Building AI that works: How to navigate the AI frontier as a product leader
Intercom’s product leaders discuss how we went about building Fin, our new AI chatbot, and what we learned about managing ambiguity, uncertainty, and risk in the fast-moving world of artificial intelligence.
The arrival of ChatGPT has completely transformed perceptions of chatbots – goodbye rigid, pre-scripted exchanges; hello dynamic, natural conversations.
Of course, we jumped at the chance to build a GPT-powered bot that could truly engage with customers, providing useful answers and resolutions to their queries with minimal setup. The result is Fin, built on OpenAI’s GPT-4 and our own proprietary technology. We’re incredibly excited by the initial results our customers are seeing – more than 400 Intercom customers are getting set up with Fin, which has already served over 250,000 answers, and some customers have seen up to 50% better resolution rates.
Getting to this point, however, has required an entirely new way of building product. It’s not easy to determine the right product-market fit in a tech landscape where things are changing every day – it‘s a balancing act of constantly assessing how to make progress based on very limited, uncertain information.
But that uncertainty is exactly what makes building an AI product so exciting, nerve-wracking and, ultimately, satisfying. In today’s episode of the Intercom podcast, we share what we’ve learned so far, how we ran the Fin beta, our sense of current market dynamics, and our predictions for the future, in so far as anyone can predict the future right now.
In today’s episode, you’ll hear from:
- Des Traynor, Co-Founder and Chief Strategy Officer at Intercom
- Rati Zvirawa, Senior Group Product Manager at Intercom
- Brian Donohue, VP of Product at Intercom
Short on time? Here are a few key takeaways:
- Striking a balance between speed, accuracy, and pricing remains a challenge in the space. What will companies be willing to pay? What sort of expectations will end users have around AI chatbots?
- By having a shared goal, empowering decision-making, and prioritizing speed, we were able to get Fin into the market quickly, cementing our position as leaders in AI for support.
- Advances in generative AI have unlocked a whole new world of product possibilities. Startups can leverage that opportunity by focusing on near-term value or picking existing problems to build on top of.
- In the rapidly evolving AI landscape, it has become increasingly difficult to ascertain a product’s competitive advantage and what differentiation will matter in the long-term.
- Customers seek trustworthy brands that will navigate the changing AI landscape so they can rely on their expertise as the market discovers where value lies.
- AI holds the capacity for democratizing access to complex software, potentially allowing users to use natural language to interact with sophisticated tools, rather than traditional UIs.
If you enjoy our discussion, check out more episodes of our podcast. You can follow on Apple Podcasts, Spotify, YouTube or grab the RSS feed in your player of choice. What follows is a lightly edited transcript of the episode.
From concept to reality
Des Traynor: Welcome to the Intercom podcast. I’m joined by two of our product leaders, Rati and Brian. We’re going to talk entirely about AI, our new product Fin, and generally what it’s like to be a product leader in the era of AI. Rati, let’s start with you. Where is Fin at? It’s been live for a few days.
Rati Zvirawa: Yeah, it’s an exciting week for us. We have over 400 customers now getting set up with Fin. And looking back at the numbers that we have with customers, we’ve served over 250,000 answers. These are powered by AI, and we’re super excited to see the results. For customers getting started, we’ve seen up to 50% better resolution rate instantly, which is so impressive.
Des: Resolution rate meaning the customer probably got what they wanted?
Rati: Exactly. Just seeing the reactions from customers and getting those questions resolved is exciting for us.
“What we optimized for initially was building something that previously was not possible, but we wanted to do it really fast and get our hands dirty”
Des: Brian, could you talk us through how Fin came about?
Brian Donohue: I can. Even though so much has happened so quickly, the reason we’ve been able to move fast is that we’ve actually been working on this for years.
Des: When you say quickly, how quick?
Brian: Well, so we’ve been working on Resolution Bot, the old version of Fin, in a way, because it’s the same core product proposition – automatically answer your customer’s questions. We’ve been working on that for years, as well as other ML products. But then ChatGPT opened everyone’s eyes, including our ML experts, Fergal and his team, to what the new tech was capable of. So, from November 30th, that was where the team was, “Let’s validate. Is this really as big a deal as we think it is?” Within a week, they were like, “This is a leap change.” And from there, it was going full-on. What we optimized for initially was building something that previously was not possible, but we wanted to do it really fast and get our hands dirty.
And so, what we optimized for was the AI-assisted inbox features. And a bunch of people have built similar stuff here. Summarization, making it more polite, all that sort of stuff. Why? Because it’s easy to do, and some of it is cool, and some of it is actually pretty useful. The big problem was, of course, hallucinations. ChatGPT is amazing, but it makes stuff up, which is a problem if you’re building a customer service product, right? And it was a couple of weeks of the team building conviction. From early February, Fergal was like, “I think we can actually get over this hallucination problem.” And then, we put those inbox features on pause and went all in on what became Fin. What was that mid-March?
Rati: Yeah, mid-March, March 14th.
Brian: On March 14th, we announced a prototype you could play with. We said we want credibility. We’re not just hand-waving products that you can see screen grabs of. You can play with this. And from there, it went from one small team working on it to, “Let’s actually build this product that we can sell.” That was kind of what brought it up to this week.
Des: What do you think made it happen fast? Was it like fingers on a keyboard? Was it the speed of decision-making? Was it simply the clarity of saying this is a priority? If you had to repeat that speed, what would you do?
“Is the product legit? Is it actually that good? That was phase one”
Brian: There were two phases of it. In the first phase, Fergal was really protective of the team because this team worked so well together, and they were so efficient at both testing the technology and testing out the product proposition of it. He was very protective, “We are able to move fast because we have this tight team.” And he was like, “I am the dictator here. No one is allowed to say anything to my team.” I Slack-channeled Rati, and I’m like, “Fergal’s probably going to kick you out tomorrow, just so you know.”
Rati: He did.
Brian: He did. You got invited back soon, and okay, you were officially welcome to the fold. I probably got kicked out at that point or something. So, it was actually quite a protective, “If we’re going to move this fast, we need to be almost a wall around the team.” Is the product legit? Is it actually that good? That was phase one. Working in there was just ripping up how you worked before and how your plans were going. Each day was, “What has changed today in the world of AI in our narrow world, and what are we learning?” Every day, you start with how much things have changed since you stopped working on this last night. And then there was phase two, which you can maybe better articulate because you went from a small team to that point of, “Okay, we are going to productize this, and that requires a lot more folks building.”
Des: The small team was validating, “Hey, the LLMs can actually do the thing we think they can do.” And then, stage two is, “Given that this bit is now ascertained, let’s build a product around it.”
Brian: Yeah, and there’s a nuance, but it’s important. It’s not, “Can the LLMs do the thing,” but rather, “Can we put our system of control around this? And will that work to solve the fundamental handicap, which is hallucinations?” This is where there are so many product judgment calls made by our technical experts. And this is, I think, why you really need your engineers as product engineers in this space. Because there are so many critical decisions that need to be made early on to figure out where there is an actual opportunity versus the state of the tech not being good enough given those handicaps.
Redefining the bot experience
Des: When did we go to beta, and how does the beta work?
Rati: Yeah, one thing I want to emphasize with folks is that the product was working on March 14th. We were already testing internally, and if people did visit our site, they could play around with it. Soon after, we knew we wanted to go into beta to validate this and to make sure it didn’t hallucinate. So, a week later, we started onboarding customers onto the beta to validate if it did what we thought it would do and if it would resolve conversations. Getting those early customers onboarded into the beta whilst we were building was key for this process.
Des: In what ways would you say betas are more essential or different from a standard meat-and-two-veg SaaS feature, like photo load or something like that?
“And you have to get to the end of the beta with confidence that the product works but a willingness to know that the landscape’s going to continue to change”
Rati: In B2B, I think there are a lot of features or capabilities you bring in. Of course, you’re starting from the problem you’re trying to validate. You might have a solid hypothesis of what the outcome will be when you go into the beta, but when you’re building and there’s a lot of change happening in the market, and it’s new to customers, and they’re excited, you have no idea what their perception is. There are different ways that customers are perceiving quality. You have to understand, from a business, are they willing to put this in front of the end users? What is the reaction from the end users? And you have to get to the end of the beta with confidence that the product works but a willingness to know that the landscape’s going to continue to change. Where do you make the call that you’ve hit the right product-market fit to start selling? So, the nature of those betas is quite different than what we typically do in B2B SaaS products.
Des: It’s also interesting to see how customers are learning how to evaluate whether or not the feature works in a sense. There are trade-offs between how accurate, how reliable, how trustworthy, how fast is it, and what it costs. And it’s really hard to understand customers’ weightings of those variables. And then, it also probably changes from a B2B versus B2C, a small number of customers that are very important to a large number of customers. Did we learn anything about how people think about this in general? What does one expect of an AI chatbot?
Rati: I think yes. One interesting thing we learned very quickly is that in the bot world, we’ve had a lot of these keyword-based bots. And for customers coming in, I think the initial perception was using keywords coming in and using this AI bot. But very quickly, what we’ve been promising in the industry is this conversational way of interacting with bots and seeing that moment where customers are excited to see their end users be served by something more conversational. I think that was a key learning that we had.
“We’d ask them, ‘What’s the most common question that comes into your support teams?’ Seeing an answer come in with their own content was the immediate ‘aha moment’ for customers”
Des: What was the aha moment for customers? What made them believe? Everyone’s been skeptical for good reasons, as in yes, if the sentence contains the keyword blah, “Oh, it’s replied blah.” It feels like we’ve crossed some perceptual cliff with our customers where they’re like, “Oh shit, this thing actually works.” What is it that causes that realization? Is it your own content?
Rati: Yeah, I’d say it’s our own content. It was interesting, all the beta calls and the inter calls with customers – we go in, there’s excitement. There are a lot of big questions, and everyone’s trying to talk about the future of AI in the calls. And then, we’d present Fin to them. They’d start asking questions. We’d ask them, “What’s the most common question that comes into your support teams?” Seeing that answer come in with their own content was the immediate ‘aha moment’ for customers. And that’s really what’s driven us. This is a key moment for customers to see that this thing works. You put your content in, and AI can really answer and solve those questions for your customers.
Des: Brian, what did you make it about?
Brian: What’s interesting and what I think is validated through some end-user testing is this before and after sense of the state of bots, or maybe it’s old and new bots. We’ve actually heard, in the end-user testing our team did, that people were like, “Oh, wait, is this an old bot?” And let’s be honest. This is our bots as well, right? And a lot of end users are like, “I don’t have a great sense of those. I’m not so excited to engage with this. There’s probably a fair amount of friction and hoops you’re making me jump through. Oh wait, is this a ChatGPT bot?” That’s different. That is something that, as an end user, I’m willing to do. So I think this will really emerge.
It started because ChatGPT has had such broad usage, so there’s this end-user perception of quality. I don’t know if people will say, “I feel like I’m talking to a human.” But I feel like I can have a normal conversation rather than, “I am talking to technology, use keywords, click these buttons.” Actually having natural dialogue is a huge shift in the perception of quality that’s going on from our customers and from end users as well as the core product proposition, which is to automatically answer your question with no setup. There’s that conversational quality which gives almost this gloss, this whole shine around the thing that really changes the flavor of this technology.
The need for speed
Des: It’s almost like the advent of ChatGPT has re-legitimized bots for the first time, such that everyone is willing to give them a new try. Let’s talk about building product in this era of AI. What’s changing, Brian? What’s it like?
Brian: It’s been quite a ride. Really since December, it’s been like, whoa, this is back to startup mode. That’s been a refrain that rang true for everyone. And what does that mean? I think the first thing is speed. We aligned on this in December and said, “We want to be the leader in AI for support.” In order for us to legitimately have that claim, as well as having the product, we need to have the speed of actually getting that product into the market. We think that, where this is at now, speed is absolutely critical. And that’s not a truism. Speed always matters. Inside here, we always talk about how can we move faster, challenging ourselves on moving faster. But just last week, I saw someone tweet about, “Hey, Apple’s not all about speed. They’re happy to come second to market but come with the best product, and then they’ll clean up.” It’s not a truism because speed has lots of risks in terms of product quality, product proposition, in terms of people’s head space and working health, and stuff like that. But we’ve all aligned on it being about speed, of moving fast for the product here.
“You can actually use speed as your anchoring purpose: ‘We want to get what we think is a game-changing product out fast'”
If everyone’s kind of aligned towards that and towards being really ambitious and aggressive, and we’re kind of willing to tear up how we work and how we should work, things are going to be messier, and you kind of rip up the process. We have too many Slack channels, but we have some amazing Slack channels that eventually exploded in size there. Now okay, you’ve got to go back and clean it up. But I think it’s the optimizing for speed.
Another critical ingredient here is an absence of politics. An absence of people saying, “Hey, I need to have a strong voice here. I want to have an opinion here.” People just aligned behind, “Hey, what can I do to help?” And if that means writing a little FAQ document here and then getting the hell out of the way, okay, that’s what I’ll do. It’s hard to know what you can do to help, but if you have that spirit, you can actually use speed as your anchoring purpose: “We want to get what we think is a game-changing product out fast.” That can really galvanize people in a way that’s actually incredibly energizing. It’s a really fun way to work, but for some people, it’s too chaotic.
Rati: We spoke a bit earlier about how we started the core ML team. We’ve got a core team of folks where there’s one decider who’s validating the core concept of Fin. We’ve now scaled to, “Okay, we’ve got a product. We want to take it to beta. We want to take it to market. We’ve got marketing folks that we need to onboard, and enablement teams that we need to work with each other. But we still want to maintain that same level of speed.” So, we’ve moved now from one core team, and how do you scale it to the rest of the group, to the rest of the business for more folks to go on whilst maintaining speed? What we tackled there was having the common goal that everyone understood where we wanted to get to with Fin. Giving people the autonomy to make decisions.
“I think that’s important when you’re building in this space with AI that’s much different from what we’ve done before – to allow for the messiness and be okay with the messiness”
Des: When you say where we wanted to get to with Fin, was that scope and timeline, was it like, “these features, this date”? Were we willing to give on anything?
Rati: Timeline was a good anchoring point for us. You can pick scope, and then let scope dictate your timeline, or you can pick a timeline, and that dictates your scope. We picked a timeline. A timeline was important, because as a business, we wanted to move fast, and we wanted to get something to market and put it in customers’ hands. So, we picked a timeline and then validated what we could get into that timeline. And that was really helpful for us to stay focused. We’ve got this timeline. We’ve got this scope that we think will be really impactful for customers. Run. Giving teams the autonomy to make decisions and accepting the messiness that comes from it. I think that’s important when you’re building in this space with AI that’s much different from what we’ve done before – to allow for the messiness and be okay with the messiness.
Brian: There’s another critical point that you touched on, which is the speed of decision-making. Everyone knows this, but like so many things, it’s easy to say and hard to do. When you have everyone aligned that speed to market is critical, when we’re aligned with what we’re trying to do, even aligned with the go-to-market goals, which I think we did better for Fin than elsewhere, when you can build that muscle of fast decision-making, it is amazing. Eoghan McCabe was able to give clarity on top, “Here’s what we’re trying to do.” That also had a ripple down effect.
“There’s no good in me running around, telling people that speed is important if it takes us six months to decide what we’re going to charge for the feature”
There are hard, knotty decisions to make and people get in and grapple with it. And instead of circling back and around, getting more information, getting more folks, and not reaching a consensus, if you get that muscle strong, it’s energizing. People can be not happy with the decision, but often, they’re just happy we’re moving because of the sense of progress. When you’re on a bus and you feel you’re moving fast, it’s a very exciting place to be. I knew there was a meeting on Monday, and I was going to be flying there. My only chance was to write up the doc, share everything I could and get ahead of it because I can’t be there at the meeting and the decision was going to be made there. I tried hard and failed. But anyway, you just get on board with, “This is the pace of decision-making we need to make, and the pace of decision-making is usually more important than the actual decisions themselves.”
Des: I totally agree. I also think it’s worth flagging for listeners that Eoghan McCabe is our CEO. On Sunday evenings, throughout this whole project, I’ve been having calls with Eoghan where we just talked through what were the unknowns and knowns, all the questions we were aware of. How early should we go? What model would we use? What would we charge? If we are slow, why on earth would you be fast? If you’re slow, why on earth would the designers or engineers move fast? It’s just impossible. There’s no point in trying to work as fast as you can if you know that, “Hey, it’s going to be another two weeks before we make the decision.” Even though I say a lot that “speed is life” and all of that, this project has been really good evidence that it actually has to be in unison. There’s no good in me running around, telling people that speed is important if it takes us six months to decide what we’re going to charge for the feature or whatever. It has to be 100% together.
A balancing act
Brian: The variables here are price, speed, latency, availability, and quality. So, you’ve got all of these variables of the LLM models on top of our system around, and how we interact and get the quality of product we want, and it’s all changing really fast. So, when OpenAI released ChatGPT with the API, which is GPT Turbo, that was 10X cheaper than 3.5. So, suddenly, the whole dynamics are shifted, and suddenly, features that before were not feasible are now like “Whoa, whoa, whoa, whoa.” The whole market here has changed wildly. Then you go back to ” Wait, what is the quality? Is the quality good enough?” And it’s actually fast. So, sometimes, the cheaper models are super fast. So, you don’t have a clear one-to-one of more expensive necessarily equals-
Des: This is one of the things we learn in beta. One of the questions we need to work out, and we’ll probably spend forever working out is, what is the trade-off between the speed of the end user getting the answer and the need for extreme accuracy? Or even just beautiful text wrapped around the answer in a sense. The right answer badly phrased in 0.1 seconds might actually beat a very elegant answer in 10 seconds in terms of what keeps the end user happy.
“How do you build with the product uncertainty of what’s happening?Everyone’s trying to figure this out”
Rati: I think that’s a challenge that everyone in this space is going to need to solve because it’s so new for businesses and customers. There are also not very many strong opinions because customers are still learning what they want in the world. Something that was really interesting throughout the beta, and even now, is trying to extrapolate from people how they perceive and value quality. Going back to that point of price and not having that separate within the company as we’re making decisions on Fin, you’re thinking about the pricing. You have to think about the product because we say, “Oh, we want to raise the quality. What’s the cost of that, and is someone willing to pay for it?” “Okay, we’re going to go to lower quality, but it’s going to be a bit faster. Are they willing to pay more or less for that?” And so, it’s a continuous cycle that you have to think through and sense-check yourself about what’s worth going forward with. Are customers willing to pay? Do they value it the same way as we do?
Des: And just given the torrential amount of change that happens, it feels like all these decisions or principles or guidelines almost are quite perishable. We might have a really strong opinion about where one must use GPT-4 versus Turbo, and two days later, we might change that opinion and revisit questions all over again. How do you manage pretty extreme change?
Brian: When I look back, I think, “Oh, I got this wrong.” Trying to optimize for speed, but then, when we got new information about the availability we’re going to have and how this was going to change things, because you’re comfortable working in a small unit, effectively, you can all get on board each day. And now, we’re like hundreds of people, and so if you make a significant change based on information from yesterday, everyone’s got to be like, “Everything’s changed, and then it changes again the next week.” It’s really hard to work at scale when the floor underneath you can have such dramatic shifts.
We were talking about the messiness and the uncertainty, and you talked about the models as a core part of this because so much is changing and shifting here, but the broader picture is so uncertain. And this is also something we’ve talked about: “How do you build with the product uncertainty of what’s happening?” Everyone’s trying to figure this out.
“If you want to be working on defining the next version of products and where this industry is going, you’ve got to be willing to jump into the chaos and hope all your limbs stay attached when you come out”
For example, is our messenger going to be irrelevant because everyone’s just going to go to Google and Google will be their new assistant to actually do all of this? Or maybe Google’s going to be irrelevant. “Actually no, they’re the ones who are going to be disrupted.” The possibilities for plausible, multi-billion company disruption… It’s even beyond that, I think, the scale of that. Is all reasonably plausible to think about. We’re in this space where the macro uncertainty for plausible alternate universe realities that could be real in six months versus what we can build right now and whether we can build that on shaky ground is a new challenge for everyone. Most people like a little bit of stability to work from, and I think everyone needs to recognize stability is not a thing anymore.
Des: It’s not an asset.
Brian: If you want to work in a stable product, that means you’re working on teams who previously handled the uncertainty, and you’re riding their coattails almost of like, “Okay, you got success, and now you have a more cushiony space to do your iterative work.” And that’s a valid place that you hope your company gets to. But I think the other side is also the innovation and the uncertainty. You can’t extract one from the other. If you want to be working on defining the next version of products and where this industry is going, you’ve got to be willing to jump into the chaos and hope all your limbs stay attached when you come out. But they generally do, right? It’s just software.
Unlocking value in the era of AI
Des: Brian, if you’re talking to an earlier-stage startup, let’s say 50-100 people, and they’re about to begin their first adventure in AI, what would you advise them? If you have to give them a direction, would it be to embrace the uncertainty, or would it be to move as fast as you can? Would it be to sit back and watch and pick your shot?
Brian: I was at some of our customer conferences and talked to some folks who are also looking at building things, and they’re like, “How do you handle all this uncertainty?” There is so much value that we are sitting on in terms of what this technology unlocks that was never before possible, and it’s buildable now. Never before has someone building product had as many options. We have way too much, and we’re trying to focus and figure out how we can balance our investment. There are a hundred things we want to do, and we’re pretty confident 50 of them are going to be built and are going to be valuable. There’s too much that’s available now.
“Stuff that was not possible before is not only possible now – a lot of it is reasonably straightforward to build”
So if you’re a startup, just choose the next step. To make a big bet on how whole product models are going to change, you need a big investment. But there’s so much that’s buildable that a company can make traction on. I feel like now is the best time to be a startup. Six months ago, remember Rob was at a show and tell, Rob from team ML, and he, with the previous version of Tech, investigated, “Can we do this?” Because we knew this was valuable and he did a two-week exploration. He’s like, “No, there’s nothing here. Sorry folks. I tried to find something, but it turns out there ain’t nothing here at all.” Six months later, I genuinely think it was under a day that he just put the wrapper around this thing to get this feature. That’s where we’re at.
Stuff that was not possible before is not only possible now – a lot of it is reasonably straightforward to build. So, I think this is an amazing time for any product builder to go after a space. Choose a near-term value to get on, because it takes time to build product and commercialize, find that near-term value that you can start thinking about where it’s going in the future, but don’t get paralyzed by the fact that we’ve got so many alternate universes in front of us now.
Rati: I would also say you don’t need to pick a new thing to do. Even if I look at what we did at Intercom when GPT launched last year, we looked at the jobs and the problems we wanted to solve, and that’s very core. You don’t need to build a new thing or find a new niche – you can find existing problems and use this new technology to solve them. And then pick one. Don’t pick a ton of things, pick one thing and go after it.
“The way we’re thinking about is thinking short term. Do we have differentiation? Hey, three months would be good. Maybe six months would be good, and we’ll see what that buys us”
Des: How do you both think about the idea of moats in business? If you’re a startup, you have a to-do list or project management, and you say, “Hey, I’ve thought of a cool thing that I think that these elements still do, so I’m going to go and do it. If it was easy for you, it’s easy for everyone”. What do you think about the degree to which you’re outsourcing the cool features? And then, ultimately, the great democratization of SaaS might happen where everyone can build all the cool things you’re building too. Maybe it’s a question beyond the scope of what you actually spend your day on, but what’s your reaction to that? The idea that these features are actually done by somebody else, and we’re just API’ing into them?
Brian: The moat and the competitive advantage is a hard question because of that. We think of what Intercom as a system has and our competitive advantages within that, with data that we have, which is a huge source potentially there. We’re not unique in having that, but we’re, in a way, a smaller group of people who have some of those characteristics. So, how can we lean on that? It’s super hard. How will it change? Will it matter or not? Will technology make that differentiation moot? Competitive differentiation is part of this as well because that will dissolve. The way we’re thinking about is thinking short term. Do we have differentiation? Hey, three months would be good. Maybe six months would be good, and we’ll see what that buys us. The reality is there’s probably not a world where you can build something for a year and you’re differentiated for a while. I don’t know if that world ever really existed, but I think that’s hard.
Des: There are still some types of moats, but I think, in general, simply building a lot of products might not be one, especially when you’re not really building – you’re just building API goals into it.
Brian: Another interesting thing, though, is that for us, as builders in this space, it’s impossible to stay on top of all the stuff that’s happening. It’s actually incredibly hard to stay on top of all the stuff happening at the Intercom project. No one can really keep up with all this stuff that’s happening. Never mind the macro industry.
“Are you betting on a product that’s going to keep getting better or on a product that hopes the world never changes?”
And what about your customers? Their job is not to stay on top of all this stuff, even though they’re now being asked to do it. I think brands will increase in importance. People say, “I need someone who’s going to do the work for me, and I know this company will figure out the good stuff and make that available to me so that I don’t have to figure it out.” So your customers can rely on you to figure out what matters from this and make sure that’s in the product. Because I would be very happy if I could just offload that whole mental, “How the hell do I think about AI?” to you.
Des: I think that’s totally true. Maybe it’s just our industry, because people are not free to change help desk every few months. It’s very messy. But I think when people are picking a tool, whether it’s like, say, project management or communications or support, they’re probably making a decision that’s at least one year in length, but probably more cause migration’s now a joke for a lot of these things. So, what they’re looking for in a brand is like, “Am I picking the right horse?” If new shit drops, will these people be able to react to it and make the most of it, or are they going to be stuck forever doing press releases that don’t amount to any software whatsoever? Are you betting on a product that’s going to keep getting better or on a product that hopes the world never changes? I think the latter category will be a bad bet at a time like this.
Brian: One other thought that’s circled around my head is like product builders. This is what you hope for. You say, “Oh, the world has changed, and we want it to be in this holy-changed space.” And it is. It’s for real. It’s a great time to be building product, and we’re lucky to be working in this space. That’s how I feel, even though it’s head-wrecking, you can’t keep up, and the ground beneath us is almost forever shaken up. But this is fun. This is why you want to build product.
“Companies who want to move in the space in AI are going to have to be okay with not knowing everything and making sure you have clear goals for your teams”
Des: What’s it like on the ground, Rati? Do you end up having to just chase down Slack channels in the night?
Rati: Yeah, I’m in Slack a lot. That is true. But I also am okay. And I think companies who want to move in the space in AI are going to have to be okay with not knowing everything and making sure you have clear goals for your teams. For us, that’s worked really well where the teams have clear goals they need to go after. I know 60% of the time what’s going on, and any key decisions are usually surface that you’ll catch. I’m sure there are things that drop, but generally, we’re going in the same direction as a team. And on the AI piece, people are excited to work in this new space. It’s a great opportunity for a lot of folks who you’ve been wanting to have this exciting new thing happen, learning how ML works, learning AI, as well as getting involved in that. So, it’s quite good on the ground. Of course, there are pain points – I’m not going to paint it as a rosy thing, but it’s exciting on the ground.
Des: I was talking to a friend, and we had this thought experiment. Imagine it all just went away. Imagine that, for whatever reason, someone deleted all of the LLMs, and we have to go back to the world of SaaS. It feels so much more boring looking back now. Because even the features that are pure SaaS 101 for us, even, let’s say, ticketing in our help center, or whatever, we were still looking at them going, “At some point, we’re going to come back to you with an AI perspective.” But if this wasn’t here, in hindsight, it feels like we were getting towards the end of the SaaS cycle. We’d worked out how to build CRMs for farmers to keep track of their roosters. How many more tail-end sort of features could we build?
Des: Forgetting about Intercom, are you excited about AI in your own life as an employee of a tech company? Are there other areas or products you would like to see AI in? Are you seeing anything interesting?
Brian: I’ll take this in a different direction, which you may want to shoot down. But for me, what’s interesting, getting a little closer to the tech, is the language ability. There are so many philosophical questions that were good for college students to rattle on about with their professors. And you go on there, and it was all academic in that sense. And this has been going on from the ’50s, actually, when it was a Touring test. What does it mean to learn? What is knowledge? What is reasoning? It’s all relevant right now.
“You learn your second language quite deliberately, but with your first language, your brain was programmed to learn it in a way that evolution built”
Here’s an example of this. Ask Fergal “Does GPT understand?” Well, no, it doesn’t understand. But I saw, in a Slack thread, “Comprehension is actually kind of remarkable.” Does a robot understand things? It doesn’t do language like we do because it just predicts blah blah blah – everyone vaguely knows how this works –, which makes no sense in our heads. But then, how did you learn your language? You had no idea. Your brain learned your language. I find myself thinking way more. I remember hearing neural networks, like programs were getting neural networks, “Yeah, right. Come on, you’re not even close.” Now I’m starting to think of my brain more like the machine and I’m like, “I have no idea how I learned language. You learn your second language quite deliberately, but with your first language, your brain was programmed to learn it in a way that evolution built. It’s all relevant now.
You could go into politics. There’s all the fear of AI, which seems reasonable, but like, humans aren’t so good at voting. I don’t know, have we really earned the right that we think we have? Because I’d welcome a little more rational voting choices in the future. You can zoom up to that level. And actually, this AI makes you more reflective of our species. We always think we’re amazing and wonderful whenever someone’s competing against us. But damn, we’ve got to live up to the promise. My head goes a lot in those directions. And then with kids, seeing how they adapt to the technology as well. That for, me, are the juicy places to go. The academic questions are now front and center. It will be a fun time to go back to college now to be able to talk about this stuff.
“And all of a sudden, do you build relationships with machines?”
Rati: Continuing the tangent. For me, it’s more the relationships. Normally, you build relationships by talking to people like this. And imagine, with the advancement of generative AI, being in conversations that you’ve had with a friend, a parent, or a partner, and then being able to have a conversation there. Can you have relationships with machines? That’s, of course, not related to B2B SaaS, but I think it’s interesting to start building the space around it.
Brian: Make the internet personal? We can productize that.
Rati: And all of a sudden, do you build relationships with machines? How does that change humans and how we interact in the world? I think that’s going to be interesting to watch.
Des: I think the infinitely easier challenge, perhaps, is to build a model of yourself. Giving one of these things access to everything you’ve ever communicated digitally in any form, everything you’ve ever spoken, your voice, your look, et cetera. You can imagine having effectively a shadow bot that does your job for you 90% of the time. And that shadow bot just knows to escalate to you whenever it does not know how to do something. You know what I mean?
Rati: We’re not far from that. I’m curious to see how open we are as a world to have this integrated into us. It’s always been a theoretical conversation about machines and AI and what it could do. So, I’m curious to see, from a personal level and everyday life, how people will actually genuinely integrate this and how it changes how we relate to each other and create connections.
Brian: What about you, Des? What are you excited about?
Des: I believe in AI in so many ways. I’m actually not super excited by the sort of text generation features that everyone who has a text area has now got the ability to expand on. And yet, I think that’s a reference to the “hello, world” of AI. You just throw this out to make sure all the endpoints are working. Where I get most excited from a future-facing point of view is that I think it’ll be a great democratizer for UI. There are so many products we use on a regular basis. Workday is one of them. Kuba is another. There are plenty, honestly, where they’re designed for the administrator in the company but not really for the end user who might be affected by the software.
“The idea is you know how to do that, and I don’t. You and I could do it, and it might take us four hours, but Brian knows how to do it in 14 minutes”
So, I’m sure we employ people who think the Workday UI is great. I’m sure somebody in the world thinks it’s the right UI. But I was in a tool yesterday to approve access for provisioning for something to somebody, and it’s not that the UI is necessarily bad from a Jacob Nielsen one-on-one point of view – the drop-downs line up, the text areas line up, et cetera. It’s more like this product is capable of so many things but my specific usage of it centers around one or two things, like requesting a day off or approving an expense or something like that, and yet it feels like I need to go on a training course to know how to do these things.
The example you’d both be familiar with is Google Analytics. You’ve both experienced it at some stage. You probably were once a certified GA or approved educator or something.
Brian: Yes, I was.
Des: You have a grant philosopher at Google or whatever. And the fact that you need a certification is almost part of my point. Because you get those certifications so you can answer questions, “Well, given this interface, tell me how to find the highest-performing referrals CPC ad word that worked for us in Norway between July and August or something like that.” The idea is you know how to do that, and I don’t. You and I could do it, and it might take us four hours, but Brian knows how to do it in 14 minutes.
“The gap between being able to express what you want to do and being able to do it will narrow to zero”
What I think genuinely we’ll see happen is a lot of these UIs will disappear for the regular folk, and they’ll just type the thing they’re trying to do. I see this all over the place today. An example is Equals, which is actually built by two former Intercomers. They’re basically doing a next-generation spreadsheet with live connections to live data, et cetera. But one of the things they can do is press command+K and start typing the thing you’re trying to do. And it will work out what you’re trying to do and then generate the Excel commands that you needed to know that you never fucking knew because none of us know Excel codes or commands. It’ll do all that for you, you hit return, and you’re done.
That’s an example of what I call democratization. Previously, all of this was only accessible to people who legit knew Excel or how to use Google Analytics. And now, all of a sudden, we’re all going to have access to the same power. The gap between being able to express what you want to do and being able to do it will narrow to zero. Whereas before, there was a big clunky thing in the middle of becoming an expert at using blah. And that expertise is no longer needed. And now I think we can all say to the machine what we want to do, and the machine’s going to do it.
I’m pretty certain it will happen because it’s so much better than the current experience of tabs, drop-downs, and mouse clicks. It will happen. And when it happens, it’s a trapdoor change for the very nature of software. I’ve talked to a lot of people, and I don’t think any of us are ready for it. I don’t think we know what it means. It’s hard to imagine UI without UI. It’s hard to imagine a lot of these things. And I don’t think UI goes away. I think people still want to see what’s happening. But the actual nature of complicated UI will fade away. Just write in English what you’re trying to do, and you’re done. That is a new era of software, and I just hope I’m retired before it happens.
Brian: Interaction design is moot.
Des: Yeah, exactly. And you know who should be all over this and somehow isn’t? Content designers. There’s a huge world coming for them if they can capitalize on it.
Rati: How quickly or who would adopt it first? When that change happens, will it happen with our generation, or is it going to be folks coming up now that changes the nature of software that it will attract?
Des: I think a lot of it depends on how accessible it is as well. If you think about text-driven interfaces, they are equivalent to audio interfaces because of Whisper, the OpenAI tech. And then you think about the future of Siri, you think about Amazon’s Echo devices and Cortana and all the rest of the gang. All of a sudden, somebody who’s not even good at computers sitting in a room can now have access to digital commands of the entire internet. Order me a taxi, send me a pizza, whatever. All of that becomes just trivial. And that changes things. I don’t want to understand it, nor do I want to say I have a clue how, but I just think the whole world’s going to be different, and it’s an exciting time for sure. Thank you so much for joining us and helping us on Fin. This has been a good chat, and we’ll see you all again soon.
Brian: Thanks, Des.