Latest / The Chief Customer Officer Podcast / EP013: Building AI Infrastructure That Actually Works

Transcript

EP013: Building AI Infrastructure That Actually Works · The Chief Customer Officer Podcast

• Jeff | GrowthCurve: Okay, welcome back to another episode of the chief customer officer podcast. Uh, today recording date is April 14th, 2026. This will be released on Thursday, April 26th. Uh, today you got me Jay, uh, coming at you solo. Jeff is out on vacation with his family. Um, enjoying some good time, uh, some beautiful weather down here in the Charleston area, uh, clear skies. We actually need a little rain. So, um, if you, our end-to-rain dances send one our way. wanted to talk a little bit today about some of the stuff that my and I are working on at Balboa, some of the things that we're building. And then as a result of that, I'm going to share some insights that we've been able to gather from some of data over the past ⁓ couple of just in terms of what we're seeing as we have conversations with many SaaS companies, product oriented companies, companies that are trying to transform digitally. We have big industrial services clients all the way down to medium and small sized SAS company. So we see a big swath of the industry right now. And we've gotten some interesting insights from our work there. but first of all, I'll tell you a little bit about what we're building. like many of us, We record every call that we have and we're a consulting firm. we just under 20 people in the firm and are on, we're all We work remotely with our customers. We're on ⁓ zoom all long. And so all of our team members utilize Fathom ⁓ is our call and call intelligence system. The actual product that records the calls, stores them, provides playback capabilities, so on and so forth. But what we found in using that system, and many of you all probably use systems like this today that are just fantastic tools, but a little bit harder to use really capture and aggregate key insights across all of the ⁓ and conversations that you're having with customers. And so we set out to build basically a customer intelligence system. We call it our call intelligence system because that's just one aspect of ⁓ sort of our centralized intelligence hub that we have for our company. But what we're doing is taking all the data out of the, all the transcripts, all the summaries, the participants on those calls. We're actually taking that data out of Fathom and we're loading it into our own central intelligence hub, which is a database. It's, it's it's a, a Postgres database. If you're familiar with that, it's an open source database system. And then there's a vector database that's associated with a vector database is basically help you do semantic search across your data. And it's super helpful for and performance when you're, when you're looking to incorporate large amounts of data into a, an AI driven type search. So we've got that. And the reason we did that is because for a couple of reasons, ⁓ number one, Everybody's having calls every day ⁓ some of those calls are sensitive. have leaders having individualized, you know, calls with individual on our team. Those calls are private and confidential. We don't want those details available in our search system or in our, from our tools are like our Claude tools to be able to query openly. So we have to actually provide a layer of security on top of some of those calls. the other, the other reason is speed organization. So, while ⁓ yes, you query across all these, these, ⁓ you know, data sets now with tools like Claude and chat, GBT, the challenge that we run into is, one of efficiency and token usage and, speed in terms of retrieval. And so, you know, my background is in data warehousing BI. I did that for years and years before I. got into management and sort of leading teams and sort of stepped away from hands on development of those types of systems. But I guess the, me, the, the concept of aggregating data and sort of preparing data for the use cases that it's intended for have never really left my head and my heart. And so, it's near and dear to me to be, to, in sort second nature to me to want to aggregate information. in a way that's going to be useful downstream for the types of queries are going to be run against it. So all that to say, you know, we've centralized a lot of this data. we've run models over it to determine what is restricted, what's not, and sort of layer on a layer of security onto what's available from ⁓ our agent, our agents basically, or, or from Claude. And then, also ⁓ put it into a storage mechanism that makes it very easy to query semantically. And what I mean by semantic query is meaning is captured across chunks of data. if I a sentence in of our call transcripts about pets, for example, and I start talking about dogs and cats, the semantic search might pick up a similar conversation around having pets that are goldfish and I don't know, gerbils and guinea pigs because all pets, right? And so that's the way semantic search works. It's a little bit different than fuzzy search ⁓ or, bullying searches that, you know, many of us may be used to from, from using all the tools we've used ⁓ over the of our careers. So anyway, but that semantic search actually gives us, me, a layer of, It gives us a layer of indirection that makes it easier to find things in our database after we've stored them. So one thing. So the Call Intelligence system, we've actually put an MCP server on top of that database, and we exposed it to our team instance. So now everybody has access to our Call Intelligence via their quad subscription. Importantly, what we'll, you know, What I think we've all started to do is create tools ourselves and use AI in a very specific way to ourselves. But one of the things that I'm increasingly trying to figure out with my team and we built this system that we built is how do we now take all the context data that we have, all of what we're learning, the documentation centralize it in a way that the entire team gets benefit from it. It's no longer good enough in my opinion. to just give everybody access to Claude or Microsoft Copilot or ChatGPT and not having a centralized context repository, or you may hear that there's another term floating around called harness right now, which is really of gaining a lot of steam, but not having any harnesses that are specific to own data sets and the types of information and knowledge that we need access to for our particular context. A large language model is only as good as the context you get it. Right. And so that's, that's sort of the goal here is to start to centralize things that our entire team can use and democratize the information that we have as, a company. So that's our call intelligence system. So far it's working really, really well. We've actually combined it with Claude skills and ⁓ skills basically Markdown files, which are like plain text files that allow you to document essentially how a thing is done. have a skill for writing a proposal. We have a skill for doing a discovery call follow-up document. Lots of skills around our project management work and how we engage with customers and how we coach our customers and how we even coach our teams. We've built skills for lot of those things. the ability to pull a call transcript from our call intelligence system with a skill and very quickly you can turn a conversation that you had on the phone or on zoom with somebody into an artifact that you can share. And that artifact doesn't look generic. doesn't look like a, you know, a chat GPT or claw generated generic document. It's actually formatted and templated the way that we want that to be for that, the purpose that it was intended. So Um, it actually looks and sounds like it's coming from me as opposed to from jet chat, GBT or from Claude. Um, so that's our call intelligence system. Super excited about that. It's, you know, it's working. We're, getting a lot of value out of it. One of the next projects we're going to do there is, you know, there's this idea floating around as well around autonomous agents that sort of continuously run in the background, iterate on your data. organize, enrich it, use to inform how they're organized, and then make updates to your entire system. you know, for those skills that I mentioned, what an autonomous agent do is continuously scan ⁓ our call transcripts, which coming into the system every day. And with every new call transcript, it could update the skills that we use. could identify best practices that people are using out in the wild that I can't even see or that our leadership team can't even see because they're buried in a call somewhere that that we aren't on. and then incorporate that into our skills and even surface it so that we can review it before, if we need a human in the loop on that. So, really interesting, when you start thinking about the iterative and recursive nature of what these agentic tools can do for us. it's really a continuous learning and evolution opportunity for ⁓ our companies and the automation and the, the knowledge basis that, we store around the business right now, I have, an agent running that is ⁓ a for our company based on all of the, the, discussions we've ever had in the thousands of calls over the past couple of years. so I'll report back on how that's going, but, ⁓ super excited about, the ability to leverage all of this unstructured data for, to build structure. for the company and tool of all of the structure without having to sit down and write a bunch of documents that nobody's gonna read and nobody's gonna keep up with. So, okay. call intelligence, that's the continuous, ⁓ autonomous ⁓ ⁓ system that we're starting to build now. The third piece is something that I'm calling Balboa OS, which is, I was really inspired by someone who was on Clairvaux's... podcast. I can't remember the name of it. I think it's how I AI is the name of the podcast. There's a woman on there who has created ⁓ actually no, I'm sorry. This was the cost goop does podcast. I it's called the growth podcast and there a woman on there who built an operating system for her team. She's a product manager. So she's working with developers, other product managers, designers, so on and so forth. But she built an entire operating system for her team. And the way she did that is essentially created a set of markdown files that documented how her team worked. The different people, their roles and responsibilities, their processes, their workflows, different systems that they integrate with. Like for example, Slack, there's a team roster in ROS that has a list of all of our teammates, their titles, their birthday, their anniversary date with Balboa, their Slack ID. their email address, all kinds of core information that we can now use from Claude or from ChatGPT or any other agent that we use in the future to basically go interact with those people on any system that we need to that's connected via MCP. So we're essentially taking everything about the company, what we sell, how we sell it, how we deliver it, the services we provide, the role definitions. way we scope, the we run operations, the way we run finance, the way we run HR, our ⁓ employee our holiday schedule. you can think of is now documented in a central repository, sort of like it as if it was a Wiki. But that repository ⁓ is ⁓ designed to distributed to the team. So every time we make an update, it automatically will download to every teammates, local and they connect to that Balboa OS folder with all of that context available to them every time that they need to sort of interact with company context while they're using quad. And that's really easy to do with quad cowork. You just connect to a folder, connect a folder to a chat, make it the default folder, and then they're using Balboa OS as part of their chat. if they have a question about, or if they have a call transcript, that is a discovery call and they want to use the discovery summary call or the discovery summary skill that we've created, they can just invoke it right there. And chances are, all they got to do is tell it to find the call and a discovery summary and Claude will go out and find the skill on its own and find the transcript on its own. So really, really stuff. What I'm trying to do is standardize a of the things that people have routine questions about. I'm also trying to a lot of the busy work that people do because the most valuable thing in our company, again, we're a consulting firm, the most valuable piece of our company is our people. And the most valuable activity that they could be doing is interacting with other people to help them understand the products that we support, the services we provide, the problem. to help them map out the problems and the solutions that we're working on together. Like that is the most valuable use of our time, not entering data in HubSpot, not tracking time, not literally documents. ⁓ The are being written as we interact with our customers now. And that's a really powerful concept. So I'll report here over the next several as we continue build. on this and expand. some really like niggie gritty tactical things that just have to get done as part of all this. Think security, think about just little things like how do these repositories get distributed? How do people get trained on how to use this stuff? Right? That's real sort of in a smaller company like ours, like we have to think about how do we deploy these things so that everybody can get use out of them? And then, and then make sure that people are using them we put a lot of time and effort into these tools. ⁓ don't want people reinventing the wheel and having to learn how to do things from scratch when a lot of tools now available to them. We actually have a AI show until every other week. Now ⁓ we have coming up on Friday. These are really This is where everybody on the team gets to show off what they've been working on with, with whatever tools they're, they're utilizing. Certainly Claude is our central at this point. But that's part of our I guess you could call it our AI center of excellence. And we see that more and more across the customers that we're working with as well, that they're wanting, needing to have an AI center of excellence to help drive the centralization of key elements the AI system. Also enablement and adoption ⁓ of AI platforms. We're spending a lot of money on them. We should be getting our money's worth and that's one way to do it. Okay. So those are some of the things we're building. I actually had my agent go out and query some of the calls that we've had over the past nine months or three months or so. So from the beginning of the year and had it sort of summarize some of the, um, some of what we're seeing across the industry. And again, we, talk with a lot of SAS companies. We talked with A number of companies that are using technology to ⁓ in their business to sort of support sales and expansion, customer retention, all the things we think about from a customer success and customer experience perspective. And so there were a few themes that popped out ⁓ of So I'll just down the list here. So theme number one, customer success ratios are really broken. right now. Every company we talk to is more and more accounts under fewer and fewer CSMs and the bar is being raised significantly. So over just the past couple of months, I've talked to companies who are no longer assigning a CSM to an account less than a hundred thousand. I've had another customer that they don't assign CSMs until an account is paying at least two hundred fifty thousand dollars. And that makes sense, Think about the, so think about the reach of an individual person in this age, ⁓ Where you have to actually reach out and contact and connect with people. If your customer's not paying you a significant amount, then ⁓ the is consultant. ⁓ It should a very high value consultant. And by the way, Like think about all the time that those consultants are spending on the phone with your customers, all the insights. This is where I want you to really think about the call intelligence, know, transcript ingestion kind of system that I just described to you because, you that's the sharp end of the stick there. But these CSMs, if they're doing their jobs well, they're industry experts, they're product experts, and they be a wealth of knowledge for your business, but that's not cheap. somebody on the phone with an enterprise customer is not inexpensive. That's you see these thresholds rising. In addition to the thresholds rising, you also see CSMs owning from dozens hundreds of accounts each, especially as you go down in the the daily market of things or the commercial side of things. You're seeing more and more. really where customer success managers are overloaded. I could go on a whole diatribe about how CSM is not the be all end all for customer success. Customer success is an operating principle, an operating standard. It's not necessarily a role. Like we provide customer success as company across our support team, our technical account management, certainly CSMs and account managers, onboarding teams, implementation, professional services, so on and so forth. the CSM role is something we talk about a lot. love to see it sort of unbundled into a set of roles that are a little bit more discreet and orchestrated across the customer journey. But at the end of the day, the, the theme still remains. We've got to out how to scale these folks. And so where, that's where ⁓ digital customer experience, digital CS in the ability to. Leverage all of that knowledge and drive it out in one to mini forums, one mini content, knowledge sharing and users in one to mini kind of model engaging on the platform. So, so critical in 2026. And I think we'll be forever at this point. anyway, scale gap CSM are sort of broken right now. Number two. So, Basically. most are sort of stuck between using AI as just an individual sidekick then more of an of an agent. of the things that I've seen more recently, actually just this I had a couple of people, I just made a post about this on LinkedIn actually. Just this week, a couple of people have showed me what their plans are for their digital CX strategy in The interesting thing is they are beginning to leverage the same idea of this autonomous loop that I talked about, the autonomous agent loop customer experience where they are listening signals coming from their data. That could be product telemetry from Pendo. could be a call recordings. It could be email trail, whatever, wherever those, those signals coming from. They're, they're listening to those. that they have an agent that's deciding what to do with those signals. Like, it signaling that a customer is unhealthy and needs some outreach? Is it signaling that a customer is healthy and needs to be signed up for the advocacy program? Whatever the case may be, we actually implement product called PendoPredict that helps make decisions around what to do based on what all of those signals are telling us. Then there's an layer where we go do the thing. Right? Maybe it's scheduling an outreach with a CSM. Maybe it's sending the customer an in-app message to guide them into an area of the product that they need to be paying attention to or a metric that they need to be paying attention to. it's an email campaign, an enablement campaign, so on and so forth. But what is the action that we're going to take? And then there's a learning layer where you actually learn what the results are of all those actions. And this is the interesting part. agent sort sends data back into the system and insights back into the system and evolves it Okay, now I think that's sort of far down the path for those of us who are in B2B. We like to make sure that what's going out to our customers makes sense. The interactions that are being put in front of our customers make sense, but at a minimum, those tools can surface what they think need to be done in our customer journey give the ability to review and either accept or deny those changes to the overall customer journey. So ⁓ I it's about to get very interesting in terms of agentic workflows that are sort of autonomous and continuously improving the way that we interact with our customers. I had something else to say about that and it slipped my mind. So, super interesting. so we'll move on, agentic loops, autonomous, the agentic loops. Keep, keep your eye on that. let's see, what else is interesting here? my goodness. So, we were just at Pindos annual conference in, ⁓ at the end of and almost to a person, we had of people come talk to us at our booth. yeah, we're, we're big, we're big guys now. We had a booth at the conference, which is cool. but had hundreds of people come talk with us and we just listened and learned from them. And, you one the most consistent things that I hear is that our data is a mess. And I know it's true. I've seen the inside of a bunch of these companies, you know, whether a SaaS company or even a bank or, you know, a FinTech business or. You name it, is the limiting factor. Organization of the data is a limiting factor for many of our businesses. And so that was just the common theme. what it's is our ability to do everything that I've talked about up until now. limiting our ability sense what's going on in the customer base and actually tie activity to an account. tie a user to an account and be able to drive outreach from them. We have all these disconnected systems, right? We've got the product itself, which contains a lot of rich information about what our customers are and are not doing at the user level, right? And the promise of SaaS was always that we were going to have into that. And I would say more often than not, we're lacking that visibility. Many of you are probably in your cars right now thinking, yep, we are too. you're alone. It's probably % of the SaaS companies that I work with, don't have the appropriate visibility into what their users are doing in their data. Then you have a secondary problem. You actually need to engage with those users on channels that are different than the product. It's fine if you may have Pindo you may drive in-app engagement through a tool like Pindo, maybe even through Braze, something like that. And that's all fine and well. but that's just one channel, right? You wanna be able to communicate with these folks on email. You wanna be able to maintain sort a consolidated history of what's going on with the account in the CRM system so that you can see it again from your agent tool because you're ⁓ querying with MCP. But the is you've gotta have into what your users are doing in product and how they're engaging with you outside of the product as well. campaigns are they responding to? What events are they coming to? There's a very simple mapping that needs to be done. And I say simple, I don't mean to diminish it because sometimes it's very, very hard. But we need to know who our tenants are in our product and what account IDs and Salesforce or HubSpot that they match to in the CRM, right? And then we need to know what user we have in our system and what contacts those map to in the CRM system. If we just had that mapping, that basic mapping across our systems, forget all the other data, I there's tons more data, support, finance data, invoices, billing, subscription information, all that kind of stuff. Just forget all that for a second. If we just had product telemetry connected up with our CRM data in a consistent way and we had a master customer file, a master user file, and we knew which users belonged to which customers, I guarantee you 90 % of us would be in a vastly different state than we are today. And you know, if you had the ability to query that from, from your, from your agentic system, or even just had the ability to write Looker reports or Tableau dashboards against that. so anyway, data is a challenge. There's a, there's a data layer, opportunity here and And I think the cool part of this is that it's never been easier to write code. It's never been easier to systems that move data around and connect data together. So actually of the things, just shameless plug here, Balboa is providing data and AI foundation solutions now is what we call it data and AI foundations, which is let's get your data into a place where you can actually do something with it. Okay. So it's a that we actually have had to move upstream from the work that we typically do with clients to make sure that that data is connected. think tools like Pendo literally three or four or five times more valuable when you have that connection between the product tenants ⁓ and CRM records. Because then you can use the telemetry data in the go-to-market and in those are from a retention standpoint, from an upsell cross-sell. even from a prospecting perspective, if you're in a PLG style kind of company. needless to say, data is the challenge, NRR is the target, but data is the underlying challenge behind that. All let's see, I'll do one more here, then I'm gonna wrap it up, because I feel like I've been talking forever and nobody needs to hear this much of my voice. We'll talk about onboarding. So onboarding is probably the highest leverage moment that you can control or that you can improve if you're trying to drive retention. I tell this story a lot. Jeff and I had a customer years ago that came to us and said, like we have a really, I was talking to the CEO of this company, a $300 million company. We have a retention problem in this one area of our business. We'd love to take a look at it. And when I asked him was, Tell me about your onboarding program. What does it look like? And he's like, I don't really know. So we time with the chief customer officer who had inherited all of this. This company had been sort of a collection of acquisitions over the years. And he said, the onboarding is actually terrible. There is no onboarding. We basically send them a link to a portal which is disorganized. a large percentage of our customers never even activate less renew after year one. So looked at the data confirmed year one was really dry. Your one renewal was really driving the, was really driving the attrition rate for this particular segment of the business. And as, ⁓ as sort as by CCO, we a look at the onboarding process and it was, it was messy. It was messy. And you know, this was an SMB type solution for folks who were in a blue collar industry, which makes that even harder, right? So we set out to improve the onboarding process and program. And we're working with another company right now who's going through a very similar transformation on their onboarding process. They have essentially digitized it, built some campaigns around it. We're now layering in in-app. guidance for that experience, connecting it with their knowledge base, connecting it with agents that are helping to drive the process. And net result of it, the old that we worked with, we improved their retention rate over 10 points by, in hindsight, by fixing onboarding program. And it's obvious, right? If people know how to get into the system and utilize the data that it provides and they're going to retain, right? It's just, this is basic. but the interesting thing about fixing onboarding is that it takes a long time for some of these results to show up in retention. The problem you always have with fixing retention problems is that it's such a lagging indicator and you don't really know if what you have done is actually moving the needle until maybe year or years from when you start making the improvements. So you've got to look for leading indicators just to give you, have to have some belief that these leading indicators will actually lead to an improved retention rate at the end of the day. So what I mean by leading indicators in this case are, one the things that you may to measure about onboarding is what percentage of your new that are coming into the onboarding process complete in seven days, 30 days, 90 days, whatever is appropriate for your business. could be 45 days. Your business, your go-to-market, your So if you at that by cohort over time, you can start to get a sense of whether it's improving or not. And then you have to have a belief. I always say that a strategy is a belief, So we believe that ⁓ If we can improve the time to value in onboarding and get customers live on the platform faster, then we believe that that is going to impact retention. We're gonna go validate that at the end of the first year cycle of this work. But if you don't have that core belief in connecting a leading indicator to a lagging indicator like retention, then it's really impossible to have a strategy. So anyway, story short, onboarding. still is probably one of the highest leverage moments. think with all these AI tools, you're seeing it now, right? It's more important because many of the AI platforms that some of you may work for or with now, they're sort of billing on ⁓ a model, right? So it's not a subscription necessarily that you're gonna sign up for and the just turns on like it used to. once the sale is made. No, that's not how it works anymore. are consumption-based models which are driving overall ARR volume, right? And that's gonna be more more common ⁓ as these applications sort of ⁓ begin to some of the more traditional software that we've had over the past 20 years. So onboarding is continue to be important. It certainly showed up in our call transcripts as one of the top themes over the past 90 days of where folks are spending time and energy to make improvements. So, okay, that's about it. I've talked to your, I don't know, 30 or 40 minutes. I hope some of this has been helpful. Jeff will be back next week. I hope, cause I don't want to do another one of these by myself again. Um, but we'd love to hear what you're seeing. If you're, um, if you're seeing something different or contrary to what I'm I'm saying here, I'd love to hear about that. So shoot me a note, jay customersuccess.io. We're also going to be at Gainsight's Pulse Conference, May ⁓ 27th, think is, that Wednesday, Tuesday and Wednesday of that week. So we're going to be trying to connect with folks there ⁓ and have some probably have a dinner as well. if you're going to be there, I'd love to hear from you. and love to see you in person in Vegas. Other that, thanks for listening. We really appreciate it. If you do get value out of this, it would be really, really helpful to us if don't mind giving us a review, a good review on or Spotify, wherever get your podcasts. Those are the two biggest ones, of course, ⁓ in our listener base. And so always appreciate you listening. I know we don't take your time lightly. We don't take the fact that we're in your ears very lightly at all. So just appreciate you listening and we'd love your feedback. Okay. hope you all have a great week and we'll come at you again next week. Take care.