Name: How Anthropic Delivers AI-first Customer Service
Uploaded: 2025-02-13T18:16:05.421Z
Duration: 59 min 12 s
Description: How Anthropic Delivers AI-first Customer Service

Transcript for "How Anthropic Delivers AI-first Customer Service": Hello, everybody. Thank you so much for joining us. We are about to get started. Welcome to our live webinar, how Anthropic delivers AI first customer service. We're about to go right into the introductions and our discussion, but just wanna go over some really quick housekeeping items first. So while I'm going through these, I already see people are doing it. We'd love it. Go ahead and introduce yourself in the chat panel on the right hand side of your screen. Let us know where you're joining from. And in that same panel, you're gonna notice there's a q and a tab. That's where you can submit any questions that you have. You can upvote existing questions that you wanna see answered. You can do that now or throughout the webinar, and we're gonna leave plenty of time at the end to answer those live. We also put some new resources in the docs tab that you can access. So one is our brand new 2025 customer service transformation report. The other is a downloadable toolkit that you can use to put a lot of what we're actually discussing today into practice. And there's also a link to the brand new, case study Intercom just released with Anthropic. So if you want to dive a bit deeper into their story and their journey, you can go ahead and check that out as well. And then the last thing I'm gonna do is open up a quick poll. I'm gonna pull it up here on the screen. You should see in the right, and there's a little poll tab there now where you can go to submit your response. We wanna hear from all of you, what are or maybe what were your biggest concerns about implementing AI into your support strategy. You're welcome to suit choose more than one, so go ahead and go in there, vote, and we'll review those results, before we go into the live q and a in a little bit. So done talking, done with all the housekeeping. With all of it out of the way, Emily, thank you so so so much for taking the time to join us today. Can you just give us a real brief introduction? Absolutely. It's great to be here with you today, Mark. Hi, everyone. My name is Emily Lambert. I lead our product support team and commercial operations team here at Anthropic. Before this, I was leading support teams at Stripe. And, excited to chat with you all today about the future of AI first customer service. I'm also based in New York City. Nice. Yeah. I forgot to mention them myself, from from joining from San Francisco. So, a lot of coverage here. I love it. Really excited for the conversation. So let's go ahead and get into it. One of the reasons I'm so excited to to hear from you and your insights is really because of your your role at Anthropic. You have this really unique front row seat to the insanely rapid changes that AI is bringing and not just in kind of the world broadly, but specifically in how it's evolving the customer service landscape. And so to kind of set the stage, I'd love to get a sense of just what customer support looks like at Anthropic. So could you maybe outline some of the key challenges you face, especially when it comes to scaling customer support? And I know there's a lot of diversity with your user base and product offerings. So we'd love to hear just what does support look like at Anthropic. Absolutely. Yeah. We do we do face a number of support challenges at Anthropic. First of all, we do have a wide variety of users. So we have free users who are using our cloud.ai product. We also have a bunch of paid users, some using our cloud team and enterprise products, others using our paid API product, and all of those users require tailored support strategies. We also face challenges in how to provide support to users at scale. So previously, our human specialists were serving both our paid and our free users, and we found that that just wasn't scalable. So we did make a change where now our free users are primarily supported through, AI first strategies. So our help center and fin, which I know we'll talk about today. And then we've also faced a number of challenges related to spikes and support volume. This is often tied to new releases. So for example, when we released our new model SONNET three five back in June, we saw really drastic week over week increases in support volume. And this was before we had FIN. So we faced major backlogs then, but we've become better equipped at dealing with those as well. We'll chat through today. Yeah. And we're gonna go into a lot of those details, but it's it's always interesting hearing even companies like Anthropic. A lot of the challenges resonate think, with a lot of other types of companies, big and small, especially around the idea of different products, different types of users, free versus paid, predictable, unpredictable spikes in volume. So we'll dive a lot deeper into into a lot of that. But I kind of have to go a bit elephant in the room. I have to ask why Anthropic, one of the leading AI companies in the world, decided to partner with Intercom instead of building your own AI support tool. So we'd love to just hear some of the factors that led you to partner with Intercom and how that's been going with you, throughout the the past few months. Absolutely. Yeah. The build versus buy decision was a big one for us. I think first of all, it came down to trust. So once Intercom began leveraging Claude, this was a huge unlock for us because we really felt like we could have the best of both worlds. We could use Anthropic's best in class, safe, reliable cloud models that we absolutely love while also benefiting from Intercom's customer service expertise and commitment to AI first support. And that expertise piece for Intercom was huge. Like, we knew that Intercom was a true expert in this field, and we wanted that. We wanted a company that was obsessed with customer service and constantly working to make Fin better, more accurate, more reliable. And then the other really compelling piece for Intercom and Fin for us was speed. At that time, we were facing massive backlogs, like, in that moment, and so we couldn't go with a solution that would take us either six months to build or six months to set up. Then we were ready to go within a week, and that was the kind of immediate impact that we needed. Yeah. I I love hearing all of that. It's great, you know, being at Intercom, but it it's really interesting to to hear about that kind of speed of innovation as well. Like, people, I think, can understand the partnership side. But, being able to go from you know, we'll talk a little bit about help center content because you've got some cool insights and tips there. But assuming that you have a lot of this content ready, the ability to go from having kind of, like, no AI frontline support to having that with Fin is is very short, especially compared to building on your own or some other tools. So, dive a bit more more into that. I wanna kind of talk a bit about the predictable versus unpredictable volume spikes, because, you know, whether you're a massive company, small startup, support leaders are often dealing with those predicted and unpredicted spikes in volume. And it's something that it's, I know, a huge factor for Anthropic. And so we'd love to hear about how Anthropic and you all prepare for predictable volume spikes, like product launches and, things like that or or new models releasing versus the unpredictable ones that can kind of catch everybody off guard. Yeah. I mean, I would say in the AI world, even something that's quote, unquote predictable still has a lot of uncertainty associated with it. So we find that even with planned new product releases and new model releases, we do have challenges in figuring out what volume is going to result and what user support needs will be. Still, for those predictable, changes, we can prepare the team and we do so. So this is really led by our amazing support operations team. But first of all, we're preparing our teammate, Finn. So we do this by updating our help center content. We are creating new snippets for Finn. Some of those snippets might just go in a folder where it's just for, you know, temporary usage, where it's just related to that specific launch. We create tailored FAQs. And then we're also preparing our human specialist team. So our support ops team is doing training for specialists and just making sure they have all of the enablement materials they need to be successful. For unpredictable spikes, that's where we really lean into AI. And we need that, you know, cushioning and buffer that Fin provides us to just ensure that even if we face a spike, we have a backlog, we can recover a lot more quickly with a lot less stress on our human team. Yeah. That that's great to hear, and it's, a a theme that we've kind of heard time and time again, which is kind of a a bifurcation of two workflows. One is the hard work of preparing help center content and snippets and making sure that you are kind of, like, operationally set up on on the technology content side. And then there's this this almost, like, internal team enablement aspect of it, which is obviously not just when you're implementing AI for the first time, but what this looks like, how we might wanna do this, making sure they understand here's what Fin is prepped for, here's what we have, here's information and resources that you have available to you to be able to to kind of handle those. And so, that's very, very similar to a lot of the other trends we've heard from from other customers with high support volume as well. Yeah. And so kind of a bit of a follow-up one to that. You know, given where you are now, you've been using Intercom and Fin for some time. How confident are you today that your team and and Fin can handle, you know, a sudden hundred or 200% increase in in support volume? Yeah. I think my answer there really depends on kind of the point in time that we're at. So we see constantly evolving, FIN resolution rates. We see very different kind of questions that we start to receive based on the particular launch that we have. And then our team is constantly fluctuating in in size as we grow. So what we are committed to is after each, spike, whether it's a %, two hundred % increase in volume, we are committed to very quickly reviewing, that volume coming in and seeing where is Fin really well equipped to handle this and where could we kind of better enable Fin to take a larger share of the volume. But we know we could still end up in a backlog situation, but we're confident that it's both gonna be less severe and just last for, a much shorter amount of time because we have that buffer in place. Yeah. And and you you had mentioned something that, a couple of other folks I brought on to different webinars have talked about, which is, their own internal systems and processes for reviewing, these different kind of, like, spikes in volume because, what we encourage people to think about is each one of these is a learning a kind of like a learning opportunity. It's a learning opportunity to figure out how you can be improving your AI system, your your human support, but also just identifying where you have gaps, especially for, like you said, even if it's a predictable release of a new product or a new feature or a big announcement. If you're in ecommerce, maybe it's a sale, you can kind of prepare, but you start to see, oh, here's some types of questions that we thought we could maybe answer, but we couldn't. So kind of like a quick follow-up on that. For that, like, apart from each volume spike, is that also a part of your kind of regular cadence? Are you going through and doing that on a regular basis as well? Yeah. Absolutely. So we, see optimizations as a part of our kind of support ops charter and what they're doing really day in and day out. So at least on a weekly basis, they're going to be reviewing, the volume trends that we're seeing, spikes that they're seeing, and just seeing, okay, are there gaps in content? Should we create a new help center article for this? Should we create more snippets from this? What is missing, and how can we best equip Fin to continually drive up that resolution rate? Love that. Yeah. We'll get a little bit more into that nitty gritty, in a bit. For now, I wanna talk more about kind of some of the initial success that you saw in this kind of broader AI adoption, and support strategy. And so, you know, Anthropic's leading AI lab. So, obviously, AI is your DNA. But I think it would be really useful for people to hear how you all went about implementing and use utilizing AI a little bit more in the weeds. So, you know, you you make the decision to to buy versus build. What were some of the initial next steps you took during kind of that initial setup and testing of of Fin before you went fully live, and what were some of the results that stood out to you? Yeah. So we we love using Claude, so we were absolutely going to involve Claude in this process. And, we use Claude to really set up the help center content. We had an existing help center, but to kind of refine that content, to also create structured snippets that could best prepare us for the FIN launch. And I'm happy to go into kind of more detail about how we use cloud a little later. But just in terms of some initial results that we saw, we were really pleased with them. So we saw a 96% fin involvement rate. We saw a climbing resolution rate. So right when we set up fin, we were at a 34% resolution rate. A couple months later, we were already at a 50% resolution rate, which we're really excited about. With Fin, we've seen tens of thousands of queries resolved around the clock, which has been really encouraging. And then I think we did a calculation this week that Fin has already saved us about 6,000 human support hours. And that's monumental for our team. That really means that my human support specialist can focus on higher impact, higher value work. And it also means that I can scale my support team nonlinearly with, support ticket volume growth, which is really important. Yeah. What so many amazing kind of, like, stats in there. And and for folks who might not, be aware, when we talk about an involvement rate, that means the number of kind of the those inbound queries that Fin is is involved with. But I I love I mean, hearing that you you turn Fin on. Right? And you started at 34%, which is still insane going up to over just over right around 50%, I think you said. Yeah. And then I I think you'd meant, you know, six thousand hours is a lot. That's amazing time that that people have back to work on, like you mentioned, the really impactful work, and that's only gonna gonna increase. You know, we're being able to respond to queries around the clock. I mean, all of it's, it's it's great initial results. And, but I you you did mention leveraging Claude and and leveraging your own tools and things like even out obviously, outside of intercom, to help with your your strategies. So let's, like, maybe dive into that a little bit more. Can you give us some insights into how your you and your team are using Claude to develop your help center content and those snippets? Yeah. Absolutely. And for this major shout out to our support ops team. Isabelle and Ashley do a great job with this, but they really started just by creating a Claude project. So within that project, they would set custom instructions, things like you are a support professional and you're building content for a support team to be used in this way. From there, they uploaded a style guide and that would have specific formatting instructions of how we like our help center to look. So use this this type of heading and use bullets here and use numbers here, added that to the project knowledge base so it's just stored anytime we need it. And then, we really got the prompting. So for a particular new feature or new change, we would share a product requirements document. We would share any marketing materials that were available. And then from there, Claude was really able to write those articles for us and then we'd review them and do, some human tweaking. But, Claude has gotten really strong at this type of behavior. In the setup process, we also use Claude to review our most commonly used support macros and then help us develop content based on those. That that's fantastic. And I I highly encourage everybody who's who's kind of listening. Don't sleep on leveraging tools like Claude to help you and your team develop content. It's, you know, you what I what I really like, Emily, is you you talked about not just sharing, you know, existing support documentation or something that one person creates, but you've got the PRD, you've got the marketing materials, all these things that, like, different teams are creating, pulling them together, and and letting Claude and and, to to kind of develop the these kind of, like, starting points, really. I mean, maybe not starting points, but relatively fine tune starting points that humans can review and then and then test. Exactly. Yeah. It definitely makes it much more scalable for us. Yeah. And you when we were chatting before, you had mentioned kind of, categorization of of different types of queries. Can you talk a little bit about how you've you've leveraged Claude to to support that as well? Yeah. This is another one of Claude's strengths. So we would provide Claude with kind of a massive list of preexisting topics that we had, and Claude's able to very easily categorize those into kind of our main topic drivers, our subtopics. And then also, leveraging cloud, we built the classifier that kind of describes each of those topics and subtopics that then we can just continue to use subsequently. Love that. So we all know that kind of, like, building any implementing any kind of AI support tool is not a set it and forget it solution. It does require kind of constant, consistent refinement. How are you and the team at Anthropic continuously optimizing Fin and your AI performance, now? And then maybe what are some of the in future future improvements that you're most excited about? Yeah. This is a constant effort as you noted. So I I mentioned we have an amazing support ops team. They're doing a weekly performance review of how Fin is doing and then constantly looking for optimizations. We've also partnered really closely with Intercom here. So we leverage a bunch of Intercom's features. We've experimented with things like the unanswered questions report where Fin is falling short, also reviewing AI suggested snippets and seeing if we should incorporate any of those into our content library. Another cool thing we did back in November, we actually participated in a company wide hackathon, and our support team participated by just really focusing on resolution rate optimization. So we use that time to have, our specialist team review and reorganize our snippets, update help center articles, just sprint on content gaps and figure out how we could, fill those. And through that hackathon, we actually, I think, doubled our number of snippets in the course of a week and also improved our resolution rate by about 5% just from that one sprint. So we do have plans to do kind of content sprints periodically throughout the year. And then the thing we One one question on that. And I just you the how how long was that hackathon? Was it like a single day kind of hack hackathon? It was, really concerted work over, I would say, like, three or four days. Yeah. I I I just wanna shout that out because I I feel like the it's a perfect I know, like, not maybe every company is doing, doing hackathons. But being able to do a quick sprint, you know, taking two to three days just to go really heads down. I mean, being able to double the number of snippets that you had, increasing a resolution rate, especially when you're handling a large volume of queries by 5%, that's a pretty massive lift. And when you think about the hours that the team gains from that and comparing it to the hours it took to invest in it, that's a pretty strong ROI. Yeah. Absolutely. We we definitely, will replicate this model throughout the year and just plan plan these, I would say, at least quarterly. Love it. Sorry. I think you're gonna continue on to maybe some of the the future improvements before I interrupted you. Yeah. The one thing I would just also note about the optimization is we care a lot about getting our overall team involved. So our human specialist involved, they are the the experts on all of this. So we like for them to be involved in both identifying the content gap and creating suggestions for, how we can improve our content to continually increase resolution rate. Yeah. That that humans being augmented by the AI and working with the AI. I love that. Yeah. That's huge. And then, Mark, you did ask about, future improvements and where we're leaning into, what we're excited about this year. I think the most exciting thing we're doing with intercom now is actually in, the actions that Fin is going to be taking on our behalf. So we are hoping to launch automated refunds by the end of the month, and this would be huge for us. Refunds right now are our biggest support volume driver, and they take a lot of human specialist time because it involves integrating with an outside system. But we are working with Intercom team to automate those and have Fin be able to handle the vast majority of those. The other priority for our team this year is in international growth and scaling, twenty four seven human support. So for that, we will be relying on Intercom's development of tools around translation in the inbox and how we can use that. Yeah. I I love that in the you you mentioned fin actions. And for those who aren't really sure, the fin actions are kind of, one of the kind of relatively newer capabilities of fin to not only answer questions that are coming in, but to then be able to go and actually take actions in other tools or systems. And so Emily was talking about somebody might come in and say, I'm having an issue and think they might get a question answered. But if somebody wants to get a refund, it used to have to take an individual person to go sign into the system and issue the refund. And, that like a lot of companies, you can usually it's like that eighty twenty rule of, you know, 20% of, like, the inbound queries are kind of causing 80% of the the time that a lot of reps are taking up. And so, start to think about well, you know, when you talk about categorization of support queries, if everybody listening, like, think through what those are, where do you see the most time being spent, and if there are different ways now, like, explore some of the options of Fin to be able to automate a lot of those actions if you, be able to integrate with those systems to save another big bucket, of time. And you also mentioned the twenty four seven global human support. That's, I I think, like, a huge value add for a lot of people that they don't necessarily think about right away. But when you start to realize you can serve serve, like, very strong, accurate, personalized answers literally around the clock and in in things like 45, 50 languages, that saves so much time for that team, like you mentioned, to be able to serve your larger enterprise customers or or really focus on the human centric support. Absolutely. Yeah. So, follow-up question kind of around that. What would you say your advice would be for support leaders to kind of balance this AI optimization with scaling the human team? Because you mentioned you are gonna be continuing to scale your your human team as well. Yeah. We are we are hiring actively now. We are, growing our human team. And I think the most important thing here is just really figuring out where your human specialists excel and what they are uniquely equipped to handle. And for us at Anthropic, that looks like, our team is great at handling the most complicated and technical questions. They're great at the questions that require a lot of customer context. And then also, they're great at serving particular strategic or enterprise users that might just wanna talk to a human, especially on particular topics. And then from there, once you have, you know, what you're reserving for humans, really trying to see if AI can handle kind of everything else. And for those places where AI can't yet handle it, asking yourself or asking your intercom team, like, how can we chip away at this and, get AI and Fin better equipped to handle these other types of issues? Yeah. I love that. And another kind of follow-up, what are some of the types of roles that you're actually specifically hiring for on the human side? So, one, you can get a shout out to anybody here that might might be interested in applying. But, also, for because I one of the things we've noticed is a lot of companies are, yes, they're hiring for kind of more of the traditional, you know, technical support specialist or enterprise support specialist, but we're also seeing people starting to hire for, you know, AI conversation designers or people whose roles are very specifically aligned to the the kind of, like, AI tools and that human AI collaboration. So what are some of the types of roles that y'all have have open right now? Yeah. Right now, our focus is on more of the traditional, you know, technical support specialist. We're hiring, an international support manager just to expand our presence outside of North America. We're, hiring a number of technical support specialists based in Dublin specifically. So, building out our EMEA presence. And then later, we will be building out our APAC presence as well. So really just trying to, like, round out human coverage around the clock. But later this year, we will also be growing our support ops team, and I think those roles will look more similar to what you expressed around, you know, experts in content management and other topics like that. Yeah. Love it. And I've seen the just a reminder for folks, go ahead and keep asking those questions. I see there's a bunch of really great ones in here. So go to that q and a tab. Go ahead and, you know, ask your questions, upvote the ones that are in there. We'll make sure we have plenty of time for that at the end. So, you know, whenever we do these webinars, some folks are deep into AI. They're using fan or they're using different AI tools in their support strategy. Some people are just starting to explore AI for customer service. So, I wanna kind of get them some insights. I would love to hear more about what you have learned and the team have learned from the journey. And so maybe what are some advice you could give to other support leaders that are maybe evaluating AI tools or considering this from, the initial decision making process to set up even into the the long term success? Yeah. I think the first piece of advice is just don't wait. Do this as soon as you can. This AI first support has been really transformational for our team, and I know that we could have prevented a lot of backlogs and user pain and user waiting time if we had done it sooner. I think the other one, which is specific to us, but I think is relevant to other customers as well, is just think really critically about that build versus buy decision. If you are choosing to build in house, just know that it's going to require a lot of expertise and then not only resourcing for development of the solution, but a lot of resourcing for the maintenance of the solution. So just know that before committing to a particular path. We've touched on this a lot, but I think the focus on content is critical. So make sure that your content is in really good shape. So you have a strong help center, strong snippets before, rushing to launch FIN. And then also know that that ongoing maintenance piece is going to continue and needs to be a priority, and you need to have, people who can dedicate time to that on an ongoing basis. And then I think to the to the last question we were talking about, just reminding your team that AI enhances human work. So this is not about replacing humans, but very much, about allowing the team to focus on the higher value, higher complexity issues and just using their skill set, in unique ways that are that are best suited for them. All fantastic advice. And I know I think before maybe we had mentioned it, but when you said that, you know, starting starting now, like, there's that old adage of, best time to start is now, which maybe you should have started yesterday, start now. I can't remember it off the top of my head. But, we've heard that a lot. And, when when y'all were starting out, like, how would you have graded, like, objectively, like, your the the status of your help center content? Like, do you do you feel like it was very strong kind of there? And I'm I'm just asking that because I think one of the things a lot of people do when they talk to us about is, well, I wanna I wanna launch quickly. I wanna get this going, but I'm I'm not really sure if my my content is is there. So what would you say to somebody who's, like, not really sure if they're, like, what the quality and coverage of their help content is? Yeah. I mean, I think what was great is in our setup process, you are able to test Fin so you can get a sense of, like, what types of questions would Fin be able to respond to based on your existing help center content. I would say we we were mostly there. But knowing that it's a constant iterative process and that you're going to continue to improve your content and improve your snippets so that Finn can keep getting better and better. Yeah. I love that. You'd you'd mentioned before, like, there's, like, there's different tactics to do that. Obviously, there's plugging it in, testing it. You can do all that, obviously, before you go live. You guys mentioned the hackathon. When you were initially starting that, was there, like, one specific person or was it, like, a team doing that? Who what was the actual process of, like, going in and testing the content? Yeah. This was really driven by our support ops team. So, specifically, Isabelle on our team, Ashley helping out and just making sure that our help center was in good shape and able to handle, especially for our high volume, support topics that we had all the content we needed on on that. Love it. And, one other one I wanted to just check-in on a little bit more is the, the the maintenance. And would you say, like, I know every company, every company's be a little bit different. But, in terms of, like, an actual process for that, do you find it's best to be doing you you're doing your weekly reviews of of things, but is there somebody that's going in, you know, every week? Is it is it like somebody who spends a bunch of hours at beginning of the month? Like, when do you when are you typically doing that kind of deep dive into the help center content and the snippets? Yeah. We we are doing that on a weekly basis. That's, like, part of the kind of roles and responsibilities of our support ops team on a weekly basis. But I would also say even on a kind of, like, daily and interaction basis, it's happening because we count on our human specialists to be our kind of eyes and ears on the ground and to see, like, if a message gets passed to them, it's probably because Finn couldn't resolve it. So we're counting on them to also have an eye to, oh, was there something missing here, like a content gap where if I just suggested this piece of content, Fin would able would be able to resolve this tomorrow. Love that. Thank you. The the deeper insights, the better. This is the the what everybody loves to hear about. Kinda wanna, like, you know, step back a little bit. So based on your your current experiences that you and the team have had with with with implementing AI into the support strategy, what are you most excited about in terms of, you know, maybe first through the lens of of AI and customer support? What are some of those trends that you're kind of, or maybe just future looking advancements that you're most excited about, over the next, you know, six months to twelve months? Yeah. I think the first one is just around AI and Fin being able to handle more advanced inquiries. So right now, I know the way Intercom thinks of it. There are kind of four or so types of support inquiries that we see, And Fin is really good at the informational ones. Those are kind of straightforward, simple, information retrieval. Fin is excellent at that. The next one, which is our next frontier with Intercom, is those action oriented ones. So ones where Fin really needs to take some action on behalf of of the user or on behalf of the support team. And the best example we have for that is those automated refunds. Really excited for for Fin to take that over. And then I think in the coming, you know, months and years, it will be Fin, getting better at the other two categories, which are around, you know, either deep technical troubleshooting with users or, inquiries that require a lot of context awareness and kind of customer specific information. And I think we're going to see advances in all of those areas, in the coming months and years. Yeah. Yeah. And, that that absolutely. I know those are things that are very top of mind, and and, I know that we have them. I didn't put it in the resources section, but, maybe we'll try and find it to share out the the different kind of ways that we've broadly at Intercom categorized the types of queries. And I think where it gets really interesting and compelling is you mentioned the the first two, the informational queries, and then these, maybe more straightforward action oriented queries. But when you're already seeing, you know, FIN resolving 50% of your kind of informational queries, it gets really exciting to start to see what that number what what the upper limit of that number looks like within those categories. So if you start to assume that you're able to handle those refunds, we're gonna start to see those resolution rates really involvement rate going up, the resolution rates going up. And even if you're not able to get to those, like, deep technical ones right away or the kind of, like, white club club enterprise style, like, support, that's still taking on a massive chunk of that support volume. So it's again, can the more Fin and these AI tools can cover in terms of the type of queries, that just means more and more time. That's for the support the human support reps to be able to, you know, work on the AI or work on those high value high value customers or the really difficult, compelling questions. Yeah. Absolutely. I think, we know as a support team in order to increase that resolution rate in more of a stepwise way, we need to chip away at those action oriented ones. Whereas informational, we're always going to be getting better and better, but each, resolution rate kind of percentage improvement is harder to come by now since we've made a lot of, initial progress. Yeah. No. Love it. And, I think so I think I maybe it interrupted you. You might have been talking about another another forward looking thing. Yeah. Other trends I think we're excited to see is just we talked about this a little already, how the human support specialist role will continue to evolve. So I think, human support specialist might end up focusing more on kind of enterprise support and those really strategic users that want that white glove service. And I think you mentioned this, Mark, there are going to be lots of new roles in the support space. So, building out our support operations teams, having people focus on AI content enablement and chatbot experts and all of that, I think that's gonna be a really growing field. And then Yeah. Yeah. So sorry. I just want another another note on that. I I again, like, that resonates a lot with what we've heard from other people. It's this specialization going deeper into the the true human support queries, this AI kind of, like, AI hybrid, like like, roles. Another trend that we've seen is, a deeper collaboration with the support team and the support reps with a lot of the other cross functional teams. And we've seen that, obviously, that that's getting deeper with, you know, the r and d side or maybe the product marketing side, but we've also seen it in kind of the customer success side, especially with certain industries when we start to think about, like, ecommerce. There's this interesting blurring going on of support into this almost, like, consultative role that's more on, like, the success side or, like, a sales side. And, I'm not sure if you have thoughts on that given given anthropics kind of, like, model, but that's just another key thing that we've seen. Yeah. Absolutely. I think, as support specialists are kind of not as focused on those really transactional quick cases, they're going to be focused more on, like, the thorniest issues for customers, and those inherently involve a lot of cross functional collaboration with product, with eng, with, sales, with, customer success, etcetera. Yep. Love it. Yep. Keep going. I promise I won't interrupt you again. No. I think the last the last trend that I think we can expect to see is just better experiences for users. I think this, you know, combination of humans and AI will create better support metrics. It's going to create better efficiency. It's gonna create more accuracy so that, humans won't have to wait for those transactional questions that they just want quick answers for. And there will be more human support time dedicated for those really thorny complex issues where you uniquely need human talents. Yeah. And talking about the user experience, I think it kind of like a smaller follow-up to that. I know that the the the audience for Anthropic is very often people who are a bit more technical, a bit more, aware of AI. But what was the kind of reception by by your, like, users and customers to interacting with AI because it's something that we've had depending on people's companies. Some companies are dealing with people who are very they're they're not very technically savvy or they're very averse, that kind of thing. And so some companies have developed special strategies for how they kind of, like, relay this information to their own customer base and the people who are interacting with Fin or AI support agents. So, how was the reception for your customers and your users with with interacting with AI and and Fin? Yeah. I would say, overall, it's been quite positive. I think, as you noted, people probably expect it from us. Like, we are going to be using AI heavily in our how we operate as a company and as a team, and, we are so proud that Fin is leveraging Claude, which we know and love and trust. I think, you know, the one thing to call out is we have seen certain, really strategic enterprise customers, maybe particular people at those companies who are less comfortable and just wanna be able to reach a human. And we've also created a special path for them. So for certain users, we actually have, like, an enterprise form that they could use where they're going to be routed directly to a human. So I think, we've seen a range of responses, but overall quite positive. Yeah. And that's that's just a great tip for anybody. If if you are in that boat where you have a lot of people who might be wanting to get or or expecting to get that white glove treatment, there is a ton of functionality built into PHIN natively that you can use to make sure that they can potentially even fully bypass Yeah. Fin or any kind of, like, AI tool and go right to a human. And that becomes a lot easier when your human support team has the time to actually be able to do that. It gives you more flexibility, but I think it's also, a great tip to think about. There might be a separate, like, form or a way or a page for those people to have access to to come in and and get that support. So that's a really great, great shout out. Yeah. We have a bunch of rules like that. Like, if we're if an, internal ant so, an anthropic employee is is forwarding something to our support team, we'll wanna bypass Fin in those cases too. So there's there's lots of great functionality for that, and we can be, very, prescriptive about how we want Fin to behave. Love that. Thank you. And so, I'm just being aware of the time. I wanna make sure we have we have so many amazing questions. I wanna make sure we can get through as many as we can. So what I'm gonna do is quickly review the results of that poll. So if people haven't, you can go to the poll tab, quickly submit your response. I'll pull that up. We'll talk through that, and then we will dive right into the q and a. So pulling this up on the screen, as a reminder, we had asked what are or what were your biggest concerns about implementing AI into your support strategy. And interestingly, this is, it it's kind of a bit across the board in a lot, but I'm seeing the kind of customer satisfaction, customer experience really kind of, like, leading the way there. You know, we've we've heard a lot about internal buy in. I think a lot of that depends on the company. If you've got a very tech forward company or or, like, maybe a CEO or a leader or director that's already sold on it, it can be easy. I see this ongoing optimization, investment, security, and privacy, but that kind of customer satisfaction and experience is one that pops up over and over and over again. And we spent a lot of time, Emily, talking through a lot of that. But I don't know. Do you have any any thoughts on on some of these, results? Yeah. I think on the customer satisfaction or experiences, you know, is obviously top of mind for me as well. And I think what gives me some peace of mind is just knowing that, Fin is going to be giving probably faster and more accurate responses for the places where Fin excels. So really those informational questions right now. And then I know that my human specialists are just going to have more time to devote to those really complicated issues as I mentioned. So to me, it does seem to be the best of both worlds and and all in line with optimizing customer experience. Absolutely. Very well said. And we're gonna go I was already taking a look at some of the questions. A lot of them start to touch on, the, like, optimization, of the content and things. So, we'll dive into some of those. So let me pull that down. Gonna open up my q and a tab here. And I'm gonna start with, the one that's got the most upvotes in there right now. It's from, Gilberto. I'm hopefully, I'm pronouncing that right. Asked, how can I optimize my knowledge base to get the most out of AI in support? We did cover this a bit, but, I mean, Natalie, do you wanna take a take another quick stab at that? Yeah. I think, this is all about and Intercom has a ton of tools for this. So there are a bunch of ways to go about it. But this is all about reviewing, the results and trends that your support team is facing. So really understanding where is Finn underperforming. So are there particular topics or, specific questions that Finn is not able to resolve yet? And then asking yourself, okay. Can I create either a new help center article, a new snippet, or tweak existing content to be able to get this answered by Fin? I think there there might be some cases where, it is like an action oriented one where you need to do a little bit more, development with a partner like Intercom to be able to handle those. But really asking yourself, like, why what is AI not resolving right now and why? Perfect. Very well said. Next one here from Divya, says, does the help center content need to be different or have different versions for human versus AI consumption? Nope. It it does not. We VIN leverages the same help center that, our our customers are visiting. Yeah. And, the one thing to shout out on the product side, I know, we have a kind of conversion of Fin that lives inside of the intercom inbox. We often just we we call it the, like, AI Copilot. It's an AI Copilot. That's, an area where if you are wanting to have guidance that is kind of like an internal source of truth for reps, it might be information on the, we were talking about refunds. If you're not automating refunds, it might be your your internal policy that you have around that. So there are tools and things that you could build and use inside Intercom and Fin to be able to provide, I wouldn't maybe say separate versions of your help documentation, but definitely guidance and resources for them that they could use to then use AI to generate a response and then to to send that out. Interesting one. This is a little bit about intercom here, from Andrew. What are your top three improvement requests for intercom that would make your life and your team's life easier? I love that. Oh, this is a good one. Yeah. I think one would be around reporting. So the way that we have customized our intercom setup, we sometimes have challenges, in reporting and figuring out, okay, exactly, like, what percentage of our volume is being handled by humans versus by a thin versus what we call workflow automations. And, we've talked to the team a lot about how, you know, our our support ops team has done a wonderful job kind of hacking that together, but just having more, customizability in terms of of reporting. That would be one. The other, which is already in the works, is translation tools. So we, are working with Intercom on the ability for, within the specialist inbox to have all of those messages translated into all of the languages available. And that's something we're excited about as our business continues to grow internationally. And then, the third one I think I would shout out is just further development around, Fin actions and, integrations with more systems like Stripe and and other other systems that we might use so we can further expand, the types of questions that Fin can handle. Yeah. I love it. It's a I'm excited to share all of that internally as well. I'm sure there's some intercom folks tuning in too. I I know, I I won't speak to the r and d side because I'll get myself in trouble, but I'm I'm highly confident a lot of those things, as you probably know, are already, in the works. And, shameless plug, we'll we'll put a note in here in a little bit, but we are doing, like, a a big kind of, like, product announcement called Built For You on March 19. So we're gonna be releasing a lot of some new features, some updates. And so if anybody's curious, we'll we'll drop the link in there in a little bit, and you can check it out. But, always love hearing that kind of feedback. The next one here is from Toby. Did Anthropic consider any other support or AI support platforms aside from Intercom, and what made Anthropic choose Intercom's fin? Yeah. So, at the time, we were already using Intercom as our support ticketing system. So, Intercom was also hosting our help center already, so I think that gave it an an advantage. And then, once Intercom switched to using Claude to PowerFinn, that was a huge, unlock for us and made our decision quite quite straightforward. We vaguely looked at some other platforms, but I think the bigger decision was that build do we build it ourselves, or do we go with FIN, and Intercom? And we are thrilled that we did. And in terms of why we chose Fin, I think it was that ability to use both Claude and Fin together and have that power combined. And then we really saw Intercom as a leader in the AI first, customer support space. I think that's how you all, kind of present yourselves, and we saw that oh, something just fell in the room. We saw a lot of thought leadership there, and we saw that, you all were constantly innovating on FIN and trying to make, resolution rate higher and responses more accurate and customer experience better. And that was the kind of innovative partner we wanted to be working with. Makes me makes you warm and fuzzy here and all that, but the, I I we we love to hear that. And, one kind of another thing to to let folks know is, on our website, if you go to, like, intercom.com/fin, there's actually a lot of benchmarking studies. One of the things that we're very proud of is the performance of fin. And we know that, like, we can sit here and tell everybody all day that, like, we're great, we're the best, but what it really matters is, like, doing a, you know, a competitive head by head, side by side comparison. So not only do we encourage everybody when you're evaluating Fin to put Fin to the test against other tools out there, check the performance. I mean, if you find that Fin is not gonna perform as well as another tool, like, then don't go with Fin. Right? But I we're confident it will. We we're publishing we publish all of our information, and I I believe we're working on some some additional, like, third party evaluations as well. So, when you're starting to think through the the technical resolution side of of the the power of the tool, definitely check out some of those. But we encourage people to do those head by head head to head comparisons as well. This next one here from Margaret. We talked a little bit about this, but one week onboarding is quick. Yes. It absolutely is. What did you do to prep launch, and how did you monitor train, and post launch? I know we talked a bit about this, so maybe just kind of, quick touch quick touch on that. Yeah. High level, we really focused on content here. So we made sure that our help center was kind of in in tip top shape, and, we then created a whole bunch of snippets. And for the snippets, we were really looking at our most commonly used support macros and, making sure that we either had help center content or snippets to cover all of those areas. And in terms of monitoring and training Fin, resolution rate is really our North Star here. So as I mentioned, we're tracking this on a weekly basis to see what trends are there and then, seeing what questions Finn is struggling with and where we can plug content gaps to to improve that. Great. And there's another one here from Carlos, that's more intercom related, so I'll I'll do my best to answer this one. As people might know, we we don't charge just for usage of FIN. We charge by resolution. We think that does a very good job aligning our incentives to be building the absolute best product, and not charging you when a response is inaccurate or wrong. If you, Carlos and anybody else interested, there is documentation in our help center that defines exactly what a resolution means and what is kind of like the the the component for that to actually be charged for. And so, I can't remember it off the top of my head, but that information is very clearly defined in there to give you a sense of what that looks like. And, the team is always working to improve this. This is something I know that we are internally always working on, which is to be able to identify not just did somebody abandon a conversation. We don't just count. You know, that that doesn't really say much. So we're we're constantly working to figure out, the level of of, like, satisfaction inside those. So there's gonna be much more to come on that that we'll be be able to announce, but definitely check out that help center. You could use Fin to find that if you needed to, to kind of define those. Let's see. This one from Vince. How do you ensure a seamless transition from AI to human support when needed? Again, I know we talked a bit about some of the the workflows and different things, but, again, if you wanna just quickly touch on that too. Yeah. I think this is all in terms of how our our visitor, and message center visitor center is set up. But, the Finn is absolutely the first line of defense for pretty much all users. And then if Finn is not able to enter a question, a, user can quite easily select, like, I need more help, and they'll know, that they're waiting for a human specialist in order to get that additional assistance. Fantastic. Another one in here from Ian. Does your team use Fin to automatically classify the category or subject of the conversation and or other attributes? If so, how has this impacted the way you route conversations, if at all? Thank you. Yeah. This is actually a place that we are actively, innovating on and working with the intercom team on. So, there are within intercom, there are, like, AI categories that we, are using to some extent, and then we also have, some of our human specialists selecting categories manually. We have used Claude a lot to figure out what are the right categories and how should we kind of, like, organize that taxonomy of topics. But this this area, I would say, is still in development and figuring out how much we leverage Fin versus our humans to do some of that classification. Yeah. And just a quick fall into that, I I think, the it's almost it's relatively easy, I say, to do this kind of broad categorization, kind of like the the types of support queries. But every company is a little bit different, and especially they're different when you start to look at the different kind of, like, routing rules and workflows and decision trees that you wanna have internally for where and how conversations get routed. And so I'd encourage people to kind of even if you're you you can use AI, there's built in tools in intercom, but you could use that to get yourself to a starting point. But I think it's a smart move to have somebody taking a look at that through the lens of having a human on your team look at it through the lens of your own business and what matters to you and what what is the the kind of nuance that you have to have there. Alright. Let's see here. This one from Emery about the, automation of the refund. So how are you planning on building logic around automated refunds in a way that prevents scamming? Interesting. Yeah. I'm not I'm not gonna share too much here just to also prevent scamming, but, I think this is about, figuring out which which rules are kind of generalizable enough that Fin can handle it while also reserving, a subset of cases where, we might be worried about fraud or scamming or whatever it may be and make sure that those are reserved for human specialists. Specialists. Yeah. That's a that's a great point. And, another kind of, like, note in the same way that you can think about, you know, when you're getting started with Finner, any kind of AI tool in your AI support agent, you know, it might even if it can't handle a % of those informational queries, even if you can get to that that first fifty percent, that's gonna be a huge win. And the same thing with things like FinActions and these automated refunds and actions that it might be taking, even if you can't, you know, maybe maybe in your world, Emily, maybe you find out that, like, 50% of the refund requests still need to be human reviewed or something like that. That's okay because you can if you can tightly define and have high confidence in the logic for, like, that one set, then you can still you're still saving a ton of time. You're you're framing things up, and you can incrementally improve. Like, a lot of this, it's that incremental improvement and and optimization. Exactly. There's one here from Emma, and I I think it's a really interesting question. And I'm not sure if I'm gonna be able to answer it, Emma, but it says some of our clients are not responding and leaving after receiving FIN's reply. Any practical advice to avoid that? FIN replies seem accurate and engaging. This is something that, if you if you contact me after this, I can maybe try and put you in touch with somebody for this, can maybe speak a better a little bit better to it. But what we've seen in the past is, sometimes I mean, I think I've done this before with other with using Fin and other platforms that use, Fin is sometimes, like, you get your answer and you just leave. Some people respond with, like, thanks that that did it. I know, one of the one of the ways is, I believe that there's ways to trigger, like, CSAT responses or, like, like, satisfaction responses automatically in the conversations that you could start to do. So that might be one way to do it. But, if you wanna, you know, shoot me an email after this, I'm I'm happy to put you in touch with somebody to to get you more information. Okay. We're getting closer to time, but we've got somebody. We're trying to get through as many as we can. Let's see. This one from Margaret, more about those kind of, like, the the sprints that you were doing, the, the structure of those. So I know that the hackathon might be a bit more of, like, a maybe not a recurring thing. Maybe it is. But if you're talking about these kind of, like, sprints that you're doing on the the the content, how are those structured and incentivized? Yeah. So this is still, a very nascent area for us. I would say the hackathon was a big push we did, and then we're gonna start, doing these probably quarterly. But I think this involves, doing an audit of your existing content so far. So, that might be through, like, reviewing an unanswered question report. That might be, by reviewing just trends and what you're seeing of, like, which categories of support has been, struggling to answer And then, systematically going through and saying, okay. Are there help center articles we need to create and then comparing that to the existing help center content you have. And similarly on snippets, are there additional snippets that you need to create, and how do you create snippets or tweak snippets so they're not repetitive? One thing we care a lot about is not just ballooning our content library, indefinitely because then it's harder to maintain. So we care a lot about making sure that the content is not repetitive and that if there's just, like, a slight tweak in a question that Fin's not able to answer, like, let's just add that specific case, to the existing content instead of just creating more and more and more content. Fantastic. And there's a couple of questions here about, headcount team size and a couple of them. We'll try and get them out with with one. So one was more about, if you're able to share the the size of the human team right now, and then another one is, was about a separate question was about how big is the team or who is taking part in those kinds of, like, hackathons and reviews. Yeah. I'm not gonna share the the exact size of our human team, but I will say, it is much much smaller than it would be without AI. So I think that piece about being able to scale nonlinearly with, support ticket growth has been huge for us. At the same time, I mentioned we are growing internationally, and we still know that we want some level of human presence twenty four seven in addition to Fin being available around the clock and in all all the languages. And then in terms of who's participating in the hackathon, we really just pulled the team to see, who is interested. I think there are certain human specialists who especially gravitate towards more of the content creation and the support ops type work, and, we just sought volunteers for that, and and a number of them were really excited to participate. I love that. Thank you. And I I think we have time for maybe one more. This is kind of about that kind of conversation design. And and for for those who don't know, when we talk about conversation design, it's a person going in and kind of building out a lot of the logic and the the things that are happening with how people are interacting with a, with Finn or with a human or a combination of both. I I know that you all are doing this, but do you have, like, a dedicated, like, maybe a role, or is it fitting into a a separate type of team right now? Is that the support ops side? It's the support ops side. Yeah. Everything related to our intercom instance. So, the kind of all of the rules on the back end, all of the flows, all of FIN optimization, all of that is handled by our support ops team. Fantastic. Thank you. And so I know we're coming up at time. Before we wrap, I mentioned it before, but, mark your calendars for our brand new built for you episode dropping on March 19. We're gonna stream that on LinkedIn live. A lot of the stuff Emily was just kind of talking about and alluding to, you might hear more about that and a lot of other really cool stuff. So, it's like a one click register. It's super easy. So you don't wanna miss that. Go ahead and check out that doc section. I'd mentioned it before, but we have our brand new 2025 customer service transformation report. It's ungated. You can just go browse it, but it's a lot of really useful insights in there. A lot of the things that we've kind of covered, are defined a little bit more in there. And there's also that kind of companion toolkit you can you can download. That's a really good starting point if you're looking to start to implement some of these things or you wanna start to look more at the optimization and and rolling out of of AI tools. That's a really good practical resource that'll help with that. We're gonna be emailing everybody after this. I think it's going out tomorrow morning, a link to the recording, all of these resources, the anthropic case study, and we also have a bunch of, cool clips, a summary of everything. So, keep your eyes peeled for that. But with that said, Emily, thank you so much for taking time out of your very busy day for the for the discussion, for all your insights. We we really appreciate your time. Absolutely, Mark. It was my pleasure. Great chatting with you today. Great. Thank you. And everybody else that joined us and tuned in and for the thoughtful questions, thank you so much, and we will see you all again soon. Bye, everyone. Bye.