The Marketing AI Show—the podcast that helps businesses grow smarter by making artificial intelligence approachable and actionable—is BACK with another episode.
You can listen now on your favorite podcast app, or keep reading for more on what to expect in this episode.
Episode 8: Kevin Walsh, HubSpot, on HubSpot's AI-Powered Features
In this week's episode, show host Paul Roetzer sits down with Kevin Walsh, Group Product Manager for AI at HubSpot. Kevin’s teams are responsible for building HubSpot's machine learning as a service platform and for integrating machine learning models across HubSpot's product lines, including SEO, content recommendations and others.
During the conversation, Walsh gives us an inside look at how HubSpot thinks about artificial intelligence, and what customers can be doing to take advantage of AI-powered features within the CRM platform. Show host Paul Roetzer also dives into HubSpot’s point of view on AI, AI applications in the platform that customers can use today to make their marketing smarter, and a little bit into the company’s product roadmap for AI solutions.
Watch the Video
Read the Full Interview Transcription
Disclaimer: This transcription was written by AI, thanks to Descript.
Paul Roetzer: Welcome to The Marketing AI Show. I am joined today by Kevin Walsh, group product manager for AI with HubSpot. Welcome Kevin.
[00:00:11] Kevin Walsh: Thank you very much for having me.
[00:00:13] Paul Roetzer: It's been a little while, man. I think we did a webinar probably about four or five months ago. So it'd be interesting to see if anything's, uh, new and exciting at HubSpot. And that's what we're gonna focus on in this episode. We're going to talk about HubSpot's point of view on AI, AI applications in the platform today. That customers can be using to make their marketing smarter, more efficient drive revenue, and hopefully get a little bit into the company's product roadmap for AI solutions, as much as Kevin is allowed to tell us.
[00:00:41] Um, but first let's talk about Kevin and, and someone who studied economics and music at Northeastern university to becoming a leader in HubSpot's AI practice, where you've been in the last seven years. So how did you go from music and economics to leading projects, uh, on AI for HubSpot?
[00:00:59] Kevin Walsh: yeah, that's a good question. It wasn't, uh, the plan all along, but it's, uh, through a lot of hard work, good timing and dumb luck. I've been able to. To land it where I am. Um, I studied music primarily in college and then added the econ actually, as I sort of like improved graduation and was like, well, shit, I probably am gonna need a job.
[00:01:21] So I might want to supplement some of the arts with something a little, uh, closer to STEM is kind of how that unfolded. When I graduated college, I really wanted to work in music tech, but I couldn't get an interview, uh, anywhere. Um, because I knew the music side, but I had never worked in software. And so technology companies are usually smaller.
[00:01:41] They're not growing super quickly in general. So it was just really hard to get an interview. Um, but I went to Northeastern in Boston. It was. 2012, 13 ish. And HubSpot was a really hot company, a really hot name, like crazy. And I joined the support team. They're hoping to just like learn as much as [00:02:00] possible about technology and the internet.
[00:02:02] And along the way, I've learned a tremendous amount about marketing sales and services and how people run their businesses. And, uh, HubSpot has just been a rocket ship of a ride. Um, I went from the support team to work on the HubSpot Salesforce integration was my first product team. Uh, before that I was like the support specialist in-house working on kind of the trickiest customer support problems, and also like working with the engineering team to help guide their roadmap and then just took over as the PM for that team.
[00:02:34] So my early roots in product management were always on like, Pretty heavy technical infrastructure projects. Um, and honestly it took me like a long time before I was working with like a proper designer on like an app that like required like a lot of like, really thoughtful sort of like user flows, um, and design elements.
[00:02:53] So I've always kind of been biased towards more of the data and infrastructure projects, um, [00:03:00] and working at an RDC or started working at HubSpot. Uh, one of the benefits that we have is our tuition reimbursement program. So when I was on the support team and as an early product manager, I was taking, uh, data science and machine learning classes through Northeastern's computer science graduate program.
[00:03:17] I did that for about two semesters and started working on, I ended up taking over predictive lead scoring, which was HubSpot's first machine learning product. Um, still our oldest product, uh, And what was interesting is when I took over break to lead scoring, it was already out and we kind of started over from scratch.
[00:03:39] And what we ended up doing was we built. HubSpot engineering at the time was going through this really, um, opinionated platform phase where everything that we were doing was about trying to extract core services that are shared across all of these apps down to single teams. And so, um, today we talk about those as the primary colors, where there's one team that's responsible [00:04:00] for the reporting platform, and there's another team that's responsible for most of the CRM platform and another team that's responsible for the messaging platform and another team that's responsible for content.
[00:04:09] It might have one more, but the idea is then you have these primary colors that are really important that you can string together to create these really like complete apps. And that's how all of the product development at HubSpot works today is how we're actually able to move really quickly. Now it took us a long time to sort of build that foundation.
[00:04:25] But now we have like this really powerful toolkit to string together, combined and launch products pretty quickly. And we'll talk about, um, how we were able to do that with. A new product conversation intelligence in a bit. Um, but I bring that up because when we went to build the second version of predictive lead scoring, we were also building the machine learning platform behind the scenes in that machine learning platform is what powers all of the production machine learning that we have.
[00:04:52] In the product today. And, uh, I'm still the day-to-day product manager for the infrastructure piece. There's [00:05:00] a few folks on the team now who are using that infrastructure to do applications, um, that are a little bit closer to customers. Uh, so that's kind of a long-winded winding way of saying, yeah, it was, I was kind of.
[00:05:15] Interested in more of the infrastructure data projects. I started taking some classes just because it seemed like an interesting way to upskill. And then, like I said, just, uh, some dumb luck and good timing. I was able to sort of like be in the right place at the right time to catch that first wave. And since then, we've just still been growing and we've, we've done a ton of projects.
[00:05:36] Uh, By nature of machine learning is pretty experimental. So a lot of stuff that we've worked or a lot of stuff that we've done, uh, it hasn't worked, but we've learned from like why projects haven't worked and we throw stuff out all the time and we, uh, we feel pretty good about our process now for building.
[00:05:52] Uh, successful machine learning products, and you can expect to see us be a little bit more bullish about sort of how we're [00:06:00] positioning and talking about, um, the machine learning power in the app. Uh, kind of as a result of we've learned a lot, and we're really confident about our ability to execute at this point.
[00:06:09] Paul Roetzer: Which is interesting to me. So people don't know my background. I know people probably listen to this podcast are aware that I own and run the Marketing AI Institute, but I also own PR 20/20, which was HubSpot's first agency partner back in 2007. So I've been working with HubSpot personally for 14 years or so.
[00:06:28] And so I've seen it grow up from. Uh, my, one of my early sales calls, whether or not, I don't know if you've ever talked about this. Kevin Halligan was actually on one of the sales calls. That's how early we were in HubSpot and Dan tire was my rep and they, they were trying to sell me on ranking for PR firm in Cleveland, like HubSpot used to sell keyword rankings was the value of the platform that was when the CMS sucked.
[00:06:54] Like the, you know, there. Product was not than what it is today, [00:07:00] but they've always been very transparent in that. Like they, they're always very realistic about where the product was. And so for people who aren't familiar with HubSpot now, again, it's a $22 billion market cap company. Most people in the marketing space are familiar, but.
[00:07:15] What is HubSpot? Cause I even know the it's positioning has been evolving a little bit in recent months about how you're going to market. So just, you know, what are the different hubs? You talked about the colors, but the customer facing, what are the different hubs that make up the platform and how do you guys kind of think about the platform from a public standpoint?
[00:07:35] Kevin Walsh: Sure. So the kind of elevator pitch on that is it's HubSpot's mission to help businesses grow better. And what that means is growth through sustainable growth, which is we want to help small and medium businesses, basically folks that are really in kind of like the 10 person to like 500 ish person size companies, um, be able to.
[00:07:56] Make as much money from their customers without providing a [00:08:00] poor customer experience. So that's sort of like sustainable growth. Um, what does that mean from the product side? Well, we have a very powerful free CRM that sits at the base of all of these, uh, product lines. And on top of that, we have hubs that are really focused around careers, basically.
[00:08:18] So we have a marketing hub or we're working to identify problems and challenges that face them. My marketing manager and director and VP personas primarily. Um, and we also have that for the sales hub, which is focused on the sales profession, service hub, which has focused on customer service profession.
[00:08:36] Uh, last year we launched the CMS hub, which was sort of splitting out the website product from the marketing product, uh, to have a full fledged, um, CMS. And you can expect to see over time, we're going to continue that launch new hubs, loosely, that follow that trajectory of like, who's the person that's buying this what's the career path, because we find that that's like a really helpful way to [00:09:00] ground, uh, the product.
[00:09:02] Strategy, you know, you always want to know when you're developing a product. It's very important that you have a very clear understanding of who your customer is. And you also want to be able to provide a product that's going to grow with that customer. So by grounding that around sort of like those career paths of marketing sales service, and then all of the other roles within a company is likely how you can expect to see HubSpot, continue to launch new hubs.
[00:09:28] And then meanwhile, like one layer down at that sort of CRM platform level, we're still adding sophistication at that level that can then be used and applied to each of those sort of key personas that are onsite.
[00:09:40] Paul Roetzer: So then how do you think about AI within the organization? So you don't have an AI team sitting in marketing hub and an AI team sitting on sales hub.
[00:09:47] You're if I'm not mistaken, your HR team is more like operational across all hubs and all areas of the organization. Is that accurate?
[00:09:55] Kevin Walsh: That's how we work today. Although, uh, we're kind of getting into the inside [00:10:00] baseball. My teams are structured. It's likely that we're going to try to, uh, Aligned sort of like within my, the artificial intelligence group, there will be sort of like a marketing persona team and a sales persona, judge help to help do that to date until really this year.
[00:10:17] And last year, that was the strategy was that we had a centralized team that was trying to, they are in small production machine running across the hubs. And that's why a lot of it was this sort of like. You know, under the covers sort of lift and the, the strategy was we wanted to just be doing machine learning in places that we thought it would help all of the customers, um, There is some sort of like organizational reality that some of the projects we choose to work on, or because we think that those projects can kind of be accomplished.
[00:10:47] And so while there might be really good fits for something like in the certain side, for example, we know that there's really strong, uh, customer support automation opportunities for machine learning within the service persona. Uh, we [00:11:00] don't have any of that in the product today, just because the priorities of the service hub aren't there yet.
[00:11:05] Uh, but meanwhile, we've done like a ton of. Customer service automation projects internally for HubSpot Spelman support team. So we know that there are a good fit problems there, and it's just a matter of prioritization and
[00:11:16] getting there.
[00:11:17] Paul Roetzer : And so just a quick step back, how do you define AI and machine learning?
[00:11:23] I mean, we have a lot, our audience is like 50% would identify as beginners and another 37% would say they're intermediate in terms of their understanding of AI terminology and capabilities. So how do you define it and kind of, how does HubSpot think about those two terms internally?
[00:11:38] Kevin Walsh: Sure. So AI is a broader term where you're delivering a product or a system that feels smart.
[00:11:46] It seems like something that would require a person to do. And there's not a technical definition for what that means in practice, what we find in industry and that HubSpot that. What passes as AI today [00:12:00] usually has some machine learning model, somewhere in the system. Um, machine learning is a specific field within statistics.
[00:12:08] And the way that I like to explain it is you're trying to mimic human decision-making with some really specific tasks. Um, so in that regard and machine learning can really be thought of as just more intelligent automation, where in the fast you might be able to say, like, you know, if the life cycle stage of a lead is equal to MQL, then rotate them to, um, sales.
[00:12:31] So that kind of automation. Uh, is great and really powerful and businesses should be adopting it, but for more sophisticated automation where it requires something like reading an email or reading a support ticket, um, that requires machine learning to help sort of automate that human tasks of like, okay, I have to read this email and identify if I should move the deal stage for something, for example.
[00:12:53] Um, so as we think about good fit opportunities to apply machine learning today across. [00:13:00] HubSpot's products. Uh, some of the lowest hanging fruit is in the automation bucket where all businesses should have pretty clear defined processes that they should be evaluating each sort of like sub steps within that process as an opportunity for automation, whether that's automation with rules or if that requires machine learning.
[00:13:18] Um, I think that that is the most approachable way to think of it that, you know, if you think about. Uh, tasks, like sending an email to send an email there's like 12 steps, sub tasks or steps that you need to do. And can we help reprove any one of those 12 sub steps, same thing with like, you know, booking a meeting.
[00:13:36] If you're a BDR trying to, uh, line up a lead for an account executive, or if that's completing a support. Case for the customer service persona. Each of those like real world business tasks can be broken down into smaller component parts and we try to automate those.
[00:13:51] Paul Roetzer: And then I, you know, that the other thing we talk about a lot with machine learning is that it's not, you set the rules and then every quarter you check in and [00:14:00] maybe adapt to rules a little bit, the whole idea of machine learning is that the machine learns from the data and can in theory, improve its recommendations or decision-making.
[00:14:09] And so I think that's a key aspect to, to think about.
[00:14:14] Kevin Walsh: Yeah it definitely is. I, I do find that, uh, you do have to babysit these projects or these processes once they're out there, they don't really self heal quite as much as folks think. Um, normally what you're doing, like if you're trying to, uh, Do some tasks like predictive and account is going to churn or not.
[00:14:36] And you find that, uh, you know, you have a model out there that's trying to predict churn and you find that it's missing a whole bunch of predictions. That's an opportunity for a machine learning model or to sort of dig into why the model was wrong in normally they find that there's some real world factor that the model is missing.
[00:14:51] And so the development cycle for that is like, okay, we noticed that we actually didn't have like contract size. As something that the process was [00:15:00] using to actually predict churn or not. And we know that people that have higher contracts are more likely or less likely to churn for some real world reason.
[00:15:07] And so the, the improvements still happens. Um, but it's just a little more abstract about that, where you still need all of that data coming in and collecting, and that becomes useful, but you have to kind of keep an eye on, on where things are failing and try to do some
[00:15:22] exploratory analysis. Yeah, I think it's a big misconception.
[00:15:26] Paul Roetzer: And I know you and I have talked about this before that like some people think AI is just this magic magic switch to like, Oh, it sounds amazing. I'll go find a tool to do email for me that has machine learning in it. Now I'm good. I can go on and do other stuff. It's like, no, the human in the loop is a very real thing and the constant need for inputs and training data.
[00:15:45] Um, for this stuff to work. And I, I remember, I think you, you had told the story one time about early lead scoring with HubSpot and the use of it internally and how there's almost this expectation of marketers and salespeople that okay. You gave me a [00:16:00] machine learning powered model. It's going to work from day one and I'm all of a sudden going to be working better leads than I was before.
[00:16:08] And then when that isn't always the case and the human realizes, Oh, I actually have to help the machine learn. It's like, Oh, forget it. I'll just go back to what I was doing. And that's a very real challenge for adoption of AI. Think. At a broader level in marketing sales and service, is this education to humans that no, you're still part of this.
[00:16:27] Like it is a human plus machine equation. Like what have you done internally? Because I know you've, you know, you've been in there from the ground up from like 2016 on when you guys probably had. A couple of machine learning engineers. Like it was not a big thing within HubSpot back in 1516. And I know you've had to do some internal education for your own uses because oftentimes HubSpot build something for internal use that then it's like, okay, this works internally.
[00:16:54] We can take this to the market too. How have you educated? Not only the development team, like the [00:17:00] engineers, but. The users at HubSpot of what AI actually enables and, and how you actually adopt it over time.
[00:17:08] Kevin Walsh: Yeah, that's a, that's a good question. It was a long saga, uh, of, um, trying lots of different ways to message this and communicate sort of like what the best fit opportunities really were.
[00:17:22] Um, The best analogy that I was able to come up with was what we were calling. The 10 interns problem was the idea that if you could define a narrow task that you could teach, you know, a rookie and intern on, on one day, how to do that task. And you could then teach another nine people how to do that same task.
[00:17:41] And if they would all reach consensus about it, then it's probably a good fit for machine learning. So for example, um, at HubSpot and one of the internal systems that we have that's related to the partner program actually is we have a web crawler. That tries to identify if a website belongs to a business that is likely to [00:18:00] sell marketing and sales services.
[00:18:01] Okay. And we actually use that to help our HubSpot sales team is split into two. We have people that sell to our direct customers, and then we have people that sell to our partners because there's a different skill set there of like reselling HubSpot and understanding all of the different services that our partners actually offer.
[00:18:18] Um, and so to help our own sales team get organized, that's why we have this web crawler that when someone fills in a hubspot.com form, we try to identify behind the scenes of that person should probably work with a partner specialist, or if they're likely just looking to use HubSpot for themselves. Um, so that's the basic, uh, Product that we're trying to build is something that classifies websites into one group or another.
[00:18:40] And in working with the folks that, uh, lead the Parker team, it was helpful to use that 10 intern analogy to be like, imagine that you're going to show somebody who just walked in off the street, a website, and you have to describe to them how to sort the websites into one group or the other. Um, that was a really useful tool because originally we were [00:19:00] hoping that we would be able to like, Scan websites and identify like sales partners as one type and CMS web development partners as another type and marketing specialists.
[00:19:09] As another type, when in reality, it's like very murky about exactly how to draw those lines. And so it's unrealistic to think that you're going to be able to clearly say somebody only sells sales services or they only do customer support sort of operations. Um, and that's how we were able to then. Take a group of websites and like sit with the team and say like, okay, let's, let's sort these ourselves.
[00:19:33] And if we find that we can't actually make this decision ourselves. Then it's not something that we're going to be able to automate. Well, um, it is something that we can sort ourselves and that's a great opportunity for machine automation,
[00:19:44] Paul Roetzer: because couldn't you go couldn't, you use some NLP function to go to their site, scrape the data, and then look at what they would call solutions, or basically build a taxonomy of what the definition of services would be.
[00:19:57] And then actually have the machine [00:20:00] extract, the services they're offering, like is that.
[00:20:04] Kevin Walsh: You have to start with, with unknown set of answers. That's like the training in machine learning. So the way that that works in practice is rather than having to go down to like the paragraph level and set rules that say like, Oh, if you see website design services, then automatically classified like this whole webpage as a CMS partner.
[00:20:23] The, the idea is that you're basically supposed to be able to say, like, show me the whole website. I don't organize it at all, but I'm going to tell you that this entire website is associated with the. Sales partner lead. Gotcha. And the magic of machine learning is then it's able to do all of that fuzzy matching.
[00:20:38] So you don't have to define those like really granular down to the word kind of rules. Um, and the truth is like, but most of the like really accurate machine learning models that we use for image recognition, self-driving cars, voice detection actually is powered by this massive army of. Human data labelers.
[00:20:57] And there's a whole industry of PR companies [00:21:00] sell data labeling services and help companies like HubSpot to connect with crowd labor across the internet, to not data so that we can use it for automated processes. Um, so that, that, like, again, gets back to the idea that I think the best way opportunities for most businesses right now are in that process automation.
[00:21:19] And I find that that's a helpful way for people to think about it too, because. Defining the process is still a really hard problem. And that requires deep business knowledge for, you know, All of the talk around like robots are coming to take our jobs from marketing sales and service and operations professionals no time soon, because those are really hard jobs.
[00:21:40] And we sort of need marketers and sales reps to think of themselves as systems architects for like these really important business processes. And then they can, once they understand that they should know that machine learning can help them sort of automate away pieces of that business process. So go ahead and think about it.
[00:21:57] Paul Roetzer: So at, at, uh, A big picture level. So I, you know, Dharmesh is a friend of mine and I've met Dharmesh and I've had many conversations about AI going back a decade. But I, you know, from a public perspective of things that Dharmesh has said, he gave a talk at Inbound. I think it was 2018 about AI. Like it was, it was, it was very focused on AI elements.
[00:22:18] And he, I think at the time said that, you know, for, for him and for HubSpot. The, the play with air, the point of view on air is that it's a seamless element of the platform and I, you know, and that it really should just be. It should just be behind the scenes, powering it, and that we don't need to necessarily be out in front talking extensively about AI and machine learning.
[00:22:42] And so, you know, having talked with the general managers of the different hubs or heard them speak at public events, I've always felt that HubSpot has been very understated in, in its work with AI, that it hasn't like necessarily, you know, you can look at earnings calls with Halligan. It doesn't come up.
[00:22:57] Like we don't. Talk about AI. We don't talk [00:23:00] about machine learning publicly. And so just like, what is HubSpot's public stance on AI? Like how, how does, how is it thought about, and how is it talked about, um, as you know, more and more marketers in theory start demanding smarter technology. Do you see a shift where maybe HubSpot starts talking more about it or has a more public point of view that it regularly talks about?
[00:23:25] Kevin Walsh: I think so. I mean, I think behind all of that, the key point is that HubSpot is first and foremost. Motivated we're on a mission to help businesses grow better. Yeah. We're not on a mission to help artificial intelligence arbitrarily. Right now we are also a modern technology company. So we're going to use modern technologies, including machine learning, to help achieve our mission of helping businesses grow better.
[00:23:48] But that should never take precedence over that core mission, um, puts me personally and like a little bit of a fun position of like trying to like marry those two things together. But, um, [00:24:00] I firmly believe that. And that's something that I personally have been motivating at HubSpot that we need to the software and the value of the products needs to stand on the merits of the value that they deliver to marketing and sales reps first.
[00:24:13] And if anybody's curious and they want to know how it works, then like sure. When we're happy to dig into the weeds on like the machine learning processes behind it, um, And we find that that's actually really resonating. It's very on-brand for HubSpot, for as long as HubSpot's been around, we've sold methodologies and software to your point of like the early days of HubSpot where, you know, the leadership was pretty candid about sort of the state of the software being.
[00:24:37] People were still willing to work with us because they believed in the methodology that we were selling him. Okay. I understand inbound marketing and these tools will help me implement it, but there's going to be some gaps. And I think people, our customers and our brand resonate with that. And so the core message for HubSpot AI is that HubSpot AI is practical.
[00:24:57] It is valuable and it is approachable. It is something [00:25:00] that you can understand. It is not. This like all knowing, being not as smarter than you, that it's going to take over your business or your processes from you. It's a tool that you should have a conceptual understanding of. We think that we have some places where it's going to really help your business, but mostly you as a marketing manager should be about generating traffic as a sales team.
[00:25:22] It should be about converting that traffic into customers as a service team, which would be about delighting those customers. And that should always be true. I think we're going to see, you'll see HubSpot become more forward. I think about, um, where. Artificial intelligence is within the hubs. Um, and we're going to talk, I think a bit about conversation intelligence is just launched on Monday, but that's kind of the first example of us getting really forward about some of the like AI messaging.
[00:25:52] Um, but I think you're going to also see other companies in this space probably pair back a bit. Uh, and I think that the [00:26:00] state of AI applied AI in business software today is a bit like where cloud was kind of like the latest. Nineties where like the fact that cloud was a selling point has totally now it's just table stakes where like it's expected that there's like cloud software.
[00:26:13] I think AI technologies will be the same thing. It will be impossible to deliver some of the value that you promise without using modern techniques. Like. Yeah, I think
[00:26:23] Paul Roetzer: That's a great point. The, the whole, the grow better. It mine is like, well, you can't, without it, like, you know, it's going to in the very near future, whether it's three years, five years, like if you're building products in this space and it's not powered in some capacity by machine learning, chances are, it's an obsolete.
[00:26:42] Solution like it's just going to be embedded in everything. So there's three main areas that at least from the website perspective, that HubSpot has really invested in AI data, cleanliness, content optimization and conversation intelligence being the new area. So let's just talk for a couple of minutes about those [00:27:00] areas, where if, again, if I'm a HubSpot customer today and I'm listening to this, it's like, Oh, I didn't, I didn't know.
[00:27:04] They had AI built into their platform. Where can I use it? How can I start today? You want to talk about a couple of the practical examples where HubSpot has built solutions that customers can use?
[00:27:16] Kevin Walsh: Sure. Yeah. Our largest sort of body of work right now is automatically related to data hygiene and data quality.
[00:27:23] Like you mentioned, um, behind the scenes, we call that project Ned for never enter data.
[00:27:29] I hadn't heard that before. There's a few things that are kind of spawning out of that theme, but. Um, some of those products include, you know, one of the first things we did there was the business card scanner, which take a picture of a business card.
[00:27:42] Uh, we use on-device optical character recognition to convert the image to strings, to text. Uh, and then we have a HubSpot built machine learning model that tries to match the right. Texts chunks from that business card picture to the right CRM properties as [00:28:00] part of our mobile app. So that's part of the free mobile app that anybody can use.
[00:28:03] Um, whenever somebody I'm renovating a house right now. And so whenever I'm like meeting with trades people and they give me their business card, it's like great to try to stump the business card scanner. It's pretty good. If you have like an unusual business card format, like maybe one of the vertical ones that are just less common, it might get tripped up, but.
[00:28:20] Check that out. You know, if you're listening to this right now and you have the HubSpot mobile app and a business card handy, scan it, see how it does, um, related to data entry problems like that. Um, another feature that we have, uh, is what we call inbox automation, but that if you have a connected inbox and we try to scan from the body of the email, uh, important contact information, like name, job, title, phone number, I think address is also one of the properties that we're looking for there.
[00:28:46] Um, the point being that we. That's one of my favorite, uh, products, frankly, because I think it encapsulates like a lot of the philosophy of why we pick the products that we do. And. [00:29:00] There's so much important information that gets exchanged in the email between, you know, your team and your customers, but there's so much work that is required to make sure that that information gets captured and organized in your CRM so that it can then be used to do more effective things.
[00:29:16] For example, if you have somebody's job title, you can get a sense of the seniority. Of the folks that are related to deals in your sales pipeline, or if you have job title, you can segment your context, your context list into seniority to work on new, not targeted promotions or something like that. Um, The other thing I loved about that.
[00:29:34] It's like, there are companies that will sell you contact information, but we think that that's a murky at best. And so rather than try to provide people, you know, personal information, like names and phone numbers, it's in the email. And so what we're using machine learning for is like a really white hat way that just like help.
[00:29:53] Your team keep track of the information that your customers are willing to giving you yourselves. It's a lot, um, more [00:30:00] HubSpot, as we say. So the business card scanner, the email parser, there's also the HubSpot insights database, which we're, you know, we're supplying information about companies. Um, and last year I think that was last year into two years ago.
[00:30:13] Uh, we launched CRM deduplication for our contacts and companies. Um, that was also kind of a capstone project for us. The real power of the CRM duplication tool is the fact that it works on. Database CRMs that are millions of records. And I don't think there's anything else on the market that is able to do duplicate beyond, you know, exact matches, uh, very large databases.
[00:30:37] That's a really hard problem in computer science to, to compare one object to every other object. And as you, as you grow past, you know, a couple of hundred thousand contact records that can begin to take literally decades of time to do that computation. So we had to. Do some, some really clever stuff to make that work.
[00:30:54] Um, and the message really, really resonates. I think there was a Gartner survey recently I can pick [00:31:00] it up, but marketing managers, self report that data quality and data hygiene issues are the biggest problem that they face. And I think a lot of the early forays for applied AI in the MarTech space, more broadly, kind of around like 2013, 14, 15 were a lot of this like predictive type stuff where the concept was, we have all this data in the CRM, so we can use that data to do more interesting things like smarter routing or personalization or scoring, et cetera.
[00:31:30] But I think over time, many. Product teams. HubSpot included discovered that the data in the CRM is really noisy. It's really messy. And that's been a problem for it forever. You hear that from Salesforce, they don't like entering data and operations teams. Managers don't have visibility to help correct the data.
[00:31:45] So they implement all of these rules to try to get the data to be better. And then the sales team hates doing that. And so solving that problem, that problem of. Can we help make sure that the data and the CRM is clean and accurate is actually going to enable some of those, like more [00:32:00] obvious use cases, like better personalization or segmentation or scoring.
[00:32:04] Um, we've been finding like a lot of success with that. And so
[00:32:06] Paul Roetzer: That's the path we're going. Can I, can I put it a, a product request? I just thought of, as you were talking, I want the LinkedIn sales navigator integration. When I have told HubSpot there is a match. I want to be alerted when that person changes jobs.
[00:32:22] Because like, if you think about your CRM database, what like 20%, maybe on average probably turns over every year by leaving their job. And if they gave you a company domain, which is what every B2B marketer wants, like we want the company domains, not the Gmail, my origami, like give me the Gmail. Cause I you're, you're going to stay in our database forever.
[00:32:41] If you gave me your Gmail. But if you gave me your company domain, it's going to within two years on average. Be irrelevant. And so I would love to have a prompt that says, this person is now at this company. Like, that'd be cool on your roadmap.
[00:32:56] Kevin Walsh: But point noted something in that direction that we don't have [00:33:00] plans for us.
[00:33:00] They don't hold me to this. But related to similar, like parsing email content that people respond to you, it's not uncommon to get an auto response. That's like, Hey, I'm out of office or, Hey, I'm actually switched companies. So for that kind of stuff, we'd like to get better at helping people organize.
[00:33:16] When people are telling you I'm switching jobs or I'm switching roles or whatever. So here's my new address and like automatically ingest and just that, but again, Oh, no plans on that. Uh, no real plans on that yet, but that would be a good fit for us to work towards someday.
[00:33:30] Paul Roetzer: Um, well, let's take a few more minutes here to talk about conversation intelligence.
[00:33:35] Cause I know this is the big thing right now, the hot launch. So tell us what, what this is, is a category for people who maybe aren't familiar with it. And then what HubSpot's play is in this space.
[00:33:46] Kevin Walsh: Yeah. So the category for conversation intelligence and which is not conversational intelligence, the spaces are a little, uh, Too closely named there, but the point is it's supposed to be intelligence about your [00:34:00] conversations.
[00:34:00] So wherever those are happening, and it's mostly aimed at sales tools or sales teams right now, we're providing a set of tools to first and foremost, help with coaching and analysis. So in practice, what conversation intelligence really means is a collection of features that are like advanced call analytics.
[00:34:21] So at launch today for the HubSpot product, which launched on March 15th, which was just a couple short days ago, um, that allows you to. Uh, record calls on HubSpot's native calling. We have a zoom integration that's in beta too. So like a conversation like this could be basically sinking from zoom and into HubSpot.
[00:34:42] Yeah. Then takes that call, audio transcribed. The audio. First thing you get from that is the ability to search. So you can go into the HubSpot search bar and the NAB at the top and search for key terms related to maybe pricing or some of your competitors names and help identify where, where some of those things are.
[00:34:59] Um, [00:35:00] In the product today, we have a full, beautiful, full screen coaching UI that allows you to share particular, you know, points in the conversation. So of a key phrase was mentioned, and you want to share that to ask for help from your manager, or if you're the manager and you want to point out, you know, really good or less than good behavior from your team, you can share that you can make comments, the UI, um, and all of that is related to the kind of like the coaching experience.
[00:35:26] Um, As it will help marketers. Uh, we are also in beta right now. Soon to be rolled out to everybody is the idea of keyword reporting so that you can tag individual sales conversations with particular keywords or topics that may have been mentioned, uh, because of the power of the HubSpot platform coming soon, you can expect that you'll be able to segment contacts based on who's had certain kinds of conversations.
[00:35:51] You can base automation based on what was said in conversations. Um, And the conversation intelligence mission was really, again, it's born out of that, never [00:36:00] entered data theme. Where what we're doing is helping people capture the information that's being exchanged between your team and your customers and prospects organize that data in the CRM through reporting through transcription, through search, and then allow you to actually action that data through routing segmentation, automation, et cetera.
[00:36:18] So we're really excited to see that product roll out. It's in. Sales of enterprise and service of enterprise. Uh, today there are some, we kind of wherever possible. We try to provide some of the functionality of the flagship up and down the product tiers. Yeah. What are some things where you can view the calls in like the free product, but if you want to actually record and transcribe information, that's in the enterprise tier.
[00:36:45] Paul Roetzer: I have not checked it out yet, but I'm anxious to do the demo. Um, you kind of alluded to this, but just kind of end on a, on this segment of the podcast with the product roadmap and, and again, you, you alluded to it through process automation, never enter [00:37:00] data. I mean, is, do you think those categorically are kind of what guides the product roadmap within HubSpot right now in terms of development of AI features?
[00:37:10] Kevin Walsh: Yeah right now. I mean, we also always are usually doing some like really experimental stuff where, you know, there's like a lot of hype right now in the AI space around like large language models, like two, three off keeps coming up. And, um, the idea of like writing, we've tried to do things like write email, subject lines or write tweets based on blog posts, or like start to write like the whole blog posts.
[00:37:34] It's all pretty experimental still. Like, it's really still like on the edge of sort of like. Applied value in research, but we're always chipping away on it. Um, meanwhile, there's so much just like more. Value to be driven on some of the less sexy applications that are more rooted in like the data hygiene and process automation.
[00:37:53] So you're likely to continue to see us add little touches of that throughout the product. And you're [00:38:00] also likely to see us do, you know, one or two more sort of like flagship AI, power launches, like conversation intelligence for the next several years.
[00:38:08] Paul Roetzer: Nice. All right.
[00:38:11]. And we're going to end with rapid fire. So here's a unique one for Kevin. I always try to come up with like one. I don't ask other people. Um, given your background in music, favorite musical instrument that you maybe previously played in, like, what are you playing today?
[00:38:26] Kevin Walsh: Uh, right now my favorite instrument is the, uh, most subsequent 25 synthesizer, which has been a quarantine toy for me. It's a. Uh, duo phonic synthesizer. That is sort of the younger brother of the classic, subsequent 37.
[00:38:45] Paul Roetzer: Um, I have no idea what you're saying right now, but I have things to be Googling. No, I love it.
[00:38:50] But what does it do? I mean, what are you using?
[00:38:53] Kevin Walsh: It says it's a synthesizer that's used in, uh, No more like electronic [00:39:00] music. It's like more of a bass sounding things. I grew up playing. I was in a garage band when I was a kid. So I went through like a guitar phase and then I went through a piano phase and my don't kind of going through up as like a technologist.
[00:39:10] What I like to do is sort of like really dig into how things work and, um, understanding. Audio synthesis is, is sort of the culmination of a lot of those things. Once you have the fundamentals of how like music and basic instruments work, it's like a totally different dimension to sort of like, uh, flex creatively. So that's been, that's been a fun sort of poppy.
[00:39:29] Paul Roetzer: So you're still staying in the music though. I mean, it is still a passion of yours and you mix it in, you know, in the free time.
[00:39:34] Kevin Walsh: Yeah. Like pursue it or put anything out. But I have some, uh, very close friends that are on labels and still touring and things, not towing these days, but yeah.
[00:39:43] Yeah. If you having gone through Northeastern's music program, yet I have made some lifelong friends that are still very involved in it. That's awesome.
[00:39:50] Paul Roetzer: All right. Um, I've asked you this before in a spotlight that you did for the Institute. So I'm gonna, uh, you might not remember your answer. I'm gonna see if your answer changed.
[00:39:57] What percentage of marketing tasks [00:40:00] will be intelligently automated to some degree, meaning there's going to be some AI in it. That's helping with process automation in the next five years, pick a percentage or a range.
[00:40:11] Kevin Walsh: Um, most of it, 75% plus yeah, 75% plus I think it, it depends on what you count as like a task is like electing a lead.
[00:40:22] A task is writing a blog, post a task is coming up with the idea task. Um, but I think we're gonna find more and more ways to help people like do the job to be done, but. Machine learning I'm pretty skeptical will ever be able to do truly creative work to really out of like nothing work. Um, I think it can do derivative work and you see that a lot in like the applied sort of like machine learning arts where it's like, go and look at like 500 paintings and then try to paint.
[00:40:52] Um, But I think for some of the most like original work, like that's, that's where people will always be needed and we'll just have a tool set that will [00:41:00] help people be better at that.
[00:41:01] Paul Roetzer: All right. Uh, which of the following marketing categories will experience the greatest disruption from intelligent automation the next five years.
[00:41:09] So pick, pick the one that you think is going to be the biggest. So I'll give you a few advertising communications, content marketing. I was just going to say conversational, but now I don't know if I should say conversation, but we'll stay with conversational customer service or email marketing or other, is there one that jumps out to you as like, yeah, this.
[00:41:29] 99% of this space is going to be intelligently automated to some degree.
[00:41:33] Kevin Walsh: Well, more of the, there's kind of like two answers to that. I think that like customer service stuff is really ripe for, uh, automation, because it's so process driven right now. That's why it's such a good fit because the steps within the process are so clear and they're so measurable.
[00:41:47] And that's where like, you can imagine like, That 10 intern, like context being applied really clearly there. I think the communication stuff is most right for like net new disruption, because as these big language models get [00:42:00] better at generating texts, like right now it's still the models talk, but they sound kind of drunk when they talk.
[00:42:08] Paul Roetzer : And so especially the longer they go, it's like after a couple more beers, it like, it really degrades when it's
[00:42:13] Kevin Walsh: spitting out. Yeah. So we had the same thing. We had an experimental project to try to like do one of those. Like a sorta like in Gmail, when you type to tries to finish your sentences, we've tried to do that in like the HubSpot CRM communicator, and then like support tickets where you're trying to fine tune the generative models to like a particular context.
[00:42:31] And we like made it talk. And, but then it was like, kind of just like spitting out jibberish and then it took us awhile to like, make it not sound crazy.
[00:42:39] Paul Roetzer: It's a hard problem, but it is, as you alluded to like, Language generation is a bit of a Holy grail right now. I mean, there is a massive arms race in terms of funding.
[00:42:50] And I know a lot of the companies that are working on this stuff right now, and it's, you're going to see leaps forward in language generation in the next like 24 months, I think.
[00:43:00] [00:42:59] Kevin Walsh: Yeah, I think you're right. I think it's like there was this moment sort of. Like maybe 10 years ago or a little less than that.
[00:43:06] When like, it seemed like computers could all of a sudden see, like every picture was like tagged on automatically. And like, they were trying to get you to be like, Hey, is this like your friend? Or is this you? And you're like, yes, that is me. What you're actually doing is like providing
[00:43:19] Paul Roetzer : Apple. Does that, if you go in, it'll say like, are they, is this you?
[00:43:23] Is this your daughter? Like it'll show and yes, you're training. You don't know that's what you're doing, but you are training. It's the same thing.
[00:43:29] Kevin Walsh: It's the exact same. User flow that we have and like the D duplicate app, or like you go in there and we, like, we think these are duplicates. And when you say yes, then we're like, okay, cool.
[00:43:37] We got that. Right. And like, we'll get better over time. Um, so just like, there was this kind of watershed moment where all of a sudden computers could see everything and everything was tagged. I think we're approaching a similar watershed moment where like, computers are going to be able to talk. They're getting better at like responding to us.
[00:43:52] Like if you use Google home or Alexa or Siri, uh, I always kind of find it fun to quiz those services. And I do find that they like are getting [00:44:00] better. Um, so I think we're approaching a similar sort of a breaking point with the generative stuff worked. The language that these things are able to generate is actually legit and usable.
[00:44:11] Paul Roetzer: That'd be great. I agree. All right. So then last one, just nice transition voice assistant. You use the most. Alexa, Google assistant Siri Cortana. Don't use them. There she is.
[00:44:26] Kevin Walsh: I have Alexa in my house, uh, on sonar system that's handy. And like, most of the lights in here are now. Uh, she's like really going, she's going.
[00:44:37] Um, so yeah, so he's an Alexa, uh, I have an iPhone too. So I do use Siri from time to time, but honestly, I'm mostly using that too. So like set timers and reminders and things. Um, I don't use a Google home. I think my TV has an option for that, but, uh, I don't have much experience with the other ones.
[00:44:55] Paul Roetzer: All right, man.
[00:44:55] Well, this has been awesome. I really appreciate the time as always. Um, we'll do it again soon, hopefully. And thanks again for being a part of it.
[00:45:03] Kevin Walsh: Yeah, thank you. And, I'm looking forward to your feedback about conversation intelligence.
[00:45:07] Paul Roetzer: When you get a chance to check it out, check it out, man. All right. This has been The Marketing AI Show. Thanks for joining us until next time. Appreciate you being a part of it. Talk soon.
Sandie Young was formerly the Director of Marketing at Ready North. She started at the agency during the summer of 2012, with experience in magazine journalism and a passion for content marketing. Sandie is a graduate of Ohio University, with a Bachelor of Science from the E.W. Scripps School of Journalism.