Everyone’s Using AI Wrong — Here’s the Real Opportunity

Welcome to the Next Wave podcast. I'm Matt Wolf and if you've been keeping an eye on how AI is changing the way we work, you're going to love today's episode. I got to sit down with Nicholas Holland, the head of AI over at HubSpot, where he spent nearly a decade driving innovation. In this conversation, we explore the massive shift happening right now from AI as a simple productivity booster to AI actually doing work on your behalf. We talk about this new frontier Nicholas calls work as a service, the rise of the super contributor in sales and marketing, and how agents, not just assistants, are quietly taking over repetitive tasks in real businesses. You'll hear a four step framework for getting your team started with AI, why structuring meeting data might be your biggest hidden asset and what it means to become an agent manager in the near future. It's practical, it's forward looking, and it's packed with insights. You're gonna love this one. So let's dive right in with Nicholas Holland.

Using only 20% of your business data is like dating someone who only texts emojis. First of all, that's annoying and second, you're missing a lot of context, but that's how most businesses operate today, using only 20% of their data. Unless you have HubSpot, where all the emails, call logs and chat messages turn into insights to grow your business. Because all that data makes all the difference. Learn more@HubSpot.com.

Well, thanks Nicholas, for joining us on the show today. It's great to have you and excited to dig in and nerd out with you about AI. How are you doing today?

Good. Thanks for having me, Matt. Appreciate it.

Yeah. So let's get into a little bit of backstory. I don't want to go too deep on this, but like, how did you get into AI? What's your relationship with AI and how did that sort of evolve into what you're doing at HubSpot right now?

Yeah, so I've been in HubSpot for almost a decade. I started off running the innovation labs out of Ireland. Then I took over the kind of content and marketing portfolios and eventually became what they call a gm. So just responsible for product and revenue over the marketing products and AI has always been something that, you know, customer platforms like HubSpot have had to do, but it's really been like on the machine learning side. Think of it simply as like scoring things. So lead scoring has always been a big part of that. When OpenAI came out with 3.0, there was rumblings in the marketing world about gen AI. And so we started to look at that. ChatGPT comes out and of course the world changes overnight. Everybody's using it, talking about it. And then we had multiple AI leaders and for whatever reason, they didn't work out. There was a point where they were looking for an AI leader and three years had gone by and marketing is being impacted the most, I think, by AI or has been. Right. And so here we are trying to serve the marketer and we had just put out a ton of value by applying AI on their behalf. And so there was a point, where do we go get an academic leader, someone who, you know, majored in AI? And our product leader came to me and he said, I think we need somebody who actually knows how to build things with AI and we would love to have you do it. And I said, you know, my specialty is in applying AI, so if we do this, we will become experts at applied AI. And he said, let's do it. And so I took over about 10 months ago. And since then, that's really been how we've just begun to drive really deep on how do we apply? We're model agnostic. We take, you know, a multimodal approach and then we apply that. You know, we're at a point now where we don't even think AI is the limiter anymore. It's like, how do you actually bake the cake?

Yeah, yeah, yeah. No, it's interesting because you mentioned marketing, right? My background was actually in digital marketing, but because AI became such a sort of integral part of what marketers do now, I think I started talking about AI as it relates to marketing. And then over time, AI got more and more popular and I went, you know what, I'll talk about AI and marketing, but I'll also just talk about AI because I just love this stuff. But I feel like marketing is probably the area where I feel like AI is really sort of impacted things the most.

Huge for the marketers that are embracing it. I'm seeing like a rebirth of marketing being fun again. Seeing a rebirth of, like, marketers kind of starting to find, you know, blue ocean again. That was like the hardest part of, like, if you roll back the clock, like two years ago, it was a red ocean for all marketers. Everything had been figured out, all techniques were basically known, all the software had started to become homogenized. Now it's like blue ocean again. You're seeing wild variances in creativity, tactics, all that stuff. Pretty, pretty cool.

Yeah, yeah. I mean, all of the really cool marketing that you've seen over the years that you're like, oh, this probably took massive teams and, you know, tens of thousands of dollars to produce has sort of been, you know, for lack of a better word, democratized. And now anybody could kind of do the really cool marketing stuff that you've seen. And you know, I love that. That's what brought me down the rabbit hole in the first place.

I don't know how you see it, Matt, when you're talking with others, but like there's like a new role emerging. I don't know what other companies call it, but we call it like the super contributor. In just the last week I've seen a marketer who runs a prospecting team now. So let that think in. Like the marketer is now running the prospecting team and it's because they're using agents and different AI things to get that done. So now you've got a marketer doing some sales type roles. I've got a sales guy that I was working with, he does all of his own product marketing work now because he's using, you know, image creation, he's doing videos, he's doing all that stuff. And so like, it's just really interesting with all of these like, capabilities coming on. Like, what does it mean to be a marketer? What does it mean to be a salesperson? Like, this super contributor is now rising where it's really like, if you have a lot of creativity and you've put in a little bit of elbow grease to learn some of these things, you got a couple of new superpowers.

Yeah, yeah. I mean, you know, guys like Sam Altman and Dario Amadai from Anthropic, they've made comments about how they think the first one person billion dollar company is coming within the next couple years. Right. That's because one person now could be the builder, the marketer, the salesperson, the, you know, the customer support person. They can do that all by using various AI tools and agents and things that are available to them. So yeah, yeah, we're entering a pretty exciting world right now.

Yeah, it's going to be cool.

So I want to talk to you a little bit about the shift of AI productivity to AI actually sort of doing a lot of the work for you and being able to use AI to build system. Because in the early days AI, like when ChatGPT first came out and even before that when we had access to GPT3 and stuff like that, it was really great at being like a productivity booster in the sense that it could Help you write some emails and you know, beef up your copywriting a little bit and some little stuff like that. But now it seems like more and more people are actually using AI to sort of build out their whole business system. So I'm curious, like, what have you seen in that realm? How are you seeing this shift from productivity to like now AI is sort of able to help run the show.

So about a year ago again, like in the early stint where I was first starting to take over, we started to notice that there's like this shift from SaaS software as a service to what we think is going to end up being was. People chuckle when we say that internally because it can seem a little cheesy, but like work as a service and work as a service can look like, you know, we've even talked about like, should HubSpot actually do marketing for people or do sales for people? It could go that far. But what we think right now is that like that is kind of the key differentiator between technology of the past, like cloud or Internet or mobile, and then AI. Those were all enabling technologies that let you go basically faster. So it's like you heard the old phrase, Henry Ford said, if I built what people asked for, I'd have given him a faster horse. Right. You know, so we actually had that internal debate, are we providing faster horses or are we providing cars? And I think that AI, when you start to shift to it, doing work is a lot more on that car standpoint. Because with cars, what came along, you had new types of professions, you had infrastructure changes with gas stations and roads, you had kind of knock on effects of cities moving further out.

Right.

I think that's the kind of equivalent that we'll see with AI. And specifically like everybody's kind of in love with the concept of agents. Agents are a type of AI that get closer to that concept of work, that it's basically something that's doing a unit of work. I think for us, what we're noticing now is like we group everything as an agent now internally. So even like assistants, like a ChatGPT, they used to go out of their way to say like, we're not an agent, we're an assistant. Right? But as ChatGPT has gotten more powerful, it's doing more work, it's doing more things autonomously. It is an agent, it's just an agent that interacts with humans. So we're kind of seeing agents that interact with humans as assistants. We're seeing agents now that basically you give it a goal and It'll go knock that out. So that's more like a task agent or a work agent. And when you look at that, you know, now you start to break it up into ways that like one, you need to find parallels as a human to think about it. So if you had chief of staff or an executive assistant, they wait for you to tell them what you need. Although some of the best ones anticipate it. But you know, you ask them to do something and they will go do it. You'll say, hey, can you take notes at this meeting? Can you let me know about this next meeting that I have coming up? Will you book this for me? That is a body of work that's starting to happen now and a lot of people are using their AI assistants to do those type of things. So now we have companies all the time ask like, how should I approach AI? One of the first things we say is give every employee an assistant. We have one that's very good for go to market roles. So marketing, sales and service. It's powered by ChatGPT. It's got all the context of HubSpot. But like we say give them an assistant and teach them how to start using that to basically help themselves. The next thing though we move into are the agents and a lot of our customers are still thinking like, what is an agent, how do I go about this, et cetera. And so we have out of the box agents that basically are low lift. You know, it's 15 minutes or less of time to have like a big body of work. And so tier one support, not just of service, but also tier one support of marketing inquiries, Tier one support of sales inquiries. That's like one of the agents we build, right? And then going from that, once you start to do that now people will start to get an idea of how it works and then they'll start to build their own kind of special bespoke agents. So we had a customer who wanted to build something called like an account handoff agent. And the easiest way to think about they just had like this little pain in the rear that was a body of work that they couldn't get sales reps to do very well, right? So I'm the sales rep, Matt, you are the service guy. You're going to deliver on what I just sold. What happens in a lot of orgs is I'm like, hey Matt, I just sold this new account, please hook them up. And you're like, I mean what did you sell them? What did you promise? Like can I get all the details and they're like, yeah, yeah, you know, I'll get those over to you. Or they'll say, can you go check out the CRM? And you're trying to piecemeal all the calls and all that stuff. So again, one of our clients was like, can I build an account handoff agent? That's brilliant.

Yeah, yeah.

It scours through all the engagements, it understands what it's trying to get done. It then puts together an account handoff. And then it's kind of like an ambient agent. Upon deal being closed, it just now does the handoff like. So it's very hard to describe that to like a regular person without going deep because you're like, that's a body of work that you didn't have an account handoff person.

Yeah.

But boy, did you need it. And at the end of the day, that's like why I think now we're not seeing like employment collapse.

Right.

That just makes that business so much better. That makes that service person not hate that salesperson, that makes the system now do this body of work. And so that will then be like a net add for that whole company. So as a long answer, but that's like what we're.

Yeah, I love that. So basically you have the salesperson, they're putting all of their communications into their CRM, you know, HubSpot for all the communications they've had with that customer in the past. Once a sale happens, then the person that actually has to handle the deliverables, they basically essentially get like a quick report on essentially all of the past communication. So I mean, you might even have stuff in there about like their birthday or like their dog's name or something. And then the person that takes it over can actually jump in and already have rapport.

Other thing that's crazy too is like, we all tend to think you're the star of your own movie.

So.

So everything you do is you think about it from your standpoint. We're seeing organizations now where on that one account they had an issue, so they called into support. They then basically were talking to team A, who has product A about this, and then they bought product B. Right now that is all very important for like whenever you hand off stuff, because that continues the cross sell, upsell conversation. You know, what things they've been upset with in the past. You might have multiple stakeholders. So it gets very complex really quickly when you begin to think through all the different possibilities. But what I love about this is like, it's a very, very human, empathetic problem. I'm now in this value chain. I would love somebody to hook me up with knowledge about how we got here. And this person has already done their job. They don't want to waste a bunch of time handing it off to me. And so I love it because it captures the moment that we're in that there's real value here. It doesn't have to be a proxy for an entire job role. The business is happier, the customer is happier. Like that's just a very like empowerment moment.

Yeah, yeah. You didn't cut any humans out of the mix. You just made the humans involved a little bit more efficient at what they're doing.

Totally.

Real quick, if you're enjoying my conversation with Nicholas, you're going to love this. Nicholas is giving us his three step framework for AI adoption. It takes you from complete AI overwhelm to clear competitive advantage in just 90 days. If you're drowning in endless AI tools, Nicholas has cracked the code on analysis paralysis. His AI adoption playbook walks you from personal assistants to team agents to custom development with zero disruption to your workflow. Get it right now for free. Click the link in the description. Now let's get back to the show. So I want to jump back to agents for a second. There's a bit of a debate, right, about like, how do you actually define an agent? I've actually seen people get really fired up about this over on X for some reason about how you actually define an agent and people getting in arguments about that's not an agent, this is an agent. So I'm just curious, like, how do you guys actually define an agent? What's the difference between that agent and the assistant?

Boy, we get fired up internally as well. One of the ways that I've learned to navigate this is, you know, inflexible thinking will be a real detriment to people during this AI phase. I think flexible thinking is going to be a real boon. And so it's not a is agent or is not agent. It's more like a spectrum on this realm of autonomy. Right. I think with an assistant we use that term to just represent the interaction mode. Like its primary job is to interact with a human. When we talk about these kind of task agents, their job is to knock out a task. You know, there's another one we talk about internally as like a customer agent. Their job is to like work with third parties, not in the company external. So you have different kind of rules of engagement with those. All of those are kind of on different realms of autonomy. But in general I think the easiest way that, like, I've been able to put it together in my brain is like, the fact that it can have some autonomy. Agency is really where this comes from. The fact that it can have some agency and begin to do work is when it starts on that spectrum of basically being an agent. The debates we have internally is. And this. I don't want to make anybody's mind blow, but it's like if you go and you ask ChatGPT, Breeze, Anthropic, whatever, if you say, make me a blog post, there will be a lot of people who will die on the Hill saying, that's not an agent, that's just a prompt. I say, okay, great, now rock with me for a minute. So then I'm like, okay, if I ask, only Breeze can do this. But if I say, breeze, make me a blog post, and it now does the same thing the first one did, but it also is context aware of all of your brand guidelines. Is that an agent? There's still some people who will dial in the Hill saying, no, that's not an agent. But then for the first time, I'm like, well, think about it. It's now made on brand content. That's. That's hard enough to even get like one of your staff members to do that, right? So then you get into. If I say, breeze, make me a blog post on brand and generate some images for it. This is why I think ChatGPT has become more flexible in their definition of it. Because now it's done the writing, it's got now some brand application, and it's done an image that's related to it. That's like really three jobs that used to be a copywriter, a media slash image, you know, layout person, and then of course the editor who had put it on brand. So this is what makes it so hard, is because at what point does it become an agent? And so there are a couple of camps. One is if it's doing any work on its own. So that, that would be all the way back to just the prompt. As the prompts get, models get better. Like, that will work. So some will basically go down the path of it's an agent whenever it effectively has a goal, can make tool selection on its own, and has context. That's like our agent leader internally right now, that's his whole thing is like, it has to have multiple tools available to it and they can do that. And so I think right now that's why people get fired up about it, is because the floor is moving up, the models are getting better and better. Like, I don't know when ChatGPT5 comes out, but, like, there's some rumors that, like, it'll now just be a super omni model. Like, it'll just be one. You won't choose a reasoning model, you won't choose a writing model, you won't choose an image model. It'll just be that. So that model will become more agentic in nature and then also over time, you'll give them more and more tools that they can choose from. So, yeah, anyways, that's a hard debate. It's a trick question for anybody that you get in this space. But, like, that's just some of the flexible thinking I have along that spectrum.

Yeah, yeah. I mean, it's very similar to the concept of AGI, Right? Like, are we at AGI already? If we do hit it, will we know? If we do, will the goalpost move?

And also depends on like, like you said, the threshold. Like, there are some people out there that are smarter than others. We probably got some AGI based on those standards we got.

Right, exactly, exactly. I mean, if you go back to like, you know, Alan Turing or someone like that, back from like the 1940s, I think he would probably say, like, you guys have AGI already if you were to ask somebody back then. Right, yeah.

If you were to look at me in high school, how smart I was then, now we had AGI, like, that was. That was definite. AI is definitely smarter than I was back then.

Yeah, yeah, yeah. I mean, when it comes to agents, the definition that I've heard quite frequently is that it needs to work through multiple tasks, but then also it should have the ability to kind of like double check its work. So if it notices that it made a mistake, it should be able to realize it made a mistake and then go back and double check it. And one thing that I found really interesting is that you can actually get models like O3 and O3 Pro, some of Chat GPT's models, to sort of act like agents Right now with a single prompt.

Absolutely.

I was giving it prompts yesterday where I was telling it to, you know, come up with five solutions for this problem and then analyze all five and then pick the best results from the five solutions that you came up with. I mean, that is going through and really thinking through all sorts of avenues and then deciding on the best avenue. You can also tell it things like, you know, solve this problem for me and then once you're done solving it, review everything you just wrote Find any logical fallacies with what you just wrote and then rewrite what you just wrote to fix any of the logical fallacies that you just found.

Absolutely.

That one prompt is acting like an agent for you because it's double checking its work and then fixing the double checking of it. So it's like. Yeah, I mean, like you mentioned, it's just a broad spectrum, I'd say. Like some people would consider what O3 does, where it will work through and pick tools to use to help you get to a solution. Pretty agentic already.

Totally, yeah. And then that's at the model layer, you know. Now what people don't know is like in one chat thread you might ask it a reasoning question, you might ask it a basic knowledge retrieval question, some knowledge already in the LLM, a fast answer, you might ask it to do an image. You know, that's why they call them the O models, because they're omni models. In my opinion, we're already way down the agentic path. So I spend less time haggling over what's an agent nowadays. And now, more to your point, I'm much more fascinated by the work that it can do. And so for our AI now, and we're still early days in this, but like I want the assistant, the breeze assistant. Its mission is to be like your executive assistant, your chief of staff, that's it. And so that's all we obsess over. And so we're doing like really cool things like with that where like it knows who you are, memory and all stuff, it basically knows your role. But it's also working really hard to like understand your calendar. It's working really hard to understand what tasks have been assigned to you. And so that to me is like, when we start to get to work, there's that. And then on the agents, you know, that we just talked about a minute ago, like the account handoff or the sales prospecting agent. Like it's something where there's a body of work that it is particularly good at and that's what we're doing. So we're trying to just really relentlessly find those bodies of work, those use cases, hammer those out with agents. And then I think what's coming next is I think everybody's going to be overloaded with agents. That's my next theory. Yeah, yeah, you know, a hundred agents working for you sounds cool. You're that billion dollar, one person company. But on the flip side, like, man, what does it look like to manage a hundred agents? Who knows?

Yeah, that'll be interesting. I mean, I've always liked the analogy of like a conductor. Right. Like you've got like your symphony and you've got your flute player and you've got your violinist and you've got that, and then you've got the conductor sort of conducting the whole thing. And I feel like we're slowly moving to the role where most people get to be that conductor and they get to conduct all sorts of different instruments and the different instruments know what they're supposed to be doing and they do that thing.

Well, that's right.

I think what really excites me about the future of AI is when AI gets a little bit more proactive. I think there's a lot of people right now that they sort of avoid AI, they don't want to use AI, that you know the ethics behind how it was trained and, you know, all sorts of reasons. Right. But I think once AI starts to get like more proactive, where it's looking at your emails, looking at your calendars, and then like, maybe one day you just get like a message on your phone that's saying, hey, Nicholas is trying to schedule a call with you. I see that you have this date open on your calendar. Would you like me to respond to him and book it in for that time? Right. Like, once agents start doing that, I think the world is on board. I think it's like, okay, I literally have a personal assistant in my pocket now.

We're close. We're closer than people think on that. We're spending a lot of time thinking about like the primitives or like the low level things. So there's this concept that's come up of these things called ambient agents. They are proactive, like you said. The concept there is like they're listening to all signals at all times and then they know what to do. So we have a pretty robust automation platform at HubSpot that kind of is baked into the whole platform. So everything that we do at HubSpot is like one unified code base. Everything we do translates to all of our product lines. So we've got this automation layer that touches everything. And so when we were building agents, we originally started off with like, give it a goal. So just think of like a master prompt. Then people wanted to give it a lot of data and context. So this was like, how do you feed it a bunch of information? Then they wanted tool selection. So can it email, can it web search, can it do that? So we did all that. So that was working really good. But then people were like, how Do I run this thing? And you're like, what do you mean? It's like right here you click run. And they're like, man, I'm not going to go into your agent page and click run every time I need it. Like this is supposed to do a job and it should do it at this moment.

Right?

This was like a really good epiphany for us for like, oh my God, think about it like this. It does a job, but all of us have a signal do it at this moment. And so we basically pulled the automation platform into the agent builder. So that for an example, you know, again I go back to something simple like account research. The first time a form is filled out, you can go do account research. You know, the account handoff. The next time it's marked as sold, it can go do the account handoff. And then there are other events that we listen to. So it's not just on objects changing stats like events. So like you can actually say when person visits this page. And we know with our customers like when they visit. So now whenever Matt visits this page, it can actually kick off a whole stream of like write an email, Matt. So good. Where you been? Haven't talked to you in forever. And like, so those ambient agents are here now again I go back to how good these are. Is going to be less and less about the AI and more and more about like who is the chef cooking the meal right now?

Right, right. So I want to chat a little bit about real world use cases. Like what are some things that like business owners listening. Obviously HubSpot has a very large audience of small and medium sized businesses. So that's primarily who's probably listening to this podcast. What are some like real world use cases that people can do like right now, today, using AI to help their business, whether it be, you know, other AI tools or HubSpot's internal tools. Just like what would be some of your best advice right now?

Well, first I do think leadership needs to make a decision. What you'll find in a lot of companies is that they'll have some forward thinker, innovator on their staff who's dabbling and playing with this. But until leadership makes a decision that they're going to embrace this kind of future of work, it's very hard for a company to move forward. So that's first. I think you should have top down buy in. Once you do that, the very first step that needs is that I think every employee needs an assistant. It can be a Gemini if it comes with Your Google account, it can be chatgpt if you want, or Claude Anthropic's good. And then of course we have Breeze for people who have HubSpot. And Breeze again is just ChatGPT powered plus all the context. But everybody needs that assistant. And then to your point, from that point forward, they should start learning how to interact with those assistants. So that's just like your regular old prompts, like, what's the weather? Tell me, I want to plan my trip to Bermuda. You know, can you help me brainstorm on this marketing idea? I'm about to contact this customer. You know, they have these objections. How would you handle it? It's like your thought partner. Once you do that, the next thing I encourage companies to do is a little strange. But it's like get all of your unstructured data into a central system. So if you're using a Fathom, a firefly, HubSpot has a note taker. Get everything, all the meetings, all that stuff recorded, get it into a system where now the assistant has access to it, the future agents have access to it. That's really critical because there's like 80% of a company's knowledge is locked up in phone calls and meetings.

Yeah.

So now you've got every employee has an assistant. You've got your data now unified with unstructured. And that's the first time that you should now start to choose an agent. I think you should start with an out of the box one first. The one that I think is going to be most ubiquitous. Kind of like websites kind of came up in the like 90s, late 90s, and it was weird when you first had a website and then became weird if you didn't have a website. I think right now it's like in that early phases, it's kind of past weird, but it's like people are starting to have agents that man the frontline support. And it's because they never sleep. You know, they're helpful. They're not trying to rush you off the phone. They're not trying to really like turn you into a sales lead. You know, there's all these kind of perverse incentives whenever you're having a tier one support. But hiring a customer agent, like for HubSpot, all you do is like point it to your website. If you have a knowledge base, you point it to your knowledge base. If you have some files you want to give it, you do that and then that's it. It's literally those three things, like upload a few files, point them to your website, point to your knowledge base, and then we're seeing that it can resolve roughly 55 plus percent of resolutions. That saves a ton of time for your employees to now work on other stuff, escalations, all that.

And it still has like a handoff kind of thing as well.

Yeah, it has a handoff. So if you say, I'm not happy with you, you know, escalate me, it'll do that. The other thing that it can do is like start to find out if you're like, want to talk to a salesperson and you want to route it to sales so you can have a very like informal conversation about products and pricing and packaging, all that stuff without feeling rushed. And then of course, when you're ready, it can do that. So that's the kind of the first agent that I recommend people do. And then after that I recommend that they then begin to bring their people along on what I kind of call like AI fluency. This goes in a bunch of different directions, but like carve out time, like one day a month for them to mess around with this stuff. There's like educational stuff that HubSpot provides, like help them come along and you'll turn around twice and then in a year, your people are fluent. All of them have an assistant, you've got an agent or more working for you. And then the software that you use day in and day out, if it has AI features like HubSpot has like over 100 AI features just baked in, they're not scared to click it. We noticed a bunch of people like AI features are coming into all these software tools and a lot of people are like, I don't know, I don't know what it does. I don't, I don't know how much it cost, you know, like, I'm busy. And so that's what we think like an AI leader looks like. Their data games on point, their employees are on point, their agents are working. I think that's it.

Yeah, I like that a lot. I think one of the things you mentioned that it's been sort of a game changer for me is I've started recording a lot of my meetings and if you actually have the ChatGPT desktop app, they actually have like a record button now too, where, where it'll record your whole meeting and then at the end of it, give you a summary, give you any to do items. So I just actually hit the little record button on my chatgpt and I don't take handwritten notes. I'm just like, Very focused, engaged on the conversation. When the meeting's over, I go, all right, what did I say I was gonna do? What did they tell me I needed to do? What are they gonna do for me?

What are my action items? And those threads, a lot of people, again, this is like, where we're all pretty, like, new to this game. But, like, Those threads in ChatGPT are kind of the same, like we learned in, like, Breeze. It's not just that moment people chat with it right now, but, like, let's say that, like, you and I are in a sales engagement. We're hanging out, we're talking. You did that recording thing. And then let's say I go cold. It happens all the time. Like, we don't talk for two or three weeks. And then I'm like, matt, can we get back together and chat again? You're like, what the hell? Like, what do we talk about? And you can go back and ask questions of that conversation. Yeah, this is like a new muscle that like tools, like granola, if you haven't seen granola, AI or, you know, like, these are new muscles that we've never even had before, which is. Wait a minute, wait a minute, wait a minute. It's not just a transcript. It's not even just like some structured follow up or task outlet. It's like I can go back and be like, what did Matt say about this? How did he feel about that? Like, that's crazy.

Yeah, it's wild. I mean, you know, I think back to, you know, whatever, 15, 20 years ago, right? Everybody knew everybody's phone number, right? Like, I knew all of my neighbors phone numbers, my parents phone numbers, like, all of my best friends phone numbers. I had them all memorized today. I don't even know my mom's phone number. I couldn't tell you what it is. It's all saved in my phone now. I don't need to remember that. I've sort of outsourced that piece of my brain a little bit.

I know.

I'm not saying people should, like, outsource, like, all of the conversations they have, but it's almost like that same thing of, like, now we have the ability to remember, like, every conversation we've ever had, in the same way that we have the ability to basically have storage for every phone number, every contact, every connection we've ever learned. And, I mean, you could do it in the real world. You've got things like the limitless pendant and the bee that just got acquired by Amazon and, you know, tools like that where you could just be walking around in the real world keeping track of all your conversations as well. And so it's really an interesting world to be able to sort of like, outsource our memory. Like, I do have concerns that maybe people will outsource too much of their memory. So I still think it's important to be very engaged in the conversations. But it is very cool to be like, I had a conversation with him two months ago. What did we talk about again? And then, you know, jump on two months later, be like, oh, how was that birthday party you were going to, you know, the day after we talked? Or whatever. It's really cool that we have that sort of superpower now that we didn't have before.

You know, it's funny. The other thing I'll tell you is, like, I think there's a lot of fascinating anthropology type things going on here. So two months later, you asked me how the birthday party is in today's world, if I'm not super into AI, I'm blown away. I'm like, oh, my God, it's amazing. He remembered, like, what a thoughtful, incredible person. And like, over in our lifetime, our younger kids may not be, like, in our lifetime, there were people like that who were just like, so, like, thoughtful. I think now if you said that, I would probably know that you had some tools. You know what's funny? This is where I think it's awesome. I wouldn't care. Yeah. The fact that you asked me about it is meaningful. We talk about it, we build a little rapport. And I go back to, again, the Delta, that only a few extraordinary people would have remembered the birthday. And now it's something that we all have. And so now you and I can have that moment. Let's say that if that's not your superpower to remember that stuff. And so there's just a lot of, like, things where I think that, like, the threshold for humanity or the bar is going to move up in a good way, which I think is pretty exciting.

Yeah, for sure. There's something that you mentioned before we hit record. You mentioned context engineering. And I hate to admit it, but that was the first time I've heard that term. Can you explain what that means? What is context engineering?

Yeah. So the AI journey that a lot of us who are in the AI industry, maybe you too, Matt. It's like you learn about a prompt, then you learn that the human language has a massive capacity for communicating information. So everybody was really learning about prompt engineering. And then there were some prompt styles that came out. There's a great YouTube video on a prompt structure called craft. C stands for context, R stands for role, A stands for the kind of action. F is the format that you want it in, and T is like the tone. And A can be for the action, dress, score, or the audience. Anyways, point being, you used to be like a master at prompting. You would fill all that out. But what has happened is that to really get good AI result, it takes a lot of work to give the right amount of context. So we started to see a body of customers that naturally, on their own, would have, like, the Mega PDF file. The Mega PDF file would be about Matt Wolf, about his podcast, about what he cares about. It's like this Mega PDF file. And then you would bring that into every AI interaction, saying like, hey, I'm looking for an idea, I'm looking for a script, I'm looking for this. And here's all the context. Well, where things started to get wild is that there's a huge variance in the context that you bring to all of these AI things. And so there's a lot of effort right now in this world of context engineering, which is, you know, you're trying to make a podcast on HubSpot. You can make a lot of content on HubSpot, a lot of the content right now. If you don't give it the right context, even with great prompting, it's not you. It doesn't sound like you. It doesn't know your history. It doesn't understand all these things that you've done in terms of previous guests. And so we started to see that, like, context is a layer above data. So everybody used to think data was this moat. Data is getting easier and easier to get, very hard. Especially, like with customers wanting to own their data. It's very hard to treat data as a moat. But context is a way that you basically pull signals out of that data. And the way we first stumbled across the first kind of context thing was companies using HubSpot kept saying, like, I don't want to have to keep telling it about the products I sell. I don't want to have to keep telling it about who our competitors are. I don't want to have to keep saying something like, who's our ideal customer profile? These are all reasonable things that a business would want a new employee to know. This is what we do. This is who we do it for. This is why we're unique. This is who we sell to. And so we were like, dang. So people were now having to constantly Copy and paste or upload that mega PDF file with that. And we were like, why don't we start to create all that context in a layer that they can do that? And so we first came across like a brand context. Then people were like, I wanted to know a lot about me. I'm Nicholas, I'm a SVP in product. I'm over AI. These people report to me, These are my peers. This is like what I do, memories about me. So then we built out like a user context model. You know, now I want to sell to Matt. Matt's a prospect. And so, you know, you go look at his LinkedIn page, you go search his blog, you know, you go find him on Reddit. And so you then build out like a whole context profile on Matt. That's a buyer context. And on and on and on and on. And so we started to find out all these contexts that really were better than people carrying these mega PDFs, and you now bring those to the table. So what's happening in a lot of the AI leadership stuff is that anybody who's working on AI solutions, including I think OpenAI and Anthropica, I think this next year is going to be all about context and memory. I think people are going to obsess over that because it will ultimately make all of the agency, all of the agents and things like that just so much better.

Yeah, I totally agree with that. I mean, I think the thing that excites me the most about these new models that keep coming out, it's less about, you know, it got a 99.3% on this benchmark instead of a 99.1%. Isn't that amazing? Like, those kinds of, like benchmark leaps, like, they're, they're impressive, don't get me wrong. Like, cool that you did that, but from like a real world consumer that's using it all the time. I love seeing the increasing token count. Like, I love seeing that Gemini now has 1 million tokens that I can use and supposedly they'll be making 2 million tokens available soon. Like, that's what really, really excites me. Because then you get these opportunities to really bake in even more and more and more context. And, you know, before we hit record, one of the things you mentioned is like, what is defensible? Well, like, like you mentioned a lot of the data isn't defensible. Almost anybody can acquire data at this point. But like, the context that you bake into the models that you're using is, I guess, what is A lot more defensible these days. I know for me, like, I have the YouTube channel and I do a lot of like, ad integrations. I've actually over many, many, many, many months dialed in the exact way I want my integrations to sound and how they want to be worded and the exact sort of like, structure of them and the beats that I want to hit on them and how do I get them under 90 seconds. And I've sort of trained over months exactly how those integrations should go. Now it's gotten to a point where a sponsor will give me an ad brief and say, here's what we want you to talk about. I literally copy and paste that brief into my sort of pre built. I use ChatGPT projects for this. So I throw it into that project and it just like spits out the perfect integration pretty much 99% of the time where I barely have to do any tweaking and nobody else has access to that specific context that I built up over time.

Listen, if somebody were to ask you to change now, like, you'll be like, you'll pry it from my cold, dead hand. And so I think that pick that maybe you do a particular outro, maybe over time you have some sort of quick hits that are a summary of this. Like, you'll build up three, four or five of these type of, you know, they're like sub assistants or kind of, you know, personal assistants, you know, custom assistance. You'll build those. They will have the context that you're looking for. And it really does then become a virtuous loop. The more you use it, the more context it gets, the more you dial it in. And then it's very, very hard to then switch. That's why it's great for the vendor and it's very, very hard for you to switch. And so, you know, I was laughing if you were back, you know, starting a product, if you ever started a company. Like, one of the things you hear is like, something you build has to be 10 times better to get people to switch. I don't know if you've ever heard.

That, but like, yeah, I've heard that. Yeah.

The thing is, is like in the world of AI right now, a lot of people are switching without just like whatever's the latest, the newest, you know, those swap, they're like very, very easy come, easy go. But I am seeing as people get more and more of that context engineering down that that bar is now moving up to five times, 10 times. And there is a world where it knows you so well that it might be harder than 10 times better to get you to switch.

No, absolutely. I think another thing that's really interesting about the world we're going into now is that even software is kind of gotten to this point now where anybody can create, like little tools that they need for their business. I think we've entered this place now where if I have a little like bottleneck in my business that I just find myself. This is something I do every day. It's repetitive. Why do I keep doing this? I can literally go to one of these tools, have it generate a little script for me and just take that off my plate. And I even feel like software and like writing little apps is no longer really that defensible because I can go and find an app that I really like. I've done this on past videos where I went and made a Spotify clone just to see if I could. Right. I went and made a feedly clone just to see if I could of some of these tools that I use in my own life. And now I don't even feel like having a software is even that defensible. So just really interesting.

Well, I will say this. I have a controversial take on that.

Let's hear it.

I do think you're right. For some people it will be hard to be defensible. However, this is like an interesting take the other way. I think you have the same access as I do as others to go by paintbrushes, a canvas. But what you paint, what I paint and what a master painter paints is wildly different. And so I think that the part that I'm also somewhat attuned to is that there's no accounting for taste in this whole thing. So I think there's another thing that's kind of fascinating happening in the world of AI tied to what you said, which is you have a little problem. There's no solution out there. This is when I talk about the bar for humanity gets better. Like you solved it. It's awesome. No harm, no foul. There's another one where there's a tool out there. There's not a lot of competition. Maybe the price is too high for the value you get. This is going to be one where you will want to build something and that's going to be a fascinating world for those type of things. There is another world where it's just so easy, you don't even look for an alternative. You build it. And I'm seeing this where people now love the tools they build. Like there's a psychological Thing called the IKEA effect, which is you built this table from ikea. You love this table, Matt. This is your table born of your own two hands. Your friend comes over and he sees a table with a crooked leg and a chip inside of it, but you love this table. And so I think there's going to be a lot of IKEA effect happening right now where a lot of people are building these apps that they're very proud of that do the job. By the way, that is key. It's like not a project, it does the job, but it's also nowhere near commercial software. It's not high taste, it's not great. And so there'll be a whole world of these like personal projects that have been enabled. But I still think there's no accounting for taste. There will be room for some of these high quality, high craft type tools. So again, the competition field out there is going to change quite a bit.

Yeah, no, I totally see that. I mean most of the little apps that I've built are ugly as hell, but they solve a problem for me and I would never try to sell access to them. They just help me with one thing, you know.

Well, I know we're running out of time here, but I do want to ask about like HubSpot and like, what are some of the like really cool AI features of HubSpot that you're excited about right now and are there any sort of upcoming AI features that maybe you can tease a little bit? Anything exciting to talk about there?

I think I'm excited. We re architected the Breeze Assistant to go a lot deeper with OpenAI and to bring a lot more of the HubSpot context. So for a salesperson, a marketer, a service person, I'm just pretty excited because I think, I don't know how you feel, Matt, but like there's a huge capacity for using LLMs out there. Like a lot of people I know they first start with one, but then over time they're using Perplexity for search, they're using, you know, Google for search, they're using ChatGPT for analysis, Claude for writing Grok for current events. Like there's like this ability to handle multiple LLMs in your life. And so I think there is going to be a world where there's going to be work LLMs. Now some of the ones I just mentioned will try to go into work as well. I'm excited for that. Like we'll have Breeze Assistant and I think it'll be great for the go to market professional A lot of the agents we put out were kind of dog water when we first put them out. Meaning like they, they just were like very, very immature. The models weren't there yet when we first put them out. The, the orchestration around it. But a couple of the ones that we put out are pretty darn good now. So the customer assistant, that tier one one is pretty good now and it is having meaningful impact to businesses. And then the prospecting agent, we just came out with the V2, it's still in private beta. But what's really good about it is that it does account research and it's really good using deep research. So it does account research. It then goes through and it's. This is crazy. It's actually deal aware. So let me tell you what I mean by that. You put a contact in there that you've worked with. It's watching the interactions and it is ready with the next communication that either you can send or it can send on your behalf. And so you send it out and let's say the person replies, let's say they get on the phone and you call them. Let's say that you have a zoom meeting in the middle of all this and then you then hang up. It's ready for the next communication that's aware of everything that just happened. So it's not like a automated sequence. It's not that it's like, it's crazy. So it's like it's all, you know, I call you up, we hang out, we talk about three or four different things. It's now ready for the next steps in those plans. So I think that's pretty good. And then the last part is, I think this next year a lot of people are going to want to curate their context. So we're going to expose a lot of that context for customers to look at. And I think there will be a rise of a new worker. Actually, I think there will be an agent manager and an agent trainer. Sometimes they'll be the same smaller companies, but in a larger company you can see an agent manager being somebody like the sales leader who says, we're going to use these agents for our sales org. And then the trainer might be either the BDR themselves, the SDR themselves, or maybe a team manager. So I think that's going to be really wild because you'll start to see people go, yeah, I've managed agents in my last job. And you're like, oh my God, you're hired. So like, I think that's awesome.

Well, I mean this has been an amazing conversation. I'm really excited to see what HubSpot rolls out in the AI Realm and the agent realm coming up. And, you know, thank you so much for hanging out and sort of nerding out about this stuff with me today. It's always a blast.

Yeah, thanks for having me, man. Appreciate.

Sam.