Speaker unverified: belief

Transcript context

…hing where we expect our AI products to be flawless and we're totally tolerant of humans showing up only speaking one language and only working six hours or whatever. It's just, it's a funny contrast. Yes, yes. That's interesting, and it's interesting you talk about product onboarding here because I think of Intercom as you guys have a house view that product onboarding should be much better. And I remember a lot of the use cases you would talk about for the original Intercom—talk to your customers through the website—was that you can have personalized nurture tracks, and it's weird that you drop people into SaaS products and just expect them to be able to use them right, and you should see how people are using the product and then give them kind of specific steers based on their usage. And it sounds like you're coming to this vision again, which is: people should have better onboarding support, people should be nurtured along based on their use case, but now kind of interactively AI-pilled. Yeah, I mean we talk a bit about the city of what is ultimately a customer agent going to be. That's what Fin will be as it grows up. It'll just become this way in which customer conversations are handled. And obviously the most direct attack here is customer service. I think every single customer touchpoint can be improved by basically immediate accurate answers available all the time. Isn't an obvious limitation—right now, you require customers to come up with a prompt. And if you look at why TikTok is so successful, it's like I would never prompt for, “I want to see videos of planes landing low over the beach in St. Martin,” but it turns out that's what I want to see. It's reacting to what customers say. Exactly. A customer has to come along and type things into the box. I mean today that's what customer service is. It's still kind of like, here's my problem and then we'll solve it. I think for sure there's obvious directions this will go as like, hey, what does a good customer look like? And maybe we can honestly infer that as well.But certainly “people like you should do things like this” is definitely an understandable domain, and then I just think working at the right level of interruptive help. You don't want to be too naggy or too pop-upy. It gets quite grating. But I think if you can get the first message right, you can sort of say, “Hey, if you come here, you're always going to get the thing you should do next or the thing that looks like you're stuck on.” If someone's on, I dunno, the renewal page and they have an error message, we know they're probably going to open the thing and we know they're probably going to say something that's got a lot of the context already there, so we can work out the right things to say and do. I think that's pretty doable. It feels like you could do a lot around, you train a model on what the customer is seeing on that webpage at that moment in time and use it to feed the answer and things like that. at's pretty doable. It feels like you could do a lot around, you train a model on what the customer is seeing on that webpage at that moment in time and use it to feed the answer and things like that. We already do a lot of that already as in customer context, so knowing that it's John and he's on the premium plan and he's on the playlist page and there's an error on the screen, that’s all useful information when it comes to—because I think people, a lot of the YC, “I could build that in a weekend” type Hacker News crowd, I think one of the things they often they're thinking every customer support query begins with like, “Hi there, my name is blah, my user number is blah. ” But actually most support conversations begin with “This is broken.” And you're like, “What's broken?” And so to solve that you need a fat reply engine that's just like, “Hey, let's chat about what's going on here.” But we realized quickly people will disengage. So any amount of extra context that when you say this is broken, and if someone says this is broken and there's a big red error box on the screen, we're like, “Well, it's probably that thing that you're talking about.” A lot of people just don't realize how deep you have to go to actually do a great job. And we do say if you install Fin today, you get 65% resolution rate after 30 days. That's shocking, but we have had to go really deep to actually get to those numbers and it involves all sorts of—every single smart thing you can think of we've had to do and then optimize and then find the right model for and all that. But one of them is customer context and that obviously answers a lot of things. What are the other smart things? nk of we've had to do and then optimize and then find the right model for and all that. But one of them is customer context and that obviously answers a lot of things. What are the other smart things? Abstraction. So I guarantee you've got no pages on your website that say “Stripe works really well for a dentistry office.” You probably don't have that in your docs. A very naive rag bot will be basically like, “Well, it doesn't say dentist and we're told not to hallucinate, so no, we don't do dentists. Sorry.” And the abstraction is in that case is well what is a dentist? It's a type of business. Does Stripe work for businesses? Can dentists be internet businesses? Well, we say we're great for internet business. So you're kind of working out what's the best risk-tolerant way to make grounded inferences without going over the cliff? That's one of, like, 27 different components of Fin. Then you've got obviously your rag and then you've got, you're like, “Hey, is an escalation appropriate at this time? Have they threatened something?” Every single type of problem, it kind of ends up you have to walk through it all to actually recreate customer service. I think a lot of times people will compare it with, “How do I reset my password? Ha! It found it!” And you're like, right, cool. That's like 0.4% of the scenarios you deal with when you're in customer service. There's a pattern. Did you ever see the movie Armageddon where—it's like Bruce Willis and Ben off or whatever—but the gist of it is they train a load of, I think it's like oil drillers to become astronauts and the comedy, the joke at the time that Ben off always says he got drunk and he recorded the voiceover for the DVD and he was like, “I always said, well, why didn't we just train the astronauts to drill oil? Surely that's an easier problem.” I think the thing that we're realizing with the AI movement is some version of what's going to happen sooner? Will AI people learn how to do CS or will CS people learn how to do AI? Thankfully, as I said, we kind of started off with CS and AI in our DNA. You say people, you mean companies in this case. Will OpenAI get better at customer service faster than Zendesk gets better at AI. Exactly, exactly. I think we were lucky in that we kind of had already backed both horses somewhere along the way.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.