Eoghan McCabe on Intercom's AI agent Fin: resolution-based pricing is the future, growth rate doubling three years running

to join the show. The founder of Intercom. Welcome to the stream. Oh yo, gentlemen. How are you? Look at that microphone. Look at the lag time. You were just You were ready for this. Terrible. No, it's good. It's fantastic. It's fantastic. I I don't want to mispronounce your name. Can you pronounce it for me?

Well, before I tell you, there are two types of people in this world I found out. There are those who guess, and then there are the rare few [ __ ] gentlemen. But you because you'll never guess. Um, it's pronounced so kind of like O H. Wait, you you right as you said it. What is it? It's pronounced own basically. O.

So as if you as if you spelled it O H N. Own. Ow. Awesome. A lot of people say Owen and I will accept that. Okay. Owen. Well, thank you for joining us. Um, great to have you.

would you mind uh just giving us a brief interview of kind of brief intro to your kind of entrepreneurial journey and the company that you built and kind of then we can dive into a bunch of hot topics.

I'm sure we'll be talking about AI venture all the top all the typical topics but I'd love to know kind of like your story. Yeah. Um in a nutshell um I moved from Dublin Ireland in 2011 a long long time ago uh to the Bay Area in San Francisco. Uh and we started intercom around then.

Intercom at that time was a generalpurpose customer communication platform. We had a cute little messenger, the first of its type. You put on your website in your app and it was great for sales support marketing use cases and it blew up.

Um, two or three years into intercom, we were the fastest growing software company since Salesforce. Slack subsequently beat us, but that was our that was our hot start. And um yeah, it's been a an incredible part of the whole like snippet trend, right?

Like you were able to boil down installation of your product to just like a few lines of HTML or JavaScript. And so the integration all of a sudden went from get a deployed engineer build on top of our API to drop this snippet into your website and you're good to go with some basic functionality. Is that correct? Yeah.

Yeah. There was there was other people doing it. There was a company called Olark years ago that had like a kind of a live chat thing. We were a little different. We were a messenger, so it was a mix between live chat and and kind of uh you know asynchronous.

Um but what was different about our approach was that you'd install the JavaScript and it would start to learn about your customers and so you had a really datari customer experience. So in many ways it was almost like a CRM for digital businesses, but at core to it was all the communication stuff.

How did the business model evolve? Were you on kind of just like a monthly plan for a little bit? Did you go based at some point? We we tried approximately 14 trillion price plans. Um most of them were unsuccessful. Yeah, but we we you know we charged per se and per message.

Um the problem with a really horizontal product like ours was was that you're trying to capture value in a dozen directions. So you have a super complex pricing model if you're really trying to do all the things. So we we did that and then eventually we picked a lane. So these days we're in customer service. Okay.

So So I'm you've tried all these business models. I have to ask you about Salesforce. Uh they're moving to this like uh ticket completion pricing, getting away from seat pricing. Uh is that a reasonable strategy? Is that the future? Like Mark Beni off kind of said, uh are you testing that?

Do you have any information on what we should expect from future pricing of these types of services in an AIdriven world? Yeah, I think it'll stick. We were the first to charge per resolution. Salesforce came to me. Oh, there we go. We're going to source. What do you guys think about this? Does it work?

Um, and you know, we like it. We like it because it's aligned with our customers. Um, we don't charge you. If we don't do work for you and every time we do work, we charge you. And when we make our technology better at doing work, we earn more money.

And right now, uh, forgive the plug, Finn, our AI agent, is the highest performing, uh, agent in service. So we more we earn more more per customer at least when you charge for resolutions than the other guys because we just do more work. Oh, interesting.

Wait, so it's highest performing not on a particular benchmark but on just how much money it makes. No. Well, yes, we because that I've been saying like that's actually the real measure of AI is how much how much economic value can it create.

And it's weird to say because we want to be like, oh, it's really good at math or whatever, but I'd much prefer like no, it actually does a job. It solves a real problem. And there's no better measure than that than economics. Yeah, absolutely.

Uh, you know, work in some senses is a bit of a commodity to a degree, right? And so, you know, if you can charge more for your work than others or if you earn more from doing work, it means that you're providing more value to the world. Yeah, totally.

So, yeah, we we have more customers than anyone else in the AI service agent space, we have more revenue than anyone else. That's very high average resolution rate, etc. Have you been in general applying uh applying AI to customer service makes a lot of sense? there's a lot of spend there.

It's just a generally good application of LLMs. Um, were you surprised at how many companies came in to try to compete uh in the category given that the intercom of AI imagine is intercom, right?

Because I feel like a lot of companies came out and just said like, oh, we're making AI agents for CX or we're making, you know, more more native experiences. But in this case, it feels like, you know, there's always been this debate, right? Is AI a disruptive in innovation or is it an extending innovation?

And um I'm sure as the models have gotten better and better, you've just gotten more excited about their potential at Intercom. So maybe talk about the kind of market dynamic. Yeah, it's been interesting. So in hindsight, I'm not at all surprised because it's a giant category.

I mean, the space we're eventually going for is all customer experience, sales, marketing, success, service. There's certainly trillions of dollars spent on salaries globally doing all of that work. So, of course, many other people are going to chase it.

I I I do think that categories are going to be more, you know, they're going to be shared in a way that they weren't in the past. There's no way one company gets half of that trillions. It's just not the case. There's going to be many different players in the space. Yeah. Um, at the time I was a little surprised.

we were first out of the gate and we had so many different advantages. Um, but I was only surprised because I had yet to realize that that times are changing. The the technology world is just so much more competitive now.

Like when we started in 2011, you pick your lane, you you probably would have a direct competitor if you were doing something interesting, but you wouldn't have 16. Like that just was not the case. And I don't have a hypothesis for why that is.

I mean, if I had to guess, I guess I would say that AI is most likely the biggest innovation the world has seen since, you know, the internet, maybe way bigger than the internet. When there are new technology changes, opportunities open left and right. And um, uh, opportunists come out of the woodwork.

I mean, opportunist in the best way, wonderful entrepreneurs. Yeah. Come out of the woodwork to go and and and and pounce on those on that moment. So, I don't know. It's fascinating. It's incredibly competitive and not just our space. Every single category in AI. I mean, think of like note takingaking.

Think of, you know, just make anything up. AI Gary Tan's fault. Let's just say it. I mean, I love the guy, but, you know, we got to we got to hold him accountable for all the competition. Wait, who is that? I was saying I was saying I love Gary Tan, but it's clearly his fault that there's so many startups now.

I mean, half half the intercom um competitors are YC companies. Yeah. I mean, you you know that like si something like three or four of the top payroll companies all went through YC and for a while there was this there was this narrative of like oh YC doesn't fund competitive companies.

It's like no they fund a ton of competitive companies. Uh but it's great. Yeah. I mean competition breeds innovation and breeds you know these huge power law outcomes. Like it's all it's all good stuff. Um but it is very it is very very competitive right now.

Um I want to hear about your evolution with artificial intelligence. take me through. Were you tinkering with the GPT3 Da Vinci API playground? Were you trying to implement that stuff? Was were there times when it was unsuccessful and then it became successful?

Is reinforcement learning and reasoning the critical turning point for you guys? Like how have the different eras of AI over the past few years played out into your product strategy? I will answer that question. I first do have to mention that I think you just heard a burp live on air. Did you hear that?

No, you didn't hear that. Okay, good. I just a lot of Laqua. It was just like a burpy moment. I heard it in my ear. I'm like, "Oh my god, we're all getting carbonated. " Yeah. Good thing we're going to cook this out, right? So, yeah. Good thing it's not live. Yeah, good thing it's not live.

Thank goodness we didn't randomly choose to live stream the show. Um so to your question I you know so there hasn't been for people in our space uh major innovation at the model layer or uh within the model providers that has uh dramatically changed how it is that that we function.

So much of the innovation is in the orchestration around the models and a lot of the ways in which we've got these massive performance gains over all of our competitors is that we've built a lot of our own AI to allow us to do more sophisticated rag and use multiple models in different types of ways.

We've needed to be very experimental. We've run something like 200 different AB tests to ek out different improvements along the way. Interesting. Going forward, the next innovations will likely come by way of these application layer companies starting to work on their own models.

I I you know, at least right now, and this space is so dynamic, so I may have a different opinion in just weeks.

At least right now, it looks like the biggest leaps forward for the verticalized AI application companies will come by way of building verticalized offerings where they have applications and models kind of working together and somehow leveraging um that interconnection.

Yeah, we need an application layer sound effect on the soundboard.

Um, I yeah, I I I wonder I imagine that there was an era where you were effectively building AI agents or AI chat bots back during the chatbot boom that were defined much more like linear business logic if then statements and you layer those up and uh that that might be derided today but it was value creative back then.

Uh I'm wondering if there is a world where um having done that work results in a better product today because you can slot LLMs in to the business logic or train a model to interact with some of the business logic because I imagine co customers and companies do want to enforce some rules and restrictions around the interactions.

And so if you already have those those those bumper lanes on the on the metaphorical uh bowling alley, you're you're going to be able to deliver a better product. But can you talk to me about the evolution of that that rules-based chatbot development to modern oh I just have an LLM that oneshots it? Yeah.

Well, you basically said all the smart things I could say, but I'll just repeat it. There was kind of three generations. We're on the third generation now.

The first generation was all the conditional stuff and we had pretty sophisticated conditional stuff where you could build big beautiful graphical diagrams that showed exactly how the customer would flow through your machine. Yeah, that was great.

Generation two uh used machine learning and it used various heruristics uh that learned from all of the human data points we have about conversations that uh happen to try and uh route people to the right answers in your knowledge base and it would pull out little snippets.

So, we had a thing we ended up calling a resolution bot um because I can share um exclusively here that Zenesk threatened to sue us because we had a um an offering called Answerbot. And little did we know apparently they also had an answer. Oh no. Um yikes. Generation 3 now is LLM Power.

And so that uh is one that uh uses you know AI as we know it today and it does away with everything in generation one and two except as you say there are some companies like us that can leverage those innovations.

So we sell to banks we sell to businesses with highly restricted uh operations in regulatory environments that don't allow them to make mistakes. You know you get it. Our customers very much want um confidence that the LLM is not going to just hiccup and bump its way in the wrong direction.

You know, hallucinate even though, you know, we've managed to squash out pretty much all the classic example would be, you know, somebody gets the LLM to refund them a huge amount for some purchase that they don't prompt injection. We heard about that at an airline.

Yeah, I think I think a big I think a big question I have is at what point is your AI actually better than an agent that would use the platform, right? Because right now over the last few years as AI got kind of good but maybe not great.

I would still find myself in that sort of cycle of being like, you know, kind of getting frustrated and just saying, "Talk to a human. Talk to a human.

" But you could eventually get to the point where somebody would actually just prefer to talk to an AI because it's instantaneous and you can really potentially move a lot faster than an agent who's working across you know multiple uh open tickets.

So I'm curious like what what is the point where it kind of flips where the average user is um you know is preferring to talk to artificial intelligence? Yeah, the world is coming uh to exactly that point for the average user, but it's there for a lot of people already.

You know, we have multiple funny instances where we see people asking for the bot. We see um requests increase when people realize that they can ask quick questions and get quick answers. Um we see um resolution rates and kind of response times uh improve dramatically for certain categories of questions.

The dirty little truth in the era of AI is that wellapplied AI is far better than humans. It is always available, you know, always stays on script, is never rude, you know, often gets the answer far more right than the humans.

We've run a couple of different studies where we've classified the answers that human agents provided using internal, you know, LLM um um products that we've built. We found that only 65% of the tickets that humans marked as resolved are actually resolved.

So the humans are way worse than they think, way worse than they report. And just anecdotally, uh you will have far kind of messier experiences that humans than you will with a very good AI agent. The problem is that still most AI agents are not very good.

Most of them are in fact cheap G cheap GPT rappers and so there is a lot of kind of crappiness. the very big consumer products that already provide [ __ ] support don't really mind but in the future won't be acceptable but we are nearing that point. Yeah.

How how has it been a challenge figuring out the right way to message interpro intercom's products given that many of your actual core DAUs on the customer side that the actual reps are people that over time like they have to be using the software and realizing somewhat slowly or maybe quickly that they aren't as good as the underlying system.

And that's kind of a difficult very human uh question crisis. Yeah. Look, the reality is there's a couple realities. One reality is that those people on the ground and even the support leaders are often not those buying these technologies.

You know, there's a vested interest in those two groups of people not adopting them. And so a lot of people adopting these technologies or CTOs even CEOs that realize their opport their their their operations and customer experiences could be dramatically improved by bringing these agents to uh to the four.

I will say that often these agents are almost sugar on top. They resolve at least right now some of the easier issues and the humans are still needed for the harder ones. They're serving onmet demand. So sometimes people will deploy Finn our agent to their free customers that they never gave support to before.

But often when they deploy it, like I mentioned earlier, people will ask more questions and so it increases demand too. So you're not really seeing, we haven't seen big mass layoffs where the humans are just obsolete. Yep. That said, all of that said, it will compete with human work.

And AI and robotics is simply going to compete with the most repetitive, most demeaning, least enjoyable, uh, least purposeful work out there. And if you talk to customer support reps, very few of them want to be customer support reps long term.

It's a great way to get into the tech world, earn a tech salary, learn what you want to do, but pretty much all tech people have their mind and eye on something else. Yeah. What uh what are your highle thoughts on the AI sales agent market?

You know, we've seen hundreds of companies at this point and the face says it all. The face the face says it all.

But but yeah, I mean I imag you you said earlier in the call that you guys will eventually get into sales and marketing and so I'm sure you have a ton of opinions on it and great customer service is often a sales role.

You're closing the customer as they ask questions and the best sales the best customer service reps I've ever worked with have always been able to turn a customer interaction into a sales interaction. Yeah.

And it's just a much different dynamic sort of taking inbound concerns or feedback with a product, trying to solve it with AI and then passing it to a human quickly versus as opposed to spamming spamming somebody with just nonsense.

Uh my my favorite one was a sales agent that was like reaching out to somebody in New York and it's like, "Hey, have you like you should try to go to the opera in Central Park. " Yeah. The Big Apple that never sleeps. I'm I'm ready to close anyway. Yeah. What's your take on the AI BDR race? I don't know.

It's it's you know you know the way they say marketers ruin all channels you know all all channels get you know exploited and trashed and spammed and completely destroyed by marketers because the job of marketing is to find you know the frontier like whites space where it's not flooded by other marketers.

So all sales and marketing is really exploiting to death all opportunities until people are thoroughly [ __ ] sick and sick to the teeth of them. So so that That's going to happen with AI, you know. Um, and it's already happening. We, you know, I've gotten tricked a couple times with some emails. I open them.

Some of them are on point and you're like, "Fuck you. This is God. I don't even know if I'm allowed to say this f word. I've said it three times. We're working on getting a bleeper, but you're we are a family friendly show. We don't swear, but we we won't hold it against you. " Okay.

Thank you for not holding against me. I will try my very best to not use it one more time. Um, but I I I I don't I don't know if it's going to work, but like you said, there's a full spectrum there. There there's a massive spectrum and Yeah.

No, it's the challenge is the the any tool that's decent will get such rapid adoption because of the internet being the best distribution channel ever and X even faster than Yeah. So, I think the issue is is any tool that's worth anything will just be used so much that it'll quickly lose any type of of real edge.

I have a followup on the cost side of things. Um, obviously some of these API calls are quite expensive especially if you're doing 03 level reasoning test time inference like everyone knows that it'll get cheaper.

It'll get baked down into A6 or there'll be specific chips that run these things really cheaply in even six months even a few years. the cost per token's dropping, but at the same time, we've heard about startups that are like, "Yeah, I burned through 500k of OpenAI credits in a couple months testing this thing out.

" So, uh, is there a world where you've actually seen an impact on your COGS or your margins in the short term?

And have you had to work with the board or investors to kind of think through that costbenefit analysis or have you always been able to kind of dance around it so that the business always feels like a software business and and feels like it has very high gross margins like any other tech company or is the fundamental gravity and the laws of physics around building uh a software company changed or at least temporarily changed?

I think it's too early to say and I think it's a really pertinent question obviously. Um when we started and we launched Finn like four months after chat GPT came out each resolution cost us $120 something cent and we we charged 99 cent so we decided to take a loss on that.

Yeah we don't share our margin at the moment but it's very very healthy now. Yeah, of course. Um I think businesses that like ours build very deep um applications on top of these base level models uh will be able to extract plenty of value because of what makes them unique.

Those that are simply GPT rappers will need to charge something closer to GPT prices, but model prices are just going to continue to come down. Yeah.

So in the same way that SAS companies were able to make 80% margins on AWS in the future, pretty much everyone will be able to make 80% margins on OpenAI or AWS when they run their own model. So I I I actually we'll see, but I don't think it's going to net out where the margin structure is completely different. Yeah.

Uh how much of you followed uh the the Clara CEO? Every every one one day he says, "Oh, we fired everyone. Now he says, "Oh, we rehired everyone. Next week, I think he just found the the ultimate attention hack is we got to have him on the show.

Got have him on the show, but it's, you know, basically just just flip-flop. Risk on. Risk off. Risk off. " Yeah. Risk on. Risk off. Nice. Um I don't know. Yeah, he's really good at marketing. Yeah, he is. Yeah. Uh what has your kind of founder journey been and where do you see the company going over the long term?

like uh from the outside I I think it's you know there's a lot of people that would see intercom and just think like oh like venture back they're going to sell and get rolled into something else. It seems like you've been in founder mode for a long time and it doesn't seem like the writing's on the wall for you to stop.

You're enjoying what you're doing. Uh what do you see your the rest of Yeah.

Anyways, the company the the biggest opportunity is still ahead in terms of being perfectly this is one of those things where we talk to a founder and we're like wait like if he's really going to work on this for like 20 years like it's going to be insane.

So yeah I'm interested to hear like what do you do you see all talk about your ambitions. Yeah. Yeah. I mean I'm 14 years in now which is long you know most people like success. Overnight success overnight success. Well, you know, we were an overnight success two years in and then we just did it for another 12 years.

That's great. Maybe we maybe we messed that up. You know, there have definitely been periods um certainly near the tail end of our first chapter where, you know, we were all tired. I actually was off for two years. I was sick and tried to run away and ended up coming back. That's a whole other show perhaps.

But um the AI thing definitely reinvigorated us. Totally. And you know, it's a you know, I won't deny that I'm 41 now. I was 26 when I started Intercom. It's a different thing when you're a little older and AI give or take is a young man's game.

Certainly now that the world is so competitive, people are working their damn asses off and a lot of our competitors, they're in their 20s. They're working literally seven days a week, 12 hours a day. We don't do that. So, we have to take advantage of the things that make us unique.

But the thing that keeps us going and that keeps me going is just the scale of potential. You know, our business is accelerating dramatically now. I think next year our growth rate, our annual growth rate will have doubled three years in a row. And these are very significant growth rates. So for Yes. Thank you.

That was that was all I was looking for. There we go. I mean I guys chill. He's still talking. He's still talking. Come on. Thank you. Thank you. Thank you. But yeah, for for for a company at scale, you know, in the hundreds of millions to accelerate at that pace is just insane.

then it's because of this AI stuff and yet very literally we've just scratched the surface. I mean if we're going to do the future of work then we're only getting started. What do you think about the early stage AI market the the rapper market? Uh people were very bearish on it.

Um then we saw OpenAI acquire Windsurf for $3 billion and it feel it felt like you know it's very special company very incredible team but it felt like it could be game on for every big tech company needs a rapper or a or a startup in every single category and so all of a sudden you could see sure maybe some of these companies don't become massive public companies but if if a founder out there is building a narrow AI solution for something that can slot very neat neatly into AWS or Azure or or Microsoft Office.

You could see kind of game on in the M&A markets and some good outcomes for the founders and maybe even the VCs. What do you think? Yeah, I mean from what I hear their M&A activity is is picking up. Certainly from our perspective, all sorts of bigger companies have started to ping us. Um it's it's kind of a thing.

You know the what I respond to this general question is is is is I just ask how how big is this market really? I mean if we are doing the future of work if you oh we lost there you go sorry if cursor wind surf are doing a big majority of future development work these things are insanely valuable.

So all of this is just way bigger than software like AI, AI applications, rappers, etc. Maybe they're 100x the total software market. Maybe they're a thousandx. Maybe that's way too much, but think about it. They're doing all of work. That's great.

I really uh VCs get, you know, so much [ __ ] but I would love for all the VCs Oh, swear jar, buddy. Yeah. Yeah. Yeah. I got to put one in the swear jar. Uh but but I would love for all the VCs to just be right this time.

like all the rounds, you know, the perplexity of 14 billion, like if you know, if they're just what if it all works out, right? Be amazing. Um, what did you do before founding Intercom? And would you do anything different or would you recommend that path to the next generation of entrepreneurs? I was starting companies.

I was doing consulting, software development, design. I don't know. There's a dozen paths, maybe 100 paths into this. At this point, I wouldn't do anything differently. You get to a certain age in life where you love all your mistakes. Yeah, maybe you like look at your younger self and you're like, "Wow, look at you.

Look what you did. " But it made you made you who you are. So, no, no regrets. And people should just follow their own pots. Follow their hearts. I love that. Follow your heart. Well, I am fantastic conversation. So excited for you and the whole team. I'm so bullish. I'm actually extremely we have these conversations.

We're just like, "Oh my god, what a beast. " Yeah. No, I mean you guys are you guys basically, you know, 14-year overnight success in the exact purpit.

Definitely in the conversation, but no, just in the perfect position to take advantage of a massive trend, but then also deliver that value back to all of your existing customers and all your new customers. Great time. Let's hit the Ashton Hall effect and thank him for joining the show. Thank you for joining the show.

Thanks for coming. Great. It's great chatting. Come back on again soon. Yeah. Yeah, we'll have to have you back on soon. There's so much more to talk about. We'll talk to you soon. Bye. And I think we have David from Box Group in the waiting room. We'll bring him in uh chat with him.

We love when a venture capitalist yaps about venture capital. It's one of our favorite activities. Uh we got a plan. I would love to get his take on what are we going to do in August because I think August might be an existential crisis for the show.

As all of you know, every venture capitalist takes all of August off and so we won't have nearly as many guests. I think we go to the south of France and just do man on the street interviews. Yeah. What do you do for like, hey, what do you do for a living? Multi-stage venture capitalist.

I think we have to have any getting in a scuba diving suit and swimming up to boats and say, "Hey, what what do you do for a living? " Let's bring him in and and we'll get his take. David, welcome