#276 Ryan Wang: How Assembled is Building the Future of AI-Powered Customer Support

AGNTCY - Unlock agents at scale with an open Internet of Agents. Visit https://agntcy.org/ and add your support. What happens when AI meets the chaos of real-world customer support? In this episode of Eye on AI, we sit down with Ryan Wang, co-founder and CEO of Assembled, to unpack how AI is transforming the future of customer service, without replacing humans. Ryan reveals how Assembled went from a workforce scheduling tool to a full-stack AI support platform used by companies like Stripe, Robinhood, and Honeylove. You'll learn how conversational AI agents are handling up to 75% of support inquiries, why voice is the next big frontier, and how AI copilots are helping human agents become 15% more productive. But this isn't just hype. Ryan shares the hard economic truths behind automation—why humans aren't going away, how companies are navigating global workforce optimization, and why hybrid AI + human systems are here to stay. This episode gives you a front-row seat into how the smartest companies are rethinking support at scale. Stay Updated: Craig Smith on X:https://x.com/craigss Eye on A. I. on X: https://x.com/EyeOn_AI (00:00) Preview and Intro (01:37) Ryan Wang's Journey from Stripe to Assembled (04:55) Launching Assembled (09:49) From Scheduling Tool to AI-Powered Support (12:11) Who Uses Assembled: Companies vs. BPOs (14:57) Building Conversational and Voice AI Agents (21:10) Competing with Zendesk, Salesforce & Crescendo (23:07) How Assembled Integrates with Customer Support Stacks (25:40) The Niche Power of Workforce Management Tech (31:16) Why the Customer Support Market Is Ripe for Disruption (33:47) How Assembled Swaps Between OpenAI, Claude & Others (37:56) Evaluating LLMs with Golden Datasets and 'Vibe Checks' (41:20) Multilingual Support and the Challenge of Europe (45:11) Industry Focus vs. Complexity Focus (47:43) Voice AI: The Next Big Frontier? (50:18) The Truth About AI Replacing Jobs in Support (54:39) The Automation Paradox: Why Labor Isn't Shrinking

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  There will be this kind of like hybrid human plus AI support that companies are gonna have to to think about. Just because you can automate support does not necessarily mean it will be economically viable to do so across the entirety of the business. If you can auto draft replies for people and then 90% of the time, they're actually just sending the message without really changing it. Then that's like the path to automation. What should go to who at what point in time? Or another way to put it, how do we get the right agent at the right time with the right answer?

• 0:28Speaker 01

  Build the future of multi-agent software with AGENCY. That's A-G-N-T-C-Y. The AGENCY is an open source collective building the Internet of Agents. It's a collaborative layer where agents can discover, connect, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for inter-agent communication, and modular components to compose and scale multi-agent workflows.

• 1:05Speaker 01

  Join Crew AI, LangChain, Llama Index, BrowserBase, Cisco, and dozens more. The AGENCY is dropping code, specs, and services, no strings attached. Build with other engineers who care about high-quality, multi-agent software. Visit agntcy.org and add your support. That's A-G-N-T-C-Y dot O R G.

• 1:38Speaker 00

  I'm Ryan Wang, I'm co-founder and CEO at Assembled and for the AI platform for superhuman support. So we help hundreds of customers from startups to enterprises with scaling great support. And customers like Ashley Furniture, like Brooks Running, like Robinhood, like Stripe.

• 1:57Speaker 00

  We started in 2018, we launched in 2020. So I remember March 2020. You know, there was a Hacker News post and like there was the World Health Organization declares COVID a global pandemic. Assembled launches and then World Health Organization again is kind of like the front page.

• 2:19Speaker 00

  But yeah, we've been, we've been, so we've been at support since before the AI era. And you know, my own background is I was at Stripe. I was a machine learning engineer. I was working on fraud detection. So that was 2014, when it was an 80 person startup. And

• 2:36Speaker 00

  you know, the trajectory was like 80 to 800 by the time I left, like 2017. And the, the thesis around support was, well, really it was an observation like at Stripe, but 80 people or so, like Patrick and John Collison, the co-founders of Stripe, like, and their apartment is just over there, like a couple of blocks from our office now in the Mission in San Francisco. They would have people over to their apartment to do support tickets, like the whole company. It was called like a support rotation. And obviously that wasn't going to scale. And at 800 people, it was like multiple products and kind of 24/7 support and like different tiers and different skills. It's just a lot of complexity.

• 3:20Speaker 00

  And so, you know, the question back then was just like, can we apply machine learning to, to draw the line between, you know, the experience at scale and when it was Patrick and John and the company and the apartment. And we kind of found that a lot of companies were kind of going through similar journeys of wanting to scale great support. Slack was an early customer of ours. You could, they have the same type of story with like Stuart Butterfield and like the whole company. And yeah, so lots of companies go through this journey. My educational background, like how I got to machine learning is I, I went to the University of Chicago and I went to study economics and, and actually econometrics was the fun part. So I got a BAMS in, in statistics, but my MS advisor was this guy, John Lafferty,

• 4:05Speaker 00

  who had done a lot of kind of like interesting theoretical work. And in particular,

• 4:12Speaker 00

  some of the work around what essentially became like one of his early students was this guy, David Bligh, who did latent Dirichlet allocation. So like topic modeling. So that was kind of how I got interested in machine learning in the first place. That was like 2012. But that's me and, you know, kind of like, kind of working backwards into, into assembled.

• 4:34Craig S. Smith?

  Yeah, yeah, but 2012, that's, that's early. So, so from Stripe, you left Stripe to start Assembled? And did you say 2018?

• 4:46Speaker 00

  Yeah, yeah, we started in 2018 and it took us a little bit of time to get off the blocks. So we didn't launch until 2020.

• 4:55Craig S. Smith?

  Right at the beginning of the pandemic.

• 4:58Speaker 00

  Yeah.

• 4:58Craig S. Smith?

  And and so things were slow presumably until 2022 or 2023, is that right?

• 5:07Speaker 00

  Well, it's it's it's interesting. So it was very scary and then immediately it was off to the races. And what I mean by that is in in there was like that immediate like, okay, you know, everything's frozen. But customer service actually benefited along with, you know, a lot of other functions. Because what happened was there were these companies that were going through like, you know, the the the Zerp and the COVID era pull forward of demand. And so companies were growing at kind of like pre-AI, you know, this is kind of wild, like unforeseen rates of growth.

• 5:47Speaker 00

  And you know, one of our early customers was Robinhood, for example. And when we, when they first started using Assembled, we were actually just a tool to replace spreadsheets for the scheduling operation. And so, you know, they had a couple hundred people in customer support. And the question was just like, what should people be working on? There's no Google Calendar, you know, that works for several hundred people. And so they had the spreadsheet of like every single person, the schedule like across across the top, like days of the week and hours of the day and who would be working on what. And then it would sum together like, okay, based on Robinhood's forecasting, here's how many people they need and here's how many people that are actually working there.

• 6:33Speaker 00

  And it was a good time for us because I don't know if you remember the kind of like roughly 2020, 2021, what happened to Robinhood like? Bitcoin, meme stocks, like, they went vertical as a business. And in particular, it was the need to hire a ton of people in customer support to keep up. And so it went from a couple hundred to a couple thousand kind of overnight. And so, you know, assembled really got started in that world of like companies going through that type of hypergrowth, need to grow their support teams to keep up. And then there was another piece of texture where historically customer support, I guess, the call center, right? And that's a pejorative term, like now we call it contact center.

• 7:16Speaker 00

  But the contact center was kind of forced to go digital. And like, if you go to like call centers or contact centers, it's huge sprawling, you know, places and they're in like rural Oregon or like Cork, Ireland, or the Philippines, or, or India. And, and, you know, all of a sudden, like, you know, they've got to change the way they approach software that like people are working distributed. And I remember reading this article in the Wall Street Journal about customer service agents in the Philippines, you know, all of a sudden they have to work from home and you could hear like chickens in the background. So, you know, it was really like they had to change the way that they were approaching software.

• 7:56Speaker 00

  So those two trends from 2020 to 2022 roughly kind of really accelerated. A lot of companies' growth and assembled like kind of got started at that kind of perfect moment in time.

• 8:08Craig S. Smith?

  So from scheduling software, essentially, how did you, what, what, what is the ladder you climbed to the current product?

• 8:20Speaker 00

  Yeah, it's, it's, um, so like the broader category is something called workforce management. And, you know, I think scheduling is doing it only a half service because if you think about the operation of a, of a large contact center. So one of our customers is one of the largest delivery apps in, in the world and they have 15,000 people, even with AI, they still today have 15,000 people in the contact center. And there's a, you know, put my economist hat back on, a supply and demand question. The demand question is how many phone calls, how many emails, how many chats are you going to get? There's a queuing theory question underneath that, or really operations research, which

• 9:04Speaker 00

  is like, how many people need to be working? If you want to handle all those calls within, say, fifteen seconds or thirty seconds. And then there's a supply question, which is the scheduling one. And then that gets more complex if, you know, you've got fifteen thousand people. It's not just like a, okay, literally like a record, but you know, there's labor laws all across the world. There's people in different locations. You need to kind of line up, you know, the hours of working. And there's actually an optimization problem. So if people work like in the industry, there's something called, it's the standard five by eight. You work five days, you work eight hours a day, or a four by ten, you work four days, ten hours a day. You could imagine this kind of like Lego blocks lining up to a demand curve.

• 9:49Speaker 00

  So like, in truth, when, when I say scheduling tool, it's this workforce management problem, which is an operations research problem, which is about quantifying demand and matching it to supply, which is actually like all the way back to machine learning. So like, we talk about AI agents today, you know, we started with like automating away support, but that wasn't the key issue that people were hitting as they were scaling in twenty through twenty twenty two. But we came back to in twenty twenty two, building AI agents on top of the data that we had, on top of the kind of like understanding now of like the deep operational complexities of large support teams. And, you know, we have two billion support tickets in our database. So it made a lot of sense when, you know, November twenty twenty two, GPT launched to say, oh, actually,

• 10:32Speaker 00

  like, now's a good time for us to take the data that we have, the understanding we have, and kind of point AI at it. And so that's kind of the synthesis of like all the way from customer-facing AI agents to this kind of like really like ML kind of support operations layer.

• 10:47Craig S. Smith?

  And, and, and so the current, I mean, are you a SaaS platform? Is that how you're, you're making yourself available? And, and, um, yeah, what's the full gamut of, of, are there a series of products? Or, or is it soup to nuts for customer support? Or how does that look?

• 11:16Speaker 00

  Yeah, when we say, like, we describe ourselves as the AI platform for superhuman support. So like, there's a SaaS part, which is that kind of like the planning layer. And that's like, you're the larger your team, probably the harder your planning problem is. I think there's some nuance there because like, as more and more companies have this kind of like human plus AI agent, like hybrid workforce, it starts to be not about headcount per se, but almost about like, kind of like the amount of volume flowing through.

• 11:46Speaker 00

  And then all the way to the AI agent, like, that's not SaaS. It is kind of like this usage-based, outcome-based type of pricing as opposed to SaaS for like the AI agent product. So, so the, the full platform, you know, is, it is for customer support. And then the piece is like, how do we price different products? Or kind of like, how do we think about the, the, the style of the product varies based on the use case?

• 12:11Craig S. Smith?

  And it's a platform, who are the customers? Are they BPOs that then are servicing, you know, customer companies? Or are they, is it the companies themselves?

• 12:29Speaker 00

  Typically, the companies themselves. So, you know, if I go back to the Stripe example, you know, they work with many different BPOs or, you know, outsourcers. And, you know, they have this kind of like meta problem of, well, on the one hand, like doing AI automation, like providing, you know, an AI agent that'll be the kind of like tier one first layer, if you chat them, or if you talk to them on the phone, or if you email them. And then within the bowels of the kind of support operation, it's like, okay, for working with those BPOs, like, how are you going to manage them? So, you know, we end up by, by way of that kind of like relationship, working closely with a lot of different BPOs.

• 13:12Speaker 00

  And, you know, we have, for example, integrations that reach into a BPO's own kind of like planning and staffing system to extract the data of like who's working when and what's the demand look like and kind of like stitch together a picture for their end clients of like, okay, what does this look like across your entire BPO network?

• 13:31Speaker 00

  But less so directly with the BPOs themselves. One of the problems that you can imagine that we're, they're solving that's kind of changing over time is if you are a

• 13:42Speaker 00

  Large business that has a lot of customers with a large support operation. And you were trying to reason about like, uh, how many people or really how many resource, like this resource allocation problem. Before you were thinking of BPOs in this context of like this cost quality curve. So the cost quality curve is like you could hire somebody and train them and they'd be really skilled. And you know, that's kind of like one part of the cost quality curve. You could outsource it. And there's different flavors of outsourcing. So like, you know, there's places like the Philippines have some cost structure. There's new places that are near shore, like you can near shore or kind of like onshore, but within like rural Oregon, for example, like there's contact centers in the U.S.

• 14:24Speaker 00

  There's budding contact center in like South Africa. One of Nelson Mandela's grandsons is like creating the contact center in South Africa because it's a good kind of like entry level knowledge work job. And so there's kind of like different points on the the cost risk quality trade off of of BPOs. And then you put AI agents in frame as well. And so really what we're helping companies do is like pick a point or pick really not a single point, but like different kind of mixtures of strategies to design their support operation.

• 14:57Craig S. Smith?

  Yeah. And the agents, are you using third party agents or are they conversational agents? What kind of agents on the on the agent end?

• 15:09Speaker 00

  Yeah, they're conversational agents. So, you know, we are this open platform that so like Intercom is a customer, for example. Like they produce Finn. So, um, you know, we help them reason about like their human workforce and their AI workforce at the same time. They even have like a head of human support and a head of AI support. But we also build our own agents. And so those are conversational agents. So, um, um, you know, some of our customers like Honeylove, for example, is an e-commerce brand. If you go to their chat bubble, if you email them. Assembled will do their responses and it's deflecting 75% of the initial contacts.

• 15:49Speaker 00

  We're also over voice. We just announced a customer of ours who's actually like a very fascinating business that like I have just been super inspired by. It's a, it's a company called Retention Express and they help membership management. And in particular, they help PE roll, like private equity rollups of car washes with their like subscription management. And, and they use this for AI voice. And so what's, what's been really fascinating in this space is like the conversational AI has quickly moved past chat all the way into voice. And like the voice technology space is just changing so, so, so quickly. And I think as consumers, all of us have been there, right? What it's like, if I need help, I'd rather just call, you know, I'm willing to wait on hold to call.

• 16:35Speaker 00

  And, and, and the AI agents have gotten good enough to like be pretty, pretty useful to people. And then also to be able to like intake information. So like, like for this particular use case, you know, we're able to resolve like membership inquiries, but for like workers compensation claims. Like that's a common voice use case. We'll say, give us the information, tell us about the incident, when it happened. And then kind of like pass that to a person to finish off the rest. So you don't really have to like spend a lot of time rehashing what happened. And then also the people don't have to spend the time like asking you these real questions.

• 17:11Craig S. Smith?

  Yeah. And but on, is the platform, I mean, the, the, the voice agents, for example, are, are you using a third party voice agent in the stack? Or, or is, is in the platform, is there the place for a voice agent? And you offer a default or an other, they, people can plug in their own? Or if they're using a third party, plug in a third party. I've spoken to a lot of people recently about voice agents and they're all

• 17:50Craig S. Smith?

  not exclusively, but they're targeting customer service because that's the most obvious market.

• 17:57Speaker 00

  Yeah, I've also seen a lot of this where people are trying to build their own. Like, we were at this Eleven Labs launch event recently for, you know, their V3 voice model. And so I think, yes, like, there were a lot of people there interested in building their own agents. So we help with, like, managing or kind of really like the orchestration layer of how to route volume to humans versus to AI. But we ourselves have built our own AI voice agent. And I think the distinction there is like, you know, a lot of people could, you know, there's this kind of like, if we put like a bunch of knowledge articles in your public

• 18:41Speaker 00

  help center and do RAG, you can answer kind of like a first layer of questions and then you can stack like a text to speech thing on top. And, you know, that'll work pretty well to answer tier one support. But, you know, I think the story with Assembled is like, that may not be something that you want to do because, like, then you get to tier two. You get to these kind of like agentic workflows that are multi-step that require you to ask for billing information and then make some decisions. And Honeylove, one of our customers, just to go back to you, you know, everybody says refunds are super easy. Refunds are not easy. Like, their policy is really complex.

• 19:24Speaker 00

  Like, you pull up the flow diagram and it's like, if you're zero to 30 days out from your order, check these particular variables and it depends on, you know, the customer loyalty and your history with them and whether or not they've already given you a discount or something like that. And if you're 60 to 90 days out, you go down like a different path. So like, these refunds, like, even just to execute that, like, simple, simple tier two workflow is actually quite complex. And so then to stack voice on top of that is still more complex. And so, you know, we talk to people who are building it themselves. Some of the questions we start to hear are like, oh, actually, you know what? Because I was in this boat myself, like, building ML tools at Stripe.

• 20:06Speaker 00

  Somebody built the first pass at this and, you know, like, you know...

• 20:12Speaker 00

  We're engineers. We get excited about the technology and then you know you you built the first pass and then you move on to the next thing. And the question becomes, how do you get from 10% automation on your initial pass to that 40% automation? With like the tier two workflows and and that's where I think a lot of companies are right now. If like between, yeah, like it's exciting and fun to build the first part and then to productize all the rest of it and to kind of keep eking out more automation gains like that's where where you probably add an assembled agent instead.

• 20:42Craig S. Smith?

  Yeah. And and so you guys are not

• 20:47Craig S. Smith?

  necessarily competing with, I mean, there's another company, Crescendo. I don't know if you've

• 20:53Speaker 00

  Yeah.

• 20:53Craig S. Smith?

  Heard of them or Sierra.

• 20:55Speaker 00

  Yeah.

• 20:57Craig S. Smith?

  You're not necessarily competing with them directly, right? You're building more of the infrastructure

• 21:07Craig S. Smith?

  for people like them.

• 21:09Speaker 00

  We are. I mean, we are competing with them. Like we have a conversational AI agent through chat that competes with Sierra, right?

• 21:18Speaker 00

  Crescendo a little less directly. But in the sense that like AI agents are competing with

• 21:26Speaker 00

  any type like BPO, like Crescendo is basically like a supercharged BPO, right? Like when they were they were previously something called Partner Hero and ironically, they were a partner of Assembled's Partner Hero. And you know, everything is kind of like fighting over what we want to provide to end customers is you have support tickets and you want to make the problem go away. Great. We'll make the problem go away. So I think everybody is doing that. We are doing AI agents to do that.

• 21:56Speaker 00

  Zendesk, Salesforce, Genesis, like every single provider of ticketing systems also has AI agents to automate it away. So so yes, we are competing, but then, you know, I think uniquely, we have this texture where kind of like Amazon, right? Like Amazon, you can use them, you can use RDS for them to run the database, or they can say, here's like a bunch of different databases that you could turn on. We're kind of like that, like we have it, or you could turn on somebody else's and we could help you manage it.

• 22:28Craig S. Smith?

  Yeah, I see. So if somebody comes to you, a company, a big multinational, what's the process to integrate you guys into their customer support? Presumably, they have a BPO and they have, you know, maybe a front-end chatbot, text chatbot, and they come to you, then what's the process of integrating or replacing chunks of their process?

• 23:07Speaker 00

  Yeah, it depends a lot on the use case. And I'll say that we've seen a lot of fragmentation on where people want to start. So, you know, our key use cases would be like AI chat agent, AI voice agent, AI email agent, AI copilot to help people be more productive, and AI workforce management to kind of like manage the entire support operation. And so if you're in any of the like AI chat, AI email, AI voice, it's like,

• 23:36Speaker 00

  give us your knowledge, give us your SOPs, your standard operating procedures that we can encode into workflows, maybe integrate some of your systems, like your billing system through a Stripe, or your order system through Shopify, and then just like you can kind of turn it on and start automating regardless of your channel. If you're AI copilot, and one of our interesting customers is one of the big design brands, and it consumer design, and they care a lot about quality. So they don't want to automate. And then they also have their own internal initiatives around automation that they want to hook it up to. And, you know, the business is in this place where they want to do more with the same. They're growing their monthly active users by 30%,

• 24:21Speaker 00

  and they're not allowed to hire people, but they don't, they care so much about quality that they don't want to just like drop automation in. So it's actually make people 15% more productive and crawl, walk, run your way, like if you can make people, if you can auto draft replies for people, and then 90% of the time, they're actually just, you know, sending the message without really changing it, then that's like the path to automation. So like, that's a lot of what we see with AI copilots. You want to start here, and you're kind of like, Collecting all the information workflows that are first making people more productive, then moving to this automation. And then workforce management, I won't talk too much about, but that's that was kind of like the earlier part of the discussion around like, oh, you have a lot of people. And now you're trying to add AI agents in, and it's pretty complex.

• 25:06Speaker 00

  Like, how are the pieces connecting?

• 25:08Craig S. Smith?

  Well, actually, to me, that's, that's what sounds unique about you guys. But is I'm not versed in this. Is there a whole workforce management world out there that, I mean, you guys are growing into the, the AI voice agent and from the workforce management. But is there a world of workforce management that you're also competing in?

• 25:40Speaker 00

  Oh, yeah. And it's a delightfully niche world, I would call it. Like, if you think about Dungeons and Dragons, like, that would be like the workforce management world. And, but by the way, we had no idea about this when we got started. Like, we had left Stripe and we were like, you know, these ML people, ML engineers, and we were gonna actually automate away support with ML. And the way that we got to workforce management was by asking, this guy Bob Van Winden, who used to run operations at Stripe, ironically, he went to Bridge, and then so he's back at Stripe.

• 26:17Speaker 00

  But we were asking, what is software that you need to buy or you hate your options? And, you know, my actually co-founder came up with that. I said, you should coin that, that should be like the, you know, product market fit question, because, you know, baked into that is something that people need to have that they probably want to buy more of, right? So he told us about workforce management and there's a workforce management conference that happens every year in Nashville. It's called the Society of Workforce Planning Professionals. It's so fun. It's a bunch of people who are really interested in like, the data, the data, the like nuts and bolts of like contact centers, right? There's a Reddit for, for workforce management. And, you know, people are like, pretty passionate about like their favorite tools.

• 26:59Speaker 00

  So, yes, we're in competition there with companies that most people haven't heard of, but are surprisingly large. Like, there's a company called Nice. If you fly United, like, the first time I saw Nice out in the wild, like, not in my day job.

• 27:17Speaker 00

  Like on the United, you know, they do the little ad before they, you know, you take off and it was like Gordon Ramsay. It's like, nice in contact. I was like, what? But that's a, that's a like 20, 30 billion dollar publicly traded company that's in this, this pretty niche space.

• 27:34Craig S. Smith?

  And it's in that end of the business is just optimizing. Schedules.

• 27:43Craig S. Smith?

  Is that, I mean, essentially that's what it is. You have a couple of thousand people and you're trying to figure out.

• 27:52Craig S. Smith?

  You know, how to have them on call or, or in a chair, you know, in front of a computer and so that there's an even flow of service. Is that right?

• 28:10Speaker 00

  That's exactly right. Yeah. I mean, it sounds so like trivial, especially because, you know, like we're AI ML people, right? And you're like, it, it, it, it must be like this super complex set of AI tools. And again, I'm very optimistic about that. That's like the part of our product suite that I spend a ton of time in as well. But like, the fundamental questions of workforce management, you're like, why are we so busy? Or, you know, how come we're so, you know, over under capacity? What happens is when you line up supply and demand, people find like, oh, a bulk of our volume comes in on Monday morning. But we're kind of like maximally staffed on Wednesdays. Right? Like there's 20% efficiency gains just from like kind of lining these things up.

• 28:55Speaker 00

  And then just is really like, it's actually complex to line these things up because people have different schedules. They're all over the world. They have different preferences. They have different skills. And so it is this kind of like multi-dimensional optimization problem, actually.

• 29:10Speaker 01

  Build the future of multi-agent software with AGNTCY. The AGNTCY is an open-source collective building the Internet of Agents. It's a collaborative layer where agents can discover, connect, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for inter-agent communication, and modular components to compose and scale multi-agent workflows. Join Crew AI, LangChain, Llama Index, BrowserBase, Cisco,

• 29:55Speaker 01

  and dozens more. The AGNTCY is dropping code, specs, and services. No strings attached. Build with other engineers who care about high-quality, multi-agent software. Visit AGNTCY.org and add your support. That's A-G-N-T-C-Y dot O-R-G.

• 30:20Craig S. Smith?

  Yeah, yeah, that's interesting. And in your case, it's applying to support staff globally. Yeah. Is there... So where do you guys go from here? Are you leaning into...

• 30:36Craig S. Smith?

  Is the customer support market so huge that there is market share for all these different companies?

• 30:47Craig S. Smith?

  Or do you end up specializing in, you know, medical

• 30:53Craig S. Smith?

  customer support or automotive customer support or something like that? Or once you've tackled customer support, I mean, it sounds like this workforce management, there are a lot of other applications beyond customer support. So what direction do you see yourself going?

• 31:16Speaker 00

  Yeah, I think AI agents and customer support should keep us busy for quite some time. You know, you referenced a couple of... Fairly large, like well-funded companies that have entered this space. And, and, and I think that, um, you know, investors don't always get this right, but it is a reflection of a fairly large market. There's hundreds of billions of dollars spent a year on labor and customer support all over the world. And just to give you a sense of it, like, you know, we're just in the early days and we have hundreds of thousands of people on our platform. And over 50% of them are outside the US. So it's just a huge labor footprint.

• 31:57Speaker 00

  And, and, and yeah, putting like a 50%, 60% automation dent in that is like, like pretty consequential from an economic perspective. So I think there's plenty of market. And the other thing that's interesting about like the history of customer support is that there are a lot of really big fragmented, like it's a big software market as well. Like, um, a lot of people don't know this. Service Cloud is Salesforce's largest product by revenue. And it has been for a very long period of time now. And, and Salesforce is not even the biggest one here. Like, you know, they're like a nice company I mentioned, Genesis, came out of Bell Labs.

• 32:41Speaker 00

  You go all the way back to the 2000s, there's a company called LivePerson that went public in the year 2000 doing AI chat, conversational AI. And so on and so forth. So like, it's a big software market, a mature software market. And one that now is kind of clearly poised for disruption. Satya talked about it, Jensen talked about it, you know, from, from, from Microsoft and Nvidia respectively. Like, um, you know, what are the applications of AI that makes most sense? Coding, customer service. But, but, but I think that there's a lot of, there's a lot of space and it's a market that, that's used to being fragmented and having a lot of players. You know, what I think is interesting is that, um,

• 33:24Speaker 00

  maybe even on like the tech infrastructure side.

• 33:29Speaker 00

  I think in, in, in customer service especially, like we talked to a lot of customers where, you know, that question like, how's this different from like a, is this just a chat GPT wrapper? Why can't we do this ourselves, right?

• 33:45Speaker 00

  And I'll say it's not like there are a lot of different elements of this and you know, we use different models for different steps for sentiment detection, you know, talked about the voice stacking like text to speech, speech to text, you know, the answer computation, using like Claude versus OpenAI, like the reasoning models and like reasoning across workflows and agentic workflows and then using like different elements of the stack for like the RAG for knowledge answers. So like it's actually like a very complicated set of architectures.

• 34:20Speaker 00

  And then like ironically, we have found that building an AI, because things are improving so fast, it actually gets harder to like the model layer, the LLM layer, it gets harder to build at the application layer.

• 34:38Speaker 00

  And what I mean by that is like, if you think about building SaaS, like it wasn't actually that hard. I think it's a little bit harder, but than than most people think, but it wasn't actually that hard. Like you're like, let's choose Amazon, like AWS, or let's choose maybe Google Cloud or Azure. They're going to be around. It's all going to be the same stuff, database, it's all like well-tested stuff. And you know, like builds around what customers need. So you don't really have to make hard technical decisions. In a world where like, what's the best LLM going to be a year from now?

• 35:09Craig S. Smith?

  Yeah.

• 35:10Speaker 00

  What's the best reasoning model going to be a year from now? What's the best voice provider going to be a year from now? No clue. Well, then we have to either architect a system that can kind of like pick and choose stuff. And this stuff, it's not directly interchangeable. Like you can't just like pick up OpenAI and like slot in Anthropic without any work. Like you have to design a system that could do that. And then you have to make technical bets. You have to kind of go read the papers and say, okay, like 11 Labs versus Cartesian is one of the like voice decisions, or like a Base10 or Together AI Unhosted. Well, Cartesian says they have a state space model and like, so where's this going to shake out, right?

• 35:47Craig S. Smith?

  Yeah.

• 35:48Speaker 00

  If you make the right bet, you're golden. If you make the wrong bet, you have this 30% tax probably for the next year or two years, either like your product is worse, or you have to go rip it, you know, rip out this key part of your infrastructure and put in another piece. So ironically, building at the application layer is harder than ever with LLMs getting better faster.

• 36:07Craig S. Smith?

  Yeah. Have you guys built your system so that you can easily swap?

• 36:14Craig S. Smith?

  L L M's or other components?

• 36:18Speaker 00

  Yeah, that's exactly right. That's exactly right. And you know, I think we were cogent enough to do that with with the foundation models. And so, you know, there was an outage one day with I forget which one it was like this, but it was a multi hour outage on OpenAI. We could just kind of like press a button. Like, we actually have a literal button in our office to press a button, like flip it to Claude. And, you know, the, the, the, the, the, the features of it are different. It's kind of like personality or, you know, its style is different and like the way you prompt it is different. Like, so that's some of the, the technical depth. And then,

• 36:58Speaker 00

  you know, everybody talks about evals. The evals, you need to know what the you can't just like turn on a model like, you know, like when, when chat GPT launches a new model, you need to know that it's going to work. And, you know, I knew that at the architectural level, but even just to illustrate, like, you know, I read this tweet from Andre Karpathy, and he was like, there's two models you should use as a consumer, four O for easy stuff and O three for hard stuff. That's it. And I was using four or five like a dummy, like I have Pro. I was using four or five. And it was like, you maybe want to use four or five if you're doing creative writing, but probably not. And I was using four or five for everything because I assumed four point five is higher than four.

• 37:36Craig S. Smith?

  Yeah.

• 37:37Speaker 00

  So it's hard to keep up with this stuff and, and to know. So yes, we, we've architected our system around knowing what's the best model, being able to slot them in easily. But, but it takes a lot of like actually upfront work to, to, to get that design right. And, and I'm certain we haven't gotten everything right.

• 37:52Craig S. Smith?

  Yeah. Yeah. Evaluation is interesting. I had a conversation with somebody else in the last day or two about that.

• 38:03Craig S. Smith?

  Because the benchmarks have sort of yellowed and curled. They're no longer really fresh because people have figured out how to train to the benchmark. So how do you evaluate a model when you guys are looking at new models coming on stream? Or do you just wait until the market has decided?

• 38:28Speaker 00

  Yeah, we have what we call golden data sets. And so we actually.

• 38:35Speaker 00

  You know, I think this is like not a new motion for for for a company's, but you know, the Palantir pioneered forward deployed engineer. You send somebody to go sit closely with the team. Now, our product thesis is that actually we shouldn't build the product for you. Over half of our automations are like built by the companies themselves rather than by us. But what these people do when they sit with you is like help you define what good means. And so it's like, we'll show them how to do evals. We'll show them how to find the like ideal form of, okay, like this is your question on how to how to introspect your refund policy. Or, you know, one of our customers is one of the largest delivery apps, the super app in Latin

• 39:19Speaker 00

  America, Rappi. And it's like, okay, if you're a courier changing your vehicle registration and you have this like complex policy, this is what good looks like. This is what not good looks like. Here are a bunch of examples. So we build out these golden data sets. And then we kind of continually run every single time we upgrade models through all these golden data sets. And we have different benchmarks for what quality means that's customer specific as well as kind of like generalized. Yeah. And then there is like a step which we actually kind of borrowed honestly from OpenAI when, you know, when they had the call like psychopathy issue. I thought it was a really interesting insight into how they do evals of their models. There's a step they call just vibe evals.

• 40:03Speaker 00

  Like I don't know when vibe entered the lexicon, but for AI, it's like just you just check the vibes, you know, you go through and you look and you're like, whoa, yeah, that's good. No good. But but I think that there is something fairly deep in that because when I was working on machine learning for fraud detection at Stripe, the secret sauce was not like, I mean, it was partially the amount of data we had and the feature engineering we did. And back then Hadoop and big data was a cool thing and, you know, how great our Hadoop cluster was and you could compute all these features. But myself and this guy, Michael Manapatt and Greg Brockman, who's now president of

• 40:46Speaker 00

  OpenAI, was CTO at Stripe at the time, like a lot of people had touched this machine learning system and Michael ended up being CTO of Notion. But just to give you a sense of like, kind of the people, you know, the intellect working on this. What we would do is just like sit there and like click through manual reviews like, okay, that transaction like, yeah, looks pretty fraudulent. Good job. Like, no, I don't think so. That's right. Like, it would just do this for hours.

• 41:12Speaker 00

  Just check vibes.

• 41:14Craig S. Smith?

  Yeah, that's interesting.

• 41:17Craig S. Smith?

  Yeah. And I mean, just as a consumer, I'll, I'll run a prompt through two or three models, you know, on two or three different tabs. And it's remarkable how different the output is. And there are some that are very clearly, some models are very clearly better on certain prompts. And so

• 41:42Craig S. Smith?

  the other question I wanted to ask about the conversational element, do you do multi-language? And do you do translation? If you have

• 41:56Craig S. Smith?

  you know, someone calling into a call center in the Philippines is English speaking and they're calling from an

• 42:06Craig S. Smith?

  Arabic speaking country. Is there a translation element so that the call center, the live agent can handle that call?

• 42:17Speaker 00

  Yeah, there is. And you get a lot of this for free with like good choice of L L Ms underneath because, you know, they are a kind of like multilingual built in. But there is like getting it, getting the translations right and actually being useful for end customers and useful for like human support agents in the AI co-pilot setting is really important. And there's actually like a bit of a, you know, again, I studied economics before I studied statistics. So like my, my, my brain goes back to there. There's a, there's an economic part of this where like 24% of our revenue is in Europe. And when you go to Europe, like you find that the fragmentation is very real. Like, they have to serve each of these individual countries.

• 43:02Craig S. Smith?

  Right.

• 43:02Speaker 00

  Each have their own regulatory frameworks and each speech their own individual languages. And so, you know, one of our customers is a travel company based out of Berlin. And you need to staff like a handful of people for each of these languages. Now, if you're following the kind of like workforce management part of it and like the operations research, like, when you're in a low volume setting, you're going to have to like overstaff just to be able to hit like a good service level. So you find these teams being like, we have five people that speak Estonian. We have 10 people that speak Finnish, right? Like it just like da, da, da, da, da. It just like adds so much complexity to the operation.

• 43:42Speaker 00

  And so when they can do translation just out of the box, whether it's for their human support agents or all the way to the end consumer, it's pretty transformative to like simplifying the operation.

• 43:54Speaker 00

  And then US companies as well, kind of expanding into Europe. I mean, we have 20% of our revenue. We've constantly been thinking about like, okay, how do we do this a little bit better? It's, it's very hard to enter Europe knowing kind of like the fragmentation. So, so, so I do think that there's something that will change there in terms of like when translation has become a lot easier. And support is, is a big part of the operation that holds people back in terms of expansion.

• 44:17Craig S. Smith?

  Do you know a company called DeepL?

• 44:19Speaker 00

  Yeah, yeah.

• 44:20Craig S. Smith?

  Yeah, I talked to them recently and they were saying that their

• 44:27Craig S. Smith?

  solution is embedded or in a lot of other customer support solutions that they're the, the translation node or nodule is, is, do you guys use them at all?

• 44:40Speaker 00

  We don't use them. I think we've competed with them once or twice.

• 44:46Speaker 00

  But we know of them and, you know, I think it speaks to the just the ubiquity of customer support. Like I hadn't appreciated that everybody would touch this a little bit. And, and so, yeah, it's a really fascinating company and set of technologies.

• 45:04Speaker 00

  But yeah, we we we've run like a different parallel track in terms of architecture.

• 45:09Craig S. Smith?

  Yeah, I mean, that's interesting what you said that that so many people from different ends of the machine learning world or generative AI world are touching customer support. I mean, as you said, even Jensen Huang was is talking about it, which speaks to how large a market it is. But yeah, it's still it's still sounds daunting, you know, to compete in that. But but how do you?

• 45:48Craig S. Smith?

  I asked this before, but do you end up specializing in an industry or in a geography? You can't, you know, the market is so large, you can't chase every sector at once. So how do you do that?

• 46:07Speaker 00

  Yeah, I think the market is very large and, you know, the the trope of you need to have like focus on your ideal customer profile has definitely been true for for assembled throughout, you know, our history. But it's less about industry. Like, there are nuances, don't get me wrong. Like, I was at a fintech conference where they're talking about, you know, actually how the regulatory framework affects like their their their stance as a business. But if you get down to kind of like the latent variables, if you will, of what kind of like differentiates different types of support teams.

• 46:44Speaker 00

  There's every first of all, every company has support, every organization even like it's not just companies. Like, we work with the Georgia Department of Human Services. Like a three one one. So every, every organization has something that looks like support. And then the, the, the variables that actually matter are how transactional is it? Like, is it high volume, high velocity? Is it super hard? Like Datadog is a customer of ours. It's technical support engineering. That's pretty hard. That's going to look different from, you know, Honeylove Retail, like B to C.

• 47:23Speaker 00

  But so like, we pretty much end up covering every industry and having segments based on each industry.

• 47:30Speaker 00

  But in truth, like it's not the industry cut that matters. It's kind of the shape of your organization. And, and that comes down to like, um, um, the velocity and the complexity of the types of issues that you're dealing with.

• 47:43Craig S. Smith?

  Yeah. So in this, uh, the, the platform and the offering,

• 47:52Craig S. Smith?

  where are you growing? And where are you spending your, your, putting your investment? Is it in the, uh, on the translation end or the conversational, the voice? Or is it, uh, more in, uh, on the workforce management or has that been solved? Where, where should we look for new developments?

• 48:17Speaker 00

  Yeah. With, with Assembled, my, my focus is on AI agents, like the AI automation. And so, you know, I think chat is a very competitive space.

• 48:28Speaker 00

  You know, you mentioned several like, uh, companies and, and then voice is actually the one where I think there's the most space, like it is the one where the technology underlying is kind of moving most quickly. But then to go back to the, to the economics of customer support.

• 48:46Speaker 00

  Human phone support costs a lot. You know, I think people talk about like $20 per call. Like some customer of ours like have been able to optimize this actually quite robustly, like all the way down to single digit dollars. So it's not like people are just like super profligate about phone support. But AI is going to make a big, big, big difference there. And it's both, it's both the automation and the cost bringing, being brought down. But the, you know, like, there's some sense in which Jevon's paradox applies here. Where, you know, you have cheaper phone support and you're going to use more of it. And, and what I mean by that is

• 49:26Speaker 00

  a lot of, again, going back to consumers, I think companies know that consumers want to talk to somebody. But it just doesn't always make enough economic sense to do that. And so that's why they put people to chat, like chat, the cost is something like 50 cents up to a dollar per case. Because like you can have concurrency, like you can have people working multiple tickets at once. And it's just like less mentally taxing. You train people for phone support, you get like, it's like, prepare yourself.

• 49:57Speaker 00

  But I think with like AI voice,

• 50:01Speaker 00

  we're seeing more companies wanting to bring back phone support. And with AI chat, if you kind of like remove some of the work, then you can actually upscale people into phone support. So there's like a lot more texture around what's happening with AI agents now.

• 50:21Speaker 00

  I think also what's interesting in, in, in the space and where I spend some of my time thinking about is like, I think we're still not off the, uh, kind of like headline of like AI is going to make people go away. You know, like there was like this kind of like the news cycle of like Klarna had, you know, introduced AI customer service and, you know, fired 700 people and they were going to get rid of everybody. And then they immediately rolled it back. And that wasn't surprising to me because support is hard.

• 50:55Craig S. Smith?

  Yeah.

• 50:57Speaker 00

  It's not like people are sitting around twiddling their thumbs and just like doing it like pressing buttons inside of a, you know, kind of like knowledge equivalent of a factory. No, it's like they are exercising judgment. And when you go, when you go shadow support agents in these contact centers from all over the world, like they're really good at what this they have all the keyboard shortcuts sorted out. They know the knowledge base backwards and forwards. They know like the weird intricacies of your product that nobody else knows. They know where the bodies are buried, right? So there's like, they're getting through this really, really quickly. And so our thesis has always been like, there will be this kind of like hybrid human plus AI support that companies are going to have to think about.

• 51:39Speaker 00

  And, and, and so like, you know, when I say like a focused on AI agents, like there's the literal higher quality AI agents, higher quality within voice where there's a great opportunity. But also like connecting it back to what happens with the human operation and that kind of like what's going to happen with the upscaling of the rest of the people. Because I think that's the missing link. Everybody's kind of focused on like this part of it, but like you gotta connect this to back to what humans are still doing in the rest of support.

• 52:07Craig S. Smith?

  Yeah. And because you're not managing call centers, right? You don't have your own call centers. No. So how do you ensure that that the call centers are, that the human agents are, are using all of the data that you're feeding them so that the customer doesn't have to repeat himself or whatever?

• 52:34Speaker 00

  Yeah, that, that is the routing layer. And at the kind of like, how do we pass the right information out of the AI conversation into a ticket that makes it to the person that might have to get, that might have to be escalated to? So that's how you kind of like skip repeating like we, we, in this like workers comp use case, we gather the information. There's an integration into like the underlying telephony system or the underlying ticketing system. So like a five nine or a Zendesk or a Salesforce Service Cloud. And you kind of have to do a little bit of like integration work to get this handoff fairly seamless. And so they, they see in context like boom, okay, here's everything that happened and it's like summarized in the form as possible. And now here's what's next for you. And by the way, here's the playbook

• 53:19Speaker 00

  if you don't know it. So I think this is where like the AI agent automation and the AI co-pilot start to collide. The other part that gets me excited, the operations research or the kind of orchestration layer or system level thinking is support is kind of like this very data-driven space. Like Moneyball is the best equivalent I could give you. Like we're used to messing around with data. And like, if you think about like designing a support operation and you know, before you had a couple of levers, like, okay, we'll hire people, we'll move people around between queues, maybe we'll cancel meetings and like shuffle people's calendars around.

• 54:04Speaker 00

  And now you can turn on an AI agent. Well, like there's this kind of like routing question, like what types of volume, what parts of the conversation, the first part, all of it, the middle part are going to people with which skills? So I think that really is like, aside from the direct experience at the system level, like the design that we are most excited about is like, okay, what should go to who at what point in time? Or another way to put it, like, how do we get the right agent at the right time, the right answer?

• 54:35Craig S. Smith?

  Yeah.

• 54:35Speaker 00

  And we're intentionally overloading the agent term there.

• 54:38Craig S. Smith?

  Do you think that agents?

• 54:42Craig S. Smith?

  I mean, the customer support experience from the

• 54:48Craig S. Smith?

  from the customer's point of view.

• 54:52Craig S. Smith?

  Right now, maybe 5% of their, I don't know what it is, I'm making that up, but of their issue is handled by an agent. You know, at companies that have deployed agents. That that will increase to where agents will be handling

• 55:14Craig S. Smith?

  75% of of their issue and then

• 55:22Craig S. Smith?

  25% will be, as you say, sort of these upskilled human agents that can handle more complex issues.

• 55:32Speaker 00

  Yeah, I think that's about right on the volume perspective, like plus or minus, you know, a handful of percentage points like I've talked to enough people and seen our own results that's like, yes, there's 75% like automated resolution of support tickets. Which, you know, kind of like perhaps implies 25% going to people, but there is a paradox that I don't know how to synthesize yet, which is that like a lot of the support leaders we talk to.

• 56:07Speaker 00

  You know, they're like, we have this bot. 75% resolution. And it's not that people are making it up, like there is some sense in which I think there's like overblown claims, but then you like inspect, you check vibes and you're like, oh, that's pretty good. Yeah, that's like a pretty good answer. It's not totally made up. And then, and then you go look at the amount of work going to humans and this is kind of like a unique vantage from our workforce management products. Like we know the total handle time of work, the total volume of work and It's kind of the same.

• 56:40Craig S. Smith?

  Yeah.

• 56:41Speaker 00

  And then you go look at like the, the cost, like the labor spend and it's like, it's the same. And, and so like all these other downstream metrics, like you would presume that if like all of this automation is happening, like this stuff is, you know, kind of like decreasing, that's not happening. So I don't know how to square these things quite yet. Um, but just to observe that, um, um, you know, that further informs our thesis that like, okay, I don't think humans are going away. We still have to unpack like why, but even with automation, people are not going away.

• 57:13Craig S. Smith?

  Is there anything I didn't touch on that you wanna, you wanna say? And, um, and what's the URL for people that wanna get in touch?

• 57:24Speaker 00

  Yeah.

• 57:27Speaker 00

  So we're just www dot assembled.com. And I've debated with people internally about whether that should be assembled.ai or assembled.com. We just stuck to our guns, assembled.com.

• 57:39Speaker 00

  You know, one last thought that I would leave you with is, um,

• 57:46Speaker 00

  we all get excited about the technology and the automation potential. And if there's one kind of like common current throughout this, whether that's the human plus AI kind of like stitching together of that frame, or, or just like what, what happens with AI adoption?

• 58:03Speaker 00

  I think the economics matter. Like many of our customers, you know, just to go back to that point of like, they have outsourced people, they have optimized their operation. It's been decades of people like coming to support, being like, okay, let's optimize this, let's like automate this, let's, so it's not new. And so, uh, just because you can automate support, does not necessarily mean it would be economically viable to do so across the entirety of the business. We are seeing people, um, start to question now, like what they're looking at ROI, okay, like, you know, the AI agent is charging 50, 60 cents.

• 58:43Speaker 00

  But hey, we're in Colombia, like, you know, and the, the wage is not that high and these people are really good at what they do. Um, you know, they can do it for cheaper. People can do this for cheaper. So, so like, I think that will be a question that comes to a lot of industries now. Um, I think support starts with it just because the data is so obvious. Um, just because you can automate it does not necessarily mean that you should.