Embracing AI: Lessons from Intercom's Journey

Thank you. It's a pleasure to be here.

I think it was a combination. So practically, I think our whole industry had been playing with the playgrounds that OpenAI were offering. And Fergal, who's our VP of AI, had been keeping an eye on it. And then obviously, November 30th, 2022, I guess ChatGPT dropped. And I think it only took four or five back-and-forths with the chatbot, ChatGPT, to realize we'd crossed some sort of perceptual cliff where now people could start talking to things and getting real answers.

Fergal played with it that evening. He posted out a tweet, he messaged me, we met the following morning, and we all kind of played with what was now possible, and we were pretty blown away. It was very obvious at the time that it was going to have short-term, very useful things for CS agents. So that evening, I had a call with Eoin. It was evening time because he's in San Francisco and I'm in Dublin. And I kind of outlined what we had seen and had that sort of—you kind of need a sort of founder intuition for this type of thing.

He was like, let's go all in. What are we waiting for? And on Monday, we started to work. We released our first features in January. I think it was Inbox AI, so doing things like classical CS jobs like summarization immediately. All this has become commonplace since then, but we were out with it pretty early. And then when ChatGPT-4—when we got our first access to that in March—that was when we started to realize, shit, this thing, whatever with helping agents, this thing can actually do the job of some agents.

So we started to build out what became Fin, the sort of chatbot that we have launched, which is putting up some crazy numbers these days. We have thousands of people using it. It's doing millions of full resolutions for people. But it was very much a—clearly we've crossed some sort of breakthrough. We still don't know where it ends, right? But it was an obvious moment, a sort of tectonic shift in what's possible. And I think genuinely customer service is very much in the kill zone of AI or whatever.

I think it's a very vulnerable space if you don't move with the times. So we felt like an urgent need to get on it. And we've always been a kind of fast-moving product org, but this really put us to our test. Yeah.

Good question. It's basically that large language models are really good at conversational back-and-forth, what I would call a give-this, answer-that type workflow, which is, read this and answer the following question, and disambiguation. Just generally speaking, what do you mean? Clarification, the pushing and prodding. A lot of previous models we played with would give up or tap out somewhere and they'd not be useful, whereas ChatGPT was both pretty quick and very good at conversational problem-solving, effectively. That's what a lot of—not all—customer service, but there's a good chunk of undifferentiated customer service that falls into that area.

Stuff like, how do I reset my password? It's not like a brand-building opportunity from which you can establish a long-term relationship. It's just a very answer-the-damn-question. And the best version of that question is not an artisanal, hand-typed, well-crafted, beautifully worded, eloquent letter. It's, here's the link to reset your password. And people actually care about immediacy more than they care about, let's just say, the personal tone in those types of things. So I just think it was very obvious at the very start there's a certain chunk of the work that's definitely doable.

And over the last while, we've just been expanding that chunk to include things like better answers, better disambiguation, multilingual. And we've got more stuff coming throughout this year that will again just increase the workload.

That's exactly what it was. The manifesto is literally what we believe about the space. And it took us quite a while to distill these things because so much was changing. But we have our sort of what we call our big beliefs, which is that CS is in need of a significant upgrade, that AI is going to be the thing to do it. There'll be winners and losers, et cetera. We've been distilling this manifesto for like two years as we've been progressively learning what AI will and won't do, and where humans are involved and aren't involved.

And ultimately, putting it out there was our way of telling the industry, right, if you choose to go with Intercom now, this is what underpins everything we build. Everything we do is set to advance this idea here. So it was really the crystallization after maybe a year or year and a half of consideration of what the hell we're all about.

Yeah, sure. So the very short summary is: when we started building Intercom, we called it a customer communication platform. It was a way of talking to your customers. Generally speaking, people talk to their customers usually for three reasons: sales, marketing, and support. And then you can add on the ancillary stuff like research or success or whatever. So for quite a while, we existed as a broad tool that could be used for any of those jobs. When Eoghan, who's our CEO, returned, he came back.

One of the points of his plan, which we can get into as a kind of reaction to the 2021 days and all that, one of the key ideas was: we have to pick one job, one specific category, and become the dominant best product, become an irreplaceable category winner. And our chosen area was customer support. So that's where we were going. We had always had sort of work in AI. We had a product called Resolution Bot, which was like a version of Fin, but you had to hand-train it.

You had to say, "Here's five examples of a question; here's what the right answer should be." That type of bot has gone through many eras, and that was just sort of the prior era of bots. When AI presented itself, it was quite clearly the future of the space that we wanted to dominate. So it might look smart. I think we moved fast, but I genuinely think that there was no other option. I think, were we sitting here without an AI strategy today, we'd be deeply worried.

It's very clear what's going to happen to most industries, but very clearly to customer service. So I think we went from a general-purpose broad tool for talking to customers, to being a CS tool, to being an AI-first CS tool.

We have this idea that if AI can answer the question, it should. And we spend a lot of time making sure that our AI, Fin, is aware of what it can and can't answer, and how certain it is. The logic for that is quite simply: an instant answer is almost always what the customer wants, and they will sacrifice a lot for an instant answer. And if you look at your own online search behavior, you go to a website, if you don't get an answer to the thing you're wondering in like four seconds, you close the tab and go back to your Google search and jump down a link or whatever.

That's just the way customers behave. So instantaneity is a really valuable thing. So we say AI-first from that perspective. But there are a few other perspectives that we think about AI-first from. Like when we're building something, when we're trying to capture a new workflow, let's say we're rebuilding our reporting. The mindset now is, can AI do this? So let's say a classic customer support leader job might be like, what new issues are occurring in our volume of support at the moment?

There are loads of ways to do that. But right now we think about how can we do that from an AI-first perspective? How can we get ahead? How can we preempt the desire and actually automatically surface the answer before it's even been asked? It's very much about how we think about our support model, but it's also kind of how I think future startups just need to think. Most of the things that we do online are like judgment calls, criteria selection from a set.

They're like maybe some degree of search or RAG sort of type stuff. But ultimately AI can do it all. And you should first ask yourself, does the UI, the buttons, the dropdowns, whatever it is I'm designing, does that make sense in an era of pure AI where you can literally just say the thing you want and see if you can get it? So I think there's a mindset shift that a lot of specifically designers and product managers, along with engineers, need to get into, which is understanding that we shouldn't be approaching AI from just simply, how can we automate a task? Or how can we automate a little step? Or how can we summarize this paragraph?

You have to actually look at how can we automate the outcome? If we can automate the outcome, let's start there. If not, let's start at the workflow. If we can't automate the workflow, then let's try and automate the steps. If we can't automate the steps, then let's try. You should start with the big picture and come downwards. And worst-case scenario, you can only automate a few steps or one workflow. But I think there's a lot of product areas that are going to be honestly deeply disrupted by what's possible with software now that wasn't always possible.

And I think it'll play out in loads of ways. The UI styles that we choose: will we go towards chat UI as opposed to natural language or old-school pointy-clicky UI? Workflows, do they even make sense? I've cited this example before, but I have a friend who runs an advertising optimization startup, and he's quickly realized most of the reasons people log into his product aren't necessary in an era of AI. The AI can generate the ads, optimize the ads, publish the ads, turn off the ones that aren't working, find the winners, iterate on the winners.

There's very little reason for a human to log in other than to get the report. But even at that, the report can be mailed out automatically. So you start to ask yourself deep questions about what is my product? And that's what I really consider to be like the AI-first mindset. So we both see customer support as being AI-first, as in if the AI can deal with it with a zero-second answer, let's do that first. But also I just think it's how software needs to be built from here on.

In our manifesto, we say we believe the future of support will be humans plus AI. And what we mean there is the AI will—so basically, some conversations will be answered entirely comprehensively by AI. How do I reset my password? Here's how. How do I get a refund? Click this link. That type of thing. Just a complete answer. Customer gets exactly what they want, and they leave. The next set will be things that the AI will attempt to answer but ultimately might fail over to a human.

And then some, the AI will just be like, "I'm not touching that. It sounds like a sales query. I'm going to hand that straight over to a human." Right? So that's the first piece. The second is there's a symbiotic relationship between the bots and the humans, if you like. Right? The AI and the human, which is when humans are in the inbox, the AI can help them. It can do things like summarization. It can look stuff up for them and all that sort of stuff.

We're investing a lot there. But then also the humans help the AI, because one of the features of Fin is you can go through all the answers it's given and be like, "Oh, you got that one wrong. Let me teach you." And you can give Fin new facts to learn from so that it doesn't make its mistakes again. So I think you'll see this kind of relationship where humans do some direct work, bots do some direct work, bots help humans, humans help bots.

That's how the whole system works for a human. For a visitor on a—let's just say, a user of one of our customers—our goal is we don't give them any of that complexity. We just ask them, "How can we help?" They say what they want. If we can do it, we do it. If we can't do it, we ask some disambiguation questions, maybe some follow-up to really learn what it is. If we can't work it out, we hand it over to a human.

When humans try to answer it, we can surface relevant articles. We can do all the stuff that would actually help. We always, at Intercom, when we started in 2011, our big first technical push was, like, messaging. People got used to messaging. iMessage launched, Snapchat launched, Facebook Messenger. So we basically built a messenger that could sit inside your product. We've always thought about this from being: how do we make it as natural as possible to talk to a business or to talk to a website or a shop or an online store?

And so we design for both sides of the conversation. We care deeply that our customers' customers have a brilliant experience through Intercom too. And that's why we'd be hesitant to offer any upfront, "Do you want to talk to the bleepy blorpy robot, or do you want to talk to the human?" You can do that. You can set up that filtering if you want in Intercom, but it's not our default mode. Our default mode would rather be: say the thing you're trying to do, and we're going to help you do it.

Don't worry about the details. That's an execution problem on the backend.

Right.

Yeah, you're totally correct. So there's been generations of shit chatbots, basically. The first generation was just button-powered bots. Like, are you trying to do X or Y? X. How are you trying to do X? One, two, three. One. It was very much like a phone tree, but in a messenger. That was Gen 1. Gen 2 was then a little bit of AI where it's like, type your query and I'll try and guess what you're saying.

And that was Gen 2, and that's what our Resolution Bot product was. And in this new generation, I think people's behavior has changed a little bit for two reasons. One, we're like 10 years into it, but two, like ChatGPT and then obviously Gemini, and who knows what will happen when Gemini goes on Android phones and Apple finally launch their thing. But I do think people are starting to realize bots are good. Like, as in, people play with ChatGPT and their expectations are growing.

So when we first launched and people thought we were a bot, people dropped immediately into what we call bot speak. So it might be like, hi, Matt, I ordered the T-shirt two weeks ago and I'd like the real blah. And as soon as they see a bot replying, they say, T-shirt refund, please, because they've given up on all the English or whatever.

They talk machine to it the same way we all learned to eventually talk machine to Google. So you got good at Googling by doing that. I think what we have seen a bit is because the conversation history is like, there can be humans and bots in the same conversation, there can be handover in both directions and all that sort of stuff. I think you're seeing a bit more people just saying they've gone back to the whole, I will tell you what you need to know and I'll trust when I hit enter, you're going to figure it out.

And I think that has taken a while, but I think we're going to see kind of a broad adoption. We're about to launch Fin working over email. So people send emails in. It'll be interesting to see how that plays out because, obviously, people tend to write emails with a degree more boilerplate or formality than they do messaging or text speak. So we'll be curious to see how Fin performs in that dynamic. It tries to match tone and style as best as it can to the inbound message.

So we'll see how that plays out. But yeah, the biggest learning for me has been how quickly the public have gotten on board with the idea that chatbots work now. Yeah. And Microsoft and Google have been helping us out a lot by running all these ads and doing all these launches to sort of say, hey, bots aren't shit anymore.

Yeah, sure. So Fin's built on GPT-4, by the way. GPT-3.5, but it didn't— it still— I remember back when we used to talk about hallucinations, like, 4 was the perceptual change for us in terms of trust and reliability. What's the difference between a thin wrapper and a thick wrapper? I also invest in companies and I talk to companies in my portfolio, and a lot of them are kind of building these thin-wrapper solutions. And I think it comes down to where do you want your differentiation?

And are you relying on this as a moat, or are you relying on this as just being basically an extra nice feature? Like, summarize this text. It doesn't matter if it's adopted or not or whatever. But there is no end of given-this-document, answer-this-question-type products in the market. And I think a lot of people will confuse that with doing the job of customer service because it looks the same, right? But I think in practice, if you're actually running a real customer service team, you have a lot of things you care about.

You care about, say, trust. You don't want your bot to go off topic. You don't want it to have a political opinion. You don't want it to express anything other than what the hell the customer is asking about. You care about reliability, you care about accuracy, you care about interaction with humans. So if you don't know, can you hand over to a human, and can the human hand back to you, et cetera? You care about reporting, like, what answers are going badly?

What answers are going good? Is the customer satisfied on the far side of this? Disambiguation. Like, the search itself. So in retrieval-augmented generation, the retrieval piece is like no joke. We've been doing a lot of work to sort of fine-tune the retrieval models here to improve the performance so we can actually get really good answers. Then you've got the whole multimodal, multilingual, multichannel working across WhatsApp, text message, or whatever. When you try to address all of those constraints, you end up building a lot of software that goes beyond the GPT style.

I built a thing that read a PDF and now it can answer four questions. That is true. But if you think you've built a sort of CS agent because you've built that, I think you probably just should brush up a little bit on what customer service is actually all about. The job is to improve the customer experience. You need to be certain that your agents are doing well. You need to be able to measure it, report on it, correct it, all that sort of stuff.

There's a lot of software that has to be built behind the scenes that goes beyond just the give-this, answer-that.

I think all of our work is pretty model agnostic. We use GPT-4 because it's best, but the very second there's a better model out there, we'll do it. I think people talk a lot about technical differentiation these days, and I personally get growing in sort of skepticism that it's actually that doable. The era of SaaS is an era of right-click, view source. Ultimately, if there's a single thing you've done, other people can do it. The job is to build a brand around having the best product.

And to do that, you have to have the freshest tech, the best-designed tech, and you need to consistently be there such that when customers choose you, they choose you not just based on where you are today, but based on your product trajectory and momentum. And I think when people adopt Stripe or Linear or Intercom or Slack or whatever back in the day, it wasn't because they were the only payments platform or project management platform. And it wasn't because, say, Stripe, you could argue, might be slightly different.

Maybe they had some commercial agreements that were hard to get. But Linear, they're building software. You can look at the UI and you can infer what's going on behind the scenes. It's just really hard to build the damn thing. And if you do start building it today, the chances are when you ship your version of it, they'll have moved on substantially. So I think the nature of moats in software, to my mind, is the following. You have the best product and you grow your brand off that, and then you attempt to transubstantiate the best brand into things like a community of people who care deeply and know how to use your product, into a software ecosystem that cares about strong integrations with your product, et cetera.

Maybe you build it, have a platform play or whatever. But the idea that some other company can't look at and copy our inbox design, that's been happening for years. All of the stuff we've built, honestly, over the years has been copied by various different chatbot incumbents. Most of them aren't still around because it's a hard thing to sustain, but that's the nature of software. And then the bigger question of where is the moat in AI? I think we're still trying to figure that out.

We're experimenting a lot internally with training our own models. We're doing all that sort of work that you'd guess we're doing behind the scenes. But I think ultimately, in the absence of any unique patent that you're going to sue everyone else for or whatever, having the best product always is probably the best strategy you have when it comes to defending your position.

Mistral. I mean, I think we care a lot about using the best product out there. We don't feel the need to be checking it every week because, frankly, you might forget to build the rest of the software that I enumerated earlier. But we have a torture test of thousands of questions and scenarios that we run any given model through. To run it through, you can do it in a reasonably automated way, but there's a lot of human eyeballing to really make sure that we're assessing things correctly too.

So given a model, let's just say Anthropic's Claude and Mistral and Llama and GPT-4 and whatever we might roll ourselves, we look at their performance against all of these things. And there's obviously trust, reliability, accuracy—all that sort of stuff are, like, our first concerns. Can it actually do the job? The secondary concerns then become things like speed and cost and all that sort of stuff as well. At the moment, we're on GPT-4 and we're happy there, but we have not moved into an optimization mode here yet.

At some stage, somebody might talk to us about anything from margins to speed or performance, whatever. And that's when we might start saying, hey, right, okay, given that we've kind of built all the stuff we wanted to build around this, now let's look at swapping in different engines and seeing if we can find the fastest one or the one that we can offer at the lowest price point or whatever. But I think that that's a second-order thing. It's basically one of those explore-then-exploit.

We are still deeply exploring how much we can transform the nature of CS with AI. When we're confident we finish that exploration, there will be an exploitation phase where we start saying, right, let's turn the dials and let's get this really fine-tuned for speed, for price, for all the variables you might think about, even just stuff like hosted in the right region, data governance, all that sort of stuff as well. So they're all to come.

It's possible. Or, I wouldn't rule that out. I also wouldn't rule out different features going through different... Depending on what we're asking, we might look for the cheaper calls or the more expensive ones or slower ones or whatever. There's also other stuff like, would we let customers choose? Would we let customers bring their own? There's loads of versions of that. We've built it in a sort of high-cohesion, low-coupling way where you can hot-swap if you want.

The challenge we have is just, we put a lot of work into making sure that we're certain on things like trust, reliability, et cetera. If customers are going to hot-swap in their version, or the reason we wouldn't just overnight flip from one to the other, is because there's reputational considerations on behalf of our customers. We want to make sure that these things don't ever do anything we don't want them to do. So we put a lot of work into guardrailing them, and that's why we wouldn't necessarily be running with a bank of 10 of them, because to own any given one of them in any period.

It is ongoing work too. It's not for free.

As I said, we have genuinely a very exhaustive torture test that we put any given model through that is almost designed to trick it into hallucinating and see where it'll fail. We've also done a lot of work in getting the models to report out on their own, how would you say, certainty of the answer. And that gives us ourselves a threshold where we say, hey, we could have higher coverage, but we would do it at a cost of more bad answers, basically.

We take the obligations of our customers' reputations quite seriously. So we don't want to have hallucinations. We put heaps of work into that. A lot of the times when Fin gives bad answers, it's not hallucination, it's actually usually just bad content, stale content on our customers' help centers, that's happened. Or somebody before, like somebody yesterday, answered this question and they gave the wrong answer, and now Fin thinks it's correct, so it's gone with it or whatever. So that's where we have that idea of human helping the AI.

So a human logs in and fixes it. But there's a lot of work done to sort of guardrail, do certainty checks, to just threshold how confident Fin should be before it answers. And then there's obviously a very rich feedback loop too. Like, if customers push back, Fin will take that as a sign that it didn't get it right, and we'll surface that to the end user and be like, hey, we read this article and answered this question like this, and the user thought it was wrong.

What did we get wrong? And the human can be like, oh, shit, I better update a help center article or something like that. So we care deeply about this. The way I think about Fin in the lives of our customers who are customer service teams is that it should move them to being that they answer questions for the first time because Fin has never seen it before, but also the last time, because once they've answered it, Fin has now seen it, and Fin should be able to answer all questions of that type, which kind of just changes the nature of it.

And hallucinations are just far less likely in a world where Fin has seen everything every human has ever said and everything that's in the help center. It's really unlikely to have to invent new stuff or pull on any sort of creative thread because we're scoping it quite tightly to private docs, public docs, what's been said before by your own CS team, et cetera. That kind of scoping means that it's not reaching deep into its own creativity. And other models will claim higher coverage for this reason, by the way, which is just interesting.

If you're willing to tolerate worse answers, you can hit a higher resolution rate. But that's a very conscious trade-off that I think people aren't honest enough about. If we tell Fin to just make it up no matter what, it would probably have more hits, but just a lot more misses too.

We have a few features launching in this area soon to customize things like branding and tone of voice and all that sort of stuff, so people can really make Fin speak like their team. So a surf shop and a bank don't have to have the same language or whatever. That's one piece. The majority of the area in which we customize is really just making sure that the customers can bring all of their data sources. That's the thing that really matters here, in our opinion.

Given that we're doing RAG, we're not yet building models per customer or tuning models per customer. We're doing RAG. So what we've spent a lot of time on is most of our customers store all the information that needs to be known in Notion or Confluence or Guru or in random PDFs that are floating around your Google Drive. Fin can imbibe or drink in all of those to get the most comprehensive answer set. It can also read private docs and knows not to surface private docs to end users, et cetera.

So that's the way in which people mostly customize it, is they just tell it as much as they can about their business, and then the vector search kind of takes care of the rest. When we think that we have tapped out in performance in our current setup, it might make sense—and I really say might because it's hard to tell what the future will hold—to start thinking about things like per-user weighting or tuning or models or any of those sort of things.

It doesn't feel like that's the barrier today. And it feels like the barrier today is making sure that Fin can speak all languages across all channels and consume all content sources. Going back to this idea of a thick wrapper, that's the stuff that we really need to get great at. And we've been doing a lot in that space too.

Yeah, great question. So Intercom is an actual CDP. We're kind of one of the first. We were out before Segment and all those with our version of this. So all our users store custom data on their users, and it would be things like what price plan is Matt paying, and how many teammates has he, and how many files has he uploaded, or whatever makes sense for your business. All that information is also fed to Fin as well, and Fin can use all that as well.

There's obvious sort of value we get out of that. So if Fin knows that you're a premium customer, it can give you the answer to your question assuming you're a premium customer. So Spotify has a different set of features for premium users than it does for regular users. And if it knows Matt is, like, a premium user, it'll give Matt the right answer. More generally, when we think about personalization, I'd sort of say there's levels of automation that we sort of work through.

You can think about this the same way we went through, say, self-driving cars. In the '80s, there was no automation. Today there's a lot of automation. We're, like, maybe L4 or something like that in support, that you go from no automation to maybe old-school bots to LLM bots, dynamic answers, the idea of looking up Matt, looking up Matt's information, and giving Matt a specific thing. Then you can go a step further and be like, let's look up Matt inside Stripe and work out where his last invoice was.

And let's answer a question based on that when he says, "Why was my credit card charged?" All the way over to, like, let's issue a refund to Matt because he expressed discontent with the service, all the way up to, like, let's proactively start a conversation with Matt. That's, like, the sort of spectrum of how much automation we can build into the CS job. Personalization is one of those sort of steps that we want to get to, which is answer this question for Matt specifically: his order history, his refund, his account plan, all that sort of stuff.

That's absolutely on the roadmap. We'll have a lot more to say about that within months or whatever. But we have genuinely a sort of full spectrum of how we're going to automate support to get from our current, say, 40%, 50% total automation up to maybe 70s, 80s, 90s, depending on the business.

No, actually. Well, the first thing we have to build, because there's two things everyone needs to know about Intercom. One is that we are a complete customer support platform. And then secondly, we have the best AI. And we ran a funny billboard about that. But the reason we had to build phone was just because people wanted to consolidate onto Intercom, but they couldn't because we didn't have a phone solution. So the first thing we had to do was just build a phone solution.

But the question you're tugging at, which is the one everyone's excited about, is, like, will Fin answer your phone calls? And the answer is not today. But absolutely, voice is still a huge amount of support volume: phone calls, or increasingly with Gen Z, voice notes as well are becoming quite popular. We've played with everything, you'd guess, Synthesia and their synthetic bots. We've played with ElevenLabs and their synthetic voices, et cetera. And in general, there's an obvious path that we're going to get to here.

I'd say the reasons we're not building it right now: one is, I think it's basically not customer demand, number one, but it will get there for sure. The latency is a little bit of a problem in a phone call, as in a good GPT-4 query might take three, four, or five seconds, and that just makes for an awkward phone call. Now, we've played with things like injecting filler, like, "Great question, Matt. Let me have a think about that while I'm running over to the server and getting the answer."

"Okay, I've got it." That type of thing can work. But I suspect, I think that's just building the wrong thing. I think if we literally just sit back for a second, all these problems will correct themselves. So I would say voice, from a point of view of robots answering the phone and hearing your query and answering it as best as they can, and if not, failing over to a human voice, I think that's all certainly going to happen. Don't have a stated timeframe for it yet.

It's definitely a valid concern. An interesting mindset that we've had to adopt at Intercom is there are cool features, and this is the first time in my career this has been true, there are cool features that we can't afford to build, right? So, as an example, summarize every conversation Intercom powers every month. That would be summarizing 500 million conversations a month. That would bankrupt the company in some amount of years, right? And we wouldn't be comfortable passing that cost directly onto our users.

So it's definitely like, how much will this cost is a new sort of step in our should-we-build-it that just wasn't there before. Related to that, by the way, so is latency, because these things are just slower than normal software. So you have to think about where and when you inject delays too. But what have we learned aside from that? I think there's a few different things. It's awesome that there are open-source models coming up that we believe we can hand over more and more of the workload to.

We haven't really gone deep down that journey yet because, as I said, we're not really in a cost optimization mode, but we're definitely growing in confidence that we do not have to pay the full highest price for every single call we make, that we have obvious paths to shedding the majority of that cost, or at least bringing it down to CPU-only, which would be effectively, in the same way we don't think about EC2 units, we shouldn't have to think about that too much. The other nice dynamic here is there are lots of players.

At the very least, all of the big players have their own offering in the space, whether it's Google or Facebook or whatever, along with OpenAI. So there's a lot of pressure on the pricing. So even with us doing nothing, everything's gotten cheaper and cheaper and cheaper. And that's without us actually doing any optimization ourselves. But all that said, when you work with large volumes of data like we do with, say, conversations, or billions of people have been, or billions of unique entities have been in contact with Intercom through its history, it's a very expensive query to say, oh, let's use some chat UI to analyze the entire conversational history for our business.

That's seriously expensive. So the dynamic we have to find is don't just build shit because it's cool. Not that that was ever really a good idea, but it's really important in this case. You have to say, where are we displacing real work that already had a significant cost attached to it, where we can do it at substantially less? And in those cases, we are very justified in charging a reasonable price. Ninety-nine cents per answer. Ninety-nine cents per answer, even for the low-end, how-do-I-reset-my-password type things. It still takes you a few minutes to do that and to close it and measure it and record it.

So I think aligning price to value is really important, and then just getting the price point correct such that you don't have a broken dynamic where people are resistant to using a feature because it's a bit expensive. You need to avoid all of that. It needs to be the case that people are turning to AI and saving money by doing so. It should not be the opposite. So there's just a bit more, how would you say, business leadership needed from a product manager working in the space.

To think about the hows, whens, and wheres of the problems we tackle, how much money is currently allocated against them, and how much you're going to cost in terms of those pricey GPT-4 calls and how they add up depending on what we're doing.

Very much top of mind here. And it's a great time for everyone to just reread The Innovator's Dilemma and to really kind of understand. And you don't really get a choice in these worlds. We can sit here and cross our fingers and hope that AI doesn't happen, but it's happening, right? And the reality is AI can answer a lot of support questions, and it can do it at a pretty cheap point. And there's no point pretending that that's not true. So there's an amount of reality that has to be embraced by anyone who's dealing with this tension between the old revenue model and what might be the new revenue model.

In practice, not everyone wants to adopt AI full beans today. People are dipping their toe. Some people, because of the depth or complexity of their service offering, might still always have a large support team. And so we need to work out a way for both. In general, if we charge you, say, $39 a seat, or we charge you $1 per answer, if that seat, that $39 seat, is displaced because the seat is no longer bought because Fin is now answering all of the work that that seat would have been doing.

As long as that number is more than 39, we're okay. But you need to have a bridge into the future. So we have to sell in the present, where a lot of people aren't necessarily ready for full AI yet, but also have a path forward. And I think we're okay in all scenarios. If all the work goes to AI or a substantial chunk goes there, we're okay and our customers are okay. If it doesn't, they can keep buying seats, and we'll continue to make the seats valuable.

I think that the challenge for startups is to— I would encourage everyone: don't rely on hope as a strategy. Take what's happening seriously and price for the new world. And for what it's worth, I do think, separate to all this, SaaS is going to be a bit more usage-based and metered anyway. I think the days of seats above all are starting to fade away. AI is going to accelerate. But I think generally speaking, people are kind of realizing that charging for value is more important than charging for empty seats.

There's a tiny bit of education. We actually consider resolution the same way they do, but you just have to explain a little bit. Let's say 1,000 conversations come into a business. Let's say Fin only touches 700 of them because it looks at 300 and goes, "I don't know what to do with that," because maybe it hasn't read the right docs or hasn't been fed the right information, or maybe they're gobbledygook or spam or whatever. It doesn't really matter. Three hundred of those it's not touching.

That's not relevant. So the first figure we'd quote is our involvement rate, which would be 70% in this case. Seventy percent of the conversations Fin jumps into. That's not what we price for. We price for when Fin has given an answer and the customer has either just closed the messenger and gone on and done the thing they wanted to do, or has explicitly said, "That answered my question." The only time we won't charge here is if the customer pushes back and says, "That's not right, this is wrong," and we hand over to a human.

That's when we don't charge. We basically charge when we gave the customer an answer that they saw and they didn't have any follow-up questions, which is exactly how CS reps are measured as well, in that no one goes chasing people being like, "Are you sure? Are you sure? Are you sure?" But that's the piece of education. Most of the time, because Fin's instant, you reply, "Here's how you reset your password." People click the link and go on about their business.

What they don't do is come back and say, "Thank you very much, Mr. and Mrs. Bot," in the same way you don't say thanks to Google after you complete a search query. So I think what we price against is literally what we believe to be the purest sort of threaded version of the thing that happens, which is, did Fin answer the question in such a way that no work was put onto your CS team? If so, then fine. We do report on things like customer satisfaction for Fin users, and our customers can see that to make sure that the Fin customers aren't more pissed off than the human ones, et cetera.

Absolutely.

To your point, we might go there, but it hasn't felt like the most important thing, in that once people understand what it's doing, they're like, "That's good enough for me." Because most of the time what they're trying to work out is like, "I don't want to buy this tool and still have to deal with the entire workload." And that's the thing they actually care about. And then also, "I don't want to disappoint or frustrate my customers." So that's why we report on CSAT.

It's a bit higher. It gets higher every few weeks. So we just shipped Fin earlier this week in 40 more languages, and that will kind of give it a higher involvement rate and probably a higher success rate too, because it was probably speaking bad English in previous cases. So that number is creeping up. I don't have the current one, but 40-ish is roughly where we're at. I will say that kind of hides—on average, I would be confident any business who turns on Fin will get at least like 25%, 30%.

We have a lot of people getting 70 and 80. It kind of depends on the simplicity of the support function that you're staffing. So if you have some sort of Pareto-style 80/20 for your support queries, as an example, like a utility provider deals with open an account, close an account, change an address, register a meter reading. Those four or five questions often account for 80%, 90% of the entire inbound. In those cases, Fin delivers exceptional results, as you'd guess, because you can target, you can deliver so much value by just getting really good at four or five things.

So the 40% thing is definitely very, very real. There are just a lot of cases where, if you have a simple enough support function, it can hit some extremely high numbers.

Internally, we have demoed these capabilities, so it's absolutely going to happen, high certainty, and we will definitely do it. If you recall the L1 to L5 of support automation, I think actions is probably one level above what we have for sale today. We are firmly going to make that happen too. It's specifically important in certain businesses where, let's just say, approval or refund requests or whatever, we can't go and do them without having a proper way to go and ping APIs. The thing that will come before actions will be third-party data lookups.

So, pulling in information and relaying it back to customers. So, why was I charged $59 on my credit card? Here's the answer. That type of thing. That's just a read. It's not an action. But actions are definitely going to happen in terms of, it's not that hard. Today, Intercom has a feature called custom actions where you can actually offer, click this button to do something, which could be refund my order or whatever.

It's the thing we have to unlock. And it's just, again, it's back to this idea of trust, reliability, and guardrails: when are we comfortable letting Fin do that off its own bat, or independently from human supervision? And I'm very confident we'll see it in months. I can't say if it's two or six, but actions is the next sort of chunk of work, if you like. Personalized answers, then actions. Then after that, you're into proactive. And I think then we're really hitting the high 90s in terms of percentages.

That's right.

AI adoption is kind of hilarious. I jump on sales calls frequently, and some larger, older institutions can be a little bit like, hey, we're just about ready to test your summarization feature. We're really excited. And then you could jump on to a new, recently Series B San Francisco startup or whatever, and they're like, when can Fin answer phone calls for me and do authentication? And we're like, whoa. So there are different approaches. The thing I notice a lot is some people are very wary of AI because human is part of their brand, or great service is part of their brand.

What we encourage those people to do is just work out a way to dip their toe in. So, like, hey, how about if Fin only works with your free customers? How about if Fin only works in the languages you don't speak? How about if Fin only works on the hours that you don't have real-time support? And in that regard, it's kind of an asymmetric upside. It's like, if it works, great, because we couldn't help that customer anyway. And if it doesn't work, well, no worries, because we couldn't help that customer anyway, right?

So I think you can always get—we can always, how would you say, seduce somebody into the full Fin experience by showing them a little bit of value in an area where it's kind of risk-free. And I think that's the adoption part we see for the bigger companies who are just more—they're AI-curious but also AI-skeptical. And that's working genuinely. And then with the startups, as I said, they're ripping it off the shelves, they're tearing it out of our hands, and they're like, why can't I point Fin at a YouTube video to read my product about yet?

And we're like, yeah, we're working on it, but give us a bit of time. I will say I really firmly believe 2024 will be the year of AI adoption. If 2022 was the year it was created and 2023 was the year we fleshed it out technically, this is go time in the market, in my opinion. And I think that's why we're excited. We have such an exciting roadmap. We're going to be pretty loud in market. Everyone will hear about it over the next while.

But I do think this is the time when every business, big or old, bank, government, healthcare, you name it, everyone who's not supposed to touch this shit, I think they're all going to be forced by the realization of what's possible to start dipping their toes in, and it's going to be an exciting time.

I mean, first of all, I'd say Eoghan returning was the catalyst for all of this. I think right-sizing costs. So we obviously had to do a RIF like every other startup that was out there. Honestly, refining the strategy, creating a lot of focus and a lot of urgency behind one thing. So moving away from being a broad tool to being a very focused customer service platform that defines the industry, and then obviously pouncing on AI as part of that. I think we also had a lot of work to do around changing our go-to-market motions, making sure that we're actually accessible and affordable and adoptable by companies big and small.

We did a lot of work there as well, changed how we kind of represent ourselves to the market, spoke to a lot more clear value. And I think in general, the framing I would use with a lot of the startups I've invested in or spoken to is, like, you have to be everything for somebody, not something for everybody. And I think that's, for us, a lot of it. We want to be businesses' first and only customer support platform, or AI-first customer support platform.

I think a lot of that was the strategy, and a lot of that is now the positioning and the marketing. And even the fact that we're having this conversation is a byproduct of the focus that Eoghan created around saying, we're going to pick one thing and only do that one thing, only talk about that one thing. I think that's what I'd encourage any startup who's kind of coming out of a lot of the ZIRP-era startups or whatever might have. I think the nature of expansive growth is it tells everyone, massively blow up your ambitions and think about all the cool things you could do.

And I think the last year or so has been one of doing something really well where you know that you can be the best. And that's what we've been doing. I think that's been what we've been all about since I returned. And I think that honestly has guided us really, really well.

It worked, I think, because, well, we had worked together before. Thirteen years is Intercom, but maybe like 15 or 16 years is the full lifespan of, say, relationship with me and Eoghan, or me, Eoghan, Ciarán, and David, all of it. So I think it was founded on a pretty strong foundation. I think we have always believed in some basic ideas. One is hierarchy. Eoghan was CEO, I was to be involved in product, Ciarán was CTO, et cetera. We always had really clear who owns what.

And I think that has held us together a load of the journey just because there's never power games at play. There's never ambiguity about who should do what. It's always very clear. Everyone knows what their job is, what they should be doing, what they shouldn't be doing, et cetera. So I think business-wise, we've always had that structure. And then obviously it's impossible to work with somebody or for somebody for like 13, 14, 15 years without actually an underlying friendship strengthening that kind of gives you the emotional backbone during hard times as well.

So I think that's how the sort of partnership, if you like, has worked. The other thing I'd just say is the desire, the emotion. If Intercom was just about making money, we would have taken the first acquisition offer or something like that. If it was just about fame, we would have quit when we got our first bit of coverage in The New York Times or some shit like that. Or if it was just about raw tech, we would have built the thing we wanted to build and be like, ta-da, we're done, right?

I think the thing that we're all really passionate about is actually making the internet better and delivering better customer service online. Our mission has always been, "Make internet business personal," really make it possible for the internet, just for doing commerce, business online, to be actually a really positive experience. And that problem has persisted for like 15 years and probably will for 15 more. But that's the thing that we're joined at the hip in chasing. It's the curse of a founder: you're both so optimistic about what's possible and then so depressed about what you see in front of you.

Right. But I think that's the permanence, this desire to make Intercom a really impactful, relevant company charging towards better internet.

So you're correct, we've been distributed forever. So when the world went remote in March 2020, was it? We weren't that shocked. And me and Eoghan, at the very least, always had a remote relationship. He was always in San Francisco, I was always in Dublin. So it wasn't like a big, "Oh my God, what do we do now?" type thing. But I think, in some sense, distribution or distributed offices was always fine for us. We had those muscles pretty warm.

Going remote, I did not like personally. And then separately, I think we're now like two days a week in the office. And I think when I look around the days in the office, there's a lot of positives to see. There's people smiling, high-fiving, there's people going for drinks after work, there's people having standup meetings by a whiteboard, really chasing down tech specs for what we're going to build. And all I see in all that is the stuff that wasn't happening when no one went to the office.

So I think I'm a big believer that face-to-face matters, not just for collaboration. It does matter for collaboration. But also I think it matters a lot. Work is such an important chunk of everyone's lives, whether they want it to be or not. The idea that you could do it and have basically no relationship with people, or that they'd be just tabs in a browser or faces in a Slack channel, I think it robs us of a real richness that society has to offer, which is, like, doing stuff with people is a great way to bond with them, to get to know them, to make friends, et cetera.

My heart used to break during COVID because people would—we're in Dublin, Ireland—join Intercom from all over Europe, relocate to Dublin, Ireland, and then they'd never see their colleagues and they'd be alone because they just moved country. And it's like, what are we doing here? So anyway, these days it's a lot more exciting just to see the office alive and vibrant. And it's the same in San Francisco too. So I'm definitely a believer that a reasonable amount of face-to-face collaboration or just meetings, in-person chats, lunches, drinks, whatever, is necessary for the sort of camaraderie that you need for a team to survive and perform and to enjoy working.

I think Europe has never failed at the opportunity to shoot itself in the foot from a tech perspective. I'm probably one of the world's biggest haters of the cookie banner initiative that has caused all of our websites to basically pop up all sorts of mindless crap that no one reads and call that progress. And somehow people in the EU celebrated that like they did. And I think I'm very worried that we're going to do the exact same thing with AI, where AI products will just, in Europe, have a load of meaningless pop-ups, or that companies just say Europe is just not worth our time.

It's just too hard to do business in. So I worry a lot about over-legislation, especially things that aren't actually focused on what end users want, but they're instead focused on what legislators think could be interesting from a legal ruling perspective or whatever. So that's the first thing I'd say. The second thing I think is, if I had one criticism of the sort of startup people who invest a lot of time in creating European startups, I think it's mostly we're really good at creating lots of startups in Europe.

I could name hundreds, or thousands, maybe not 1,000, but certainly hundreds of performant, let's just say post-10 million ARR or credible businesses that are growing and delivering value online or whatever. I think we twist ourselves in knots about the idea that any one of them gets particularly big. I don't see a line of sight to a trillion-dollar company in Europe, full stop. And I think that's because governments or the EU or whatever will pounce on it because it seems like just deeply incorrect or whatever.

And because of that, I think we'll always be chasing the shadows of Silicon Valley, which is disappointing. But there are some signs of optimism. I think Mistral is going to be exciting. I think it's a really relevant player in the AI space. And then there are other, just in a totally different sort of corner, there are folks like the team at Amo in France who are building, I think, a really interesting alternative take on the future of social networking. And I'm really excited to see how that plays out too.

So there's a growing amount of ambition. Usually it spins out of European founders who sold their previous company for a significant, let's say hundreds of millions, to a US company, and have now realized, why did I have to do that? So I hope that we'll see Europe support these companies more, encourage them to be Fortune 10 trillion-dollar businesses based in Europe, not penalize them, not legislate them, not banish them to cookie banner AI pop-up hell. And maybe then we'll actually start to very credibly see a super strong, thriving EU tech ecosystem.

Right now, as I said, we've got hundreds of great startups, and that's a really necessary start, but it's not sufficient. I really want to see trillion-dollar companies.

Cool. Thanks so much, Matt. I really enjoyed it too.