of building a custom brittle integration that speaks to every single platform in a different way, you just have this standardized language now. Yeah, and I think that's, you know, it's a good comparison with an API because I think a lot of people might confuse those two, so API obviously application programming interface. The API is more of a developer is pre-defining every connection, right? It's a little bit more, it's a little less agile. There are fixed endpoints. It's a custom integration per the service that you're using it for. The API is a little bit more dynamic. The AI is deciding basically what information to call and to give you. The AI is like asking you, hey, you know, what can you do automatically? You're connecting once versus reconnecting, and I think the MCP is a little bit more user-friendly as well, so. A hundred percent, and I think another big breakthrough for the MCP. I think the best way to describe it is like being grounded in real-time data. So if you didn't have an MCP, like you could do an integration or an API, you could pull data from a CRM, from product usage data, whatever, but you're not sure that that information might be static in a CRM, and it's not currently up to date. So by actually having grounded data, you're making sure that it's the most real-time data, which is super important when you're thinking about GTM data, whether it's signals, whether it's like telemetry and product usage data, whether it's from the CRM, etc. So that's like the big difference from you with the MCP versus any type of like, you know, custom integration into a specific platform. Yeah, that's a good call out. Sweet. So let's actually go through a play now that we're going to query an AI agent with or via an MCP server. So John, tell us a little bit more about the play here today. Yeah, so the example we're going to be going through today, and this kind of harkens back to what my Eureka moment was with discovering the usefulness of the MCP is a churn risk prevention play. So the problem is that churn rarely announces itself in one system, you know, sentiment between, you know, you and your client, you know, might might sour over the course of many calls, it might be one call, you know, usage dips, replies via email, turn a little bit more terse. And your renewal date is sitting above you like a doomsday clock and you don't really know how to reach out to these accounts at scale with the level of specificity that they need to avoid that churn. So this play really can use MCP to give you the tools for a full outbound, you know, go to market play motion by ranking your accounts based on a risk score with just a few prompts. And I think, I think this makes sense to kind of turn into an infographic. We might have already done this, but this is for SDRs, BDRs or revoxlators. You could be anywhere, anywhere, you know, on the hierarchy of a sales team and find this useful. So again, the point is you're trying to give your book of business a gut check. You find it difficult to get a bird's eye view of the accounts that could be at risk of churn, especially those headed for renewal. So if I was to launch this play right now, my first prompt would be to pull accounts from my book of business. I'm the owner and or I'm pulling accounts from my sales reps books of business. That include relevant first party data, right? So a lot of time the MCP can either connect directly through, you know, your CRM or through a software as a service platform like zoom info that counts to CRM data that has a native integration with your CRM. So you would basically like pull together your book of business. And maybe like first party data, I would include is usage data, right? What products are they currently using? What products are they not currently using? Maybe look at their ACV and TCV to help prioritize them and filter those accounts to ones that are approaching renewal. So you can do that in a prompt or two. You can basically say, hey, you know, I want to, I want you to give me a rundown of my my book of business. I want you to only include accounts that are, you know, 30 days away from renewal, let's say. And then the LLM through the MCP is going to understand that request and give you a breakdown of your accounts. So that's that's problem number one. That's kind of setting the stage with first party data prompt number two. And this is my preference. You can interchange a few of these, but you would, you know, again, if you're using a call recording software like chorus or gong, ask the LLM to give you specific data points from the recent calls you had with those accounts that you just brought in. Ask it to highlight pain points and only give you accounts that have identified, you know, problems or risk areas via live calls. You can do this with emails too. So that's number two. I think that's the biggest asset that you can use for a reengagement campaign is like literally baking into an email like, hey, here's what this client said to me three months ago, this is why they're pissed off. And this is why I want to win them back. Like that data is gold. The prompt number three, I would say, is asking the LLM to pull in other third party signals. Like again, like competitive intent through Ziminfo or, you know, G2 or trust radius views. I want to know if the account at risk is actively researching, you know, the product, our product, my product with other competitors because that might add them to another tier of risk if they're actively looking to turn. And then prompt number four would be to score these accounts on maybe a maybe a three tiered scoring system, maybe a one through five, maybe zero through a hundred. I would say like, you know, highly critical medium critical or low risk, something like that. You can ask the LLM to take the call sentiment, their first party product usage, their renewal proximity and their competitive intent research as like a weighted scoring model. Like you can use the LLM to create that calculation for you or you can tell it, hey, I want, you know, I want an even split or I want 30% weight to be placed on call sentiment, 20% on email sentiment, so on and so forth. So you can basically create a scoring system yourself with the LLM. And then you get, you know, that ranked prioritized list of accounts. That's that's the leading into the final prompt here is like, you know, prompt five is draft draft emails to, you know, the top 10 accounts that landed in the critical risk bucket. And then you have a full sales play through five, you know, five sentences, five paragraphs that you've sent to this LLM. Like, you know, you can pick your most at risk accounts and immediately draft the outreach to them. And then, you know, you can, you can send those to a cadence or a campaign and, you know, kick off your outreach. And I would say an optional prompt, what I would do is as prompt number six would be, maybe create a dashboard for, if you're an SDR or a BDR, create a dashboard for your, for your sales leader and let them know, hey, on Monday afternoon between, between one and two PM, I created a sales play. And here are all the accounts that I identified as high risk. Here is what I am doing to win them back. And here's the first email. I've already drafted it. So like, you're not only creating more scalability for yourself and your company, but you are, you're, you're, you're being specific. You're, you're developing hyper personalized, you know, emails and call scripts for yourself. So, and you're, you're promoting your own ability to use an MCP in a successful way. So to sum it up, the churn risk play, I think, is a really popular play. I'm a lot of, a lot of of clients. that I've worked with have used it and found success from it, so that's the number one use case I would give, but there are so many more that we will go through in future episodes. Florn, what are your reactions that have you ever used, like a churn risk calculator in your life as a former, former salesperson? Yeah, 100%. I've never myself been in account management or CS, but this is one of my favorite plays. We've done some variations of this at most, at the last two companies. I've been at and it's been super helpful. And one of the reasons I love it so much is because one, whenever we talk about GTM plays and GTM tech, I tend to find that there is a bit of an echo chamber, especially like in the sales world. So there's so many use cases that are highlighted on LinkedIn or on different social media platforms about the sales side of things before accounts actually closed. But as we all know, I think making sure that we're managing risk is super important. I think it's something that should be talked about much more in this space because especially as companies are growing, ultimately like expansion is super important, not just managing churn. And I think there's a lot of fantastic data that are in GTM tech systems that can help you not just prevent churn, but actually expand accounts. And another point I wanted to double down on here. So there's a lot of platforms like Gainsite, there's the CSM type platforms that do churn risk scores. But John touched on this briefly and I think it's super important. Churn rarely announces itself in one specific system, right? So again, real-time data is super important here. We talked about this earlier in the podcast. So sentiment is going to sour on calls. Usage is going to dip in your telemetry platform. There's going to be replies and emails that where the language is starting to change a little bit, right? And then there's obviously going to be a ton of data in your CRM. So via MCP, you can get this real-time data and it's all in one context layer. And I think that's the advantage of doing it in these types of systems as opposed to just having something in one of those CS platforms, although obviously they're fantastic as well. And I'll give you another example. So one of the two metrics that we found super helpful in preventing a lot of risk, especially on my last company, when you ended up seeing usage spikes dropping in the product, obviously that's something that any one platform could tell you. But then another really key indicator was champion job changes. So if you have these like two things happening at once, we usually would have like an SDR that worked on customer accounts as per or or account managers. So when multiple, when two of these metrics would dip, that would automatically send an alert and obviously some context to the reps. And then we can get ahead of it and make sure that we're not just reacting, but we're actually proactive about churn. So fantastic play here. So obviously, John, the next question here is like, what is going to make all this possible? And I think there's kind of like a few different bullet points that we should get through here. But first is like, what you know, what do you actually need to build from a technology perspective? So as we've alluded to a little bit here, an MCP connection is ultimately only going to be a strong as the later queries. And I think a lot of people have had experiences, maybe not necessarily with like GTM data, but when you think about, you've probably seen examples on whether it's like extra different social media platforms. It's like, oh, I asked, I asked the AI agent, like, what, you know, what 2+2 wasn't it gave me like five or something? And it hallucinates. So I think people naturally are going to be a bit, a bit skeptical of data until they're, they're very confident that the layer underneath is going to be really strong. And they have like multiple examples that they've kind of gone through to make sure that that is the case. So with this specific play, you're going to assume that your MCP is already going to be connected to conversational intelligence. Obviously, the product usage feed, the emails that you are sending through Gmail or whatever back and forth, obviously your CRM contract data. And I think underneath that sits the same requirements as pretty much every play in the series. You need result identity. So a fused risk score only means something if the account itself is the one entity across all those different four systems. Or otherwise, you're going to have similar to some of the other like outbound pleas we talked about. You're going to have information in different accounts, different contacts. And you're not getting the full context. And that's going to become a problem when you're actually when the system is calculating the score. And as with anything, data needs to be verified and current. And we already know the step that 70% of contact data does get old quite quickly. So that context layer is going to be massively important. So whether you build that layer yourself or actually connected to one that already exists, I think the requirements ultimately don't really change here. Yeah, the pillars of the MCP are your data, the connection to the conversational intelligence platform, which probably a little less likely to hallucinate because it's pulling from transcripts. But you should still still give that a gut check. So yeah, you're totally right. And sometimes you need to specifically ask it not to hallucinate or to not treat a null value as an opportunity to fill a blank with something that it just pops into its creative mind. Because that's where I've come into issues before is like, hey, if you know, if four out of five accounts have, you know, product usage data for my CRM and one of them doesn't, it might try to fill in that last gap and you don't want to do that. So be really specific with your prompts and check the information. You always need to do a layer of QA. I think that this still deserves a human in the loop. Yeah, 100%. And I think I talked about this concept a little bit earlier, but think about like the grounding test for AI. So I think there's two very important questions that you need to ask. So one is, can you actually trace a specific output back to a verifiable source and then when your underlying data is going to change, does the output actually change with it? Right. So that's like a test that you need to put in there. So I think if you get 3S is there, then you're grounded. And then if you get any type of no, then obviously your agent might be guessing and you need to make sure that that's something that you're fixing. Yep. Yeah, definitely need a gut check. I think that's a good test. And then, you know, take a beat. I think there's one caveat here. You know, grounding an agent in CRM alone is really grounding it in, you know, what your internal team has captured. Some of that probably is manual. And I know we've kind of went through this before, but like B2B data decays. And it decays quickly. So if you have that third party source that's connected in with the MCP, that is key, right? You know, using that platform that is constantly updating with refreshed contact data, with refreshed account data, like use that data to your to your advantage. And don't just assume that you need to pull data from your CRM the entire time, right? Like use it for enrichment. Make sure that you're gut checking the list of contacts that you're pulling from CRM with the data that the third party data provider, you know, that you're using has. So I think that's another area where you can kind of, you know, give yourself a QA and a gut check. 100%. I think that's a really good point, John. And the last two final things that I think are super important to making sure this type of play works is one change management, people in ownership, and then two is going to be processes. So how are you actually thinking about implementing this? And again, it can be daunting to try to get that whole process set up in the beginning, but I don't think you have to do that. You can do it step by step. So I think one with people in ownership. So the play should have its owners, right? Like who actually owns what part of the play? So AM is going to be running the daily triage. You're going to have CS leadership, which is owning very critical, you know, save, save plans, essentially inside some kind of SLA. I'd probably suggest 24 or 48 hours for that. And then you need somebody that actually owns recalibration, so log in the outcome. So is this, are these risk scores actually indicating true turn risk? And if that's not the case, then obviously the the prompts and the scoring system, or maybe like what actual data points you're using in that return risk score need to change. So yeah, I mean, to kind of summarize that, the rollup fails if all of this is just pretty much landing some kind of channel and like no one's actually taking action or reads what's happening there. And then the last point, which I mentioned was around process, so how to implement it. Again, I don't think you have to boil the whole ocean in this point of view. What I would recommend here is start with like two very simple signals. I think when you're thinking about turn risk, usage data plus like renewal proximity is going to be super important. And then run daily prompts for a couple of weeks, try to actually save one deal. And then when you get that, you then have like a business case that hey like this is something we should implement across across the whole organization start to think about other signals that you can add inside of that churn risk score and I think that's the best way to start that's usually how I think about building out any specific initiative right starting with step-by-step yeah so for the 52% of people who are not sure or undecided about the usefulness of MCP I think the message from both of us would be to dip your toes in you know to the shallow end at first don't dive in right away don't boil the ocean I think it's just something that you need to try to convince yourself how useful it is but yeah I mean I think that that's a that's a good that's a good closing note foreign anything that you would want to leave people with in terms of like a specific action item yeah of course so I think just that want to recap this episode I think we saw that obviously about half the market is not unsure or uncomfortable to call an AI agents via MCP for a number of different reasons so hopefully today we give you some very concrete examples of what you need to do with underlying layer to make sure that it is running properly and we provided you with a nice little play here to build out a turn risk score and make sure that you're getting ahead of predicting churn in your book of business so in terms of my challenge to sales leaders so if there's one thing that I'm going to challenge any GTM leader to do differently on Monday after hearing this ask yourself or the people on your team maybe do a bit of a brainstorming session what four signals would actually predict churn in your book and think about whether you would actually be able to see those four different signals in one platform today or whether they're still you know scattered across different places in which makes the context much more difficult to find and once you have that again don't boil the whole ocean start with a few signals build out that score test and iterate and I think you'll be in a much better place thank you to John and every single person for listening please subscribe share with whoever owns retention and renewals in your GTM function and we're looking forward to seeing you on the next episode thank you so much see everybody